Current Evidence Claims

Role: current benchmark/readiness claim snapshot. Status: current owner for numeric evidence claims, supersession, supports / material_limits report projection, material claim boundaries, and confirmed scope-boundary remediation pointers.

This is the current-claims snapshot for benchmark and readiness numbers that are easy to over-read when old dated ledgers still say "current" in their local context. It is not a command ledger and it does not replace source reports.

Snapshot review date: 2026-06-15. Row-level run_date, metric IDs, and source-report dates are authoritative for promoted rows added after that full-file review date.

Claim Reviewer Card

current_status: current owner for promoted benchmark/readiness numbers; dated reports remain bounded evidence until a row below promotes or supersedes them.

can_say: AIppocampus has source-backed continuity evidence for the named rows, within each row's run_date, cohort, claim_level, supports, and material_limits.

cannot_say: do not claim new benchmark quality, public readiness, live host behavior, private-history quality, or official external scores from a historical report title alone.

owner_routes:

Current numeric claims: docs/evidence/current-claims.md#current-claim-snapshot.
Readiness status and gaps: docs/evidence/readiness/stage-0-5-readiness.md.
Benchmark report routers: docs/evidence/benchmarks/reports/README.md.

claim_safe_next_action: open the relevant claim row and source report first; use docs-health only after deciding whether the row needs promotion, supersession, or a no-open-followup note.

next_verification_command:

python tools\aippocampus\docs\check_docs_health.py --json

Detailed Evidence Index

Rules:

A value is current only for the run_date, cohort, and claim_level named in its row.
Dated evidence remains in docs/evidence/readiness/public-readiness-verification.md and detailed benchmark methodology remains in docs/evidence/benchmarks/design/memory-decision-benchmark-plan.md.
Benchmark priority, run-profile, and claim-boundary ownership remains in docs/evidence/benchmarks/design/benchmark-priority-map.md.
Benchmark maturity, sample-size floors, and the separation between contract_gate_ok and quality_gate_ok remain in docs/evidence/benchmarks/design/benchmark-maturity-gates.md.
Closed benchmark/evidence issue states such as harness-ready, pilot-run, contract-smoke, blocker-recorded, and completed-score remain in docs/evidence/benchmark-evidence-maturity.md. Use that ledger before reading an issue title as a completed score.
Stage-level positive claims and launch boundaries remain in docs/evidence/readiness/stage-0-5-readiness.md.
New public-facing benchmark reports should project measured_result, supports, and material_limits first. Keep cannot_claim as a short legacy compatibility or durable-boundary field, not as the default reader foreground.
Demo caveats in docs/guides/demo-scenarios.md are claim-boundary inputs, not standalone benchmark proof.
AMemGym adapter smoke currently has no Current Claim Snapshot row; its native-score, source-backed overlay, diagnosis, utilization, and cost/latency boundaries stay in docs/evidence/benchmarks/amemgym.md until a dated result owner upgrades a claim.
STATE-Bench Agent Learning feasibility now has a blocker/defer Current Claim Snapshot row, not a score row. Adapter/readiness and one-domain preflight evidence stay in docs/evidence/benchmarks/state-bench-agent-learning.md; the 2026-06-14 decision rechecked the official source and records that the locked GPT-5.4 evaluation client endpoint/deployments are still not configured in this environment.
PersonaMem / PersonaMem-v2 currently has no Current Claim Snapshot row and should not be run as a retrieval-only quality claim. A first public-safe Ficus MVP exists for schema/lifecycle/hard-mask fixtures, but it is not a PersonaMem score or broad profile-quality row. The readiness boundary stays in docs/evidence/benchmarks/personamem-readiness.md.
LoCoMo text-QA currently has a fixed-reader answer/latency harness, but no Current Claim Snapshot row yet. Keep its dry-run as schema/privacy evidence only; a dated fixed-reader provider report is required before citing answer quality.
LongMemEval-S now has a first dated fixed-reader provider answer/latency baseline on a 25-question public cohort. Keep it separate from the larger 500-question retrieval-only rows and from official LongMemEval judge claims.
LongMemEval-S now has 500-question rows that separate the current source-worker-surface proxy, the opt-in LLM query/candidate upper bound, and the contract-aware full-source semantic-scope warming path. Treat the older 100-question query/cache row as progress evidence. The full-source warming row improves sidecar coverage but does not improve fused R@10, so it is a measured bottleneck, not a product-quality closeout.
LongMemEval-S now also has a 500-question source factual-alias closeout row. It improves bounded candidate evidence coverage to 463/479 = 0.9666, keeps fused R@10 at 422/479 = 0.8810 with zero fused regressions, and records zero hot-path provider calls. This closes #1323/#1327/#1424 only as bounded source-side retrieval evidence, not as answer generation, official QA score, or source truth from aliases.
LongMemEval-S now has a #1327 source-window coverage diagnostic that separates candidate-missing misses from reranker-visible misses and records a bounded candidate-coverage projection plus a rejected naive large-radius negative control. It does not make route candidates source truth or justify dumping wider foreground context by default.
LongMemEval-V2 currently has an official-harness pilot decision and text-only Memory adapter contract, but no Current Claim Snapshot row yet. Keep the decision report and any tiny dry run as integration evidence only; a dated official-harness run is required before citing V2 answer accuracy, LAFS, or latency quality.

Confirmed Scope Boundaries (Expected Null Results)

Start here before opening dated report history. These rows are the current reader-facing ledger for open and resolved remediation routes that are easy to over-read as either broad product failure or broad product proof. A confirmed scope boundary is still real evidence: keep the original result visible, then state the condition where the result does and does not apply.

Continuous-memory short complete-spec boundary

Current evidence state: The 2026-06-09 #960 product-remediation rerun preserves the expected-null boundary: lower_bound_passed=false; the historical row keeps decision_label=no demonstrated memory advantage, and the public interpretation is no demonstrated net advantage over modeled fresh-context spec loop. The product path now reports product_change_status=implemented_rerun for source-miss fallback: source-required packets with no reopenable ref ask for a minimal source anchor instead of guessing or silently dying.
Remediation route: #960 is closed as the first remediation/rerun slice; a future repeated-evidence row would need a new owner before superseding this boundary.
Reader boundary: Do not claim continuous-memory superiority until a later dated row supersedes this evidence; also do not read this as evidence that source-backed recall has no value in context-loss, cross-thread, post-compaction, or long-horizon scenarios.

Track B private source-evidence retrieval

Current evidence state: The current selected private slice is 100/100 top-5 after the #963 dynamic-source ranking repair; the pre-repair same-day run exposed 5 top-k pruning misses, all with the gold source present in the raw candidate pool.
Remediation route: #963 is closed; the fix is reflected in the current snapshot row.
Reader boundary: Do not turn the selected slice into broad real-user gate quality, semantic completeness, or public benchmark score.

Retrieval score-fusion calibration

Current evidence state: The 2026-06-10 public fixture reports 5/5 expected top matches across exact-quote guard, question-tracking semantic bridge, wrong-stance lure suppression, explicit vector-unavailable fallback, and missing source-join rejection. The 2026-06-14 source-joined routing decision keeps text-first defaults and defers vector prefilter/local embedding adoption.
Remediation route: #309 closed as a bounded routing/fusion decision owner.
Reader boundary: Do not treat this as default vector-prefilter adoption, local embedding adapter evidence, live answer quality, source truth, or broad vector/graph/rerank consumer safety by default.

AMemGym official live-provider score blocker

Current evidence state: AMemGym has adapter/protocol/overlay evidence but no Current Claim Snapshot metric row yet; the 2026-06-11 #1232 resume attempt used #1229 provider-budget preflight and a topped-up OpenRouter account, then diagnosed the post-top-up failure as an OpenRouter provider-route block: the required OpenAI-family routes fail even on a harmless fixed prompt, while non-OpenAI OpenRouter routes still accept that prompt. Outputs remain partial: overall is still partial at 6/20 user items, upperbound is still partial at 38/882 choice evaluations, and random is complete.
Remediation route: amemgym-official-live-provider-1232-blocker-2026-06-11.md closes #1232 as the current provider-route-blocked report; amemgym-official-live-provider-1083-checkpoint-2026-06-10.md closed #1083; amemgym-official-live-provider-blocker-2026-06-09.md closed #958.
Reader boundary: Do not present protocol output, partial OpenRouter output, bounded subset output, resume/checkpoint output, provider-budget adoption, route-preflight output, or overlay diagnostics as an official live-model score.

E2E50 public behavior-pack / private-local blocker split

Current evidence state: The 2026-06-12 #279 update makes the checked-in 50-case public-safe behavior pack the primary executable E2E50 path: it covers silent constraints, rejected-route avoidance, transient-concern extinction, superseded currentness, scope-limited constraints, summary-overhang traps, no-remember negatives, and source reopen before risky action. The private/local #1086/#279 follow-up remains useful dogfood: the wide scanner finds 23/20 candidate seeds, but local annotation retains only 7 control/seed cases against the 20-case private target.
Remediation route: python benchmarks/aippocampus/benchmark_e2e50_silent_constraint.py --json is the public behavior-pack replay; e2e50-private-local-seed-followup-2026-06-10.md closes #1086 as a private/local blocker report; e2e50-private-annotation-readiness-2026-06-10.json is the optional private scorer replay.
Reader boundary: Do not make private-history case scarcity the main public E2E50 blocker. Also do not claim private-history behavior lift, completed private 20-case quality, representative or live 50-case E2E50 quality, or live host behavior from the public pack.

React VCS lexical near-miss false positives

Current evidence state: The current 2026-06-09 row reports 60/60 gold true positives and 30/30 explicit-cue hard negatives suppressed.
Remediation route: #961 is closed; the fix is reflected in the current snapshot row.
Reader boundary: Do not use older 2026-06-04 false-positive evidence as the current source-disambiguation state.

Progressive recall route follow-through gaps

Current evidence state: The current row reports route_actionability_rate=1.0 and eligible source_reopen_follow_through_rate=1.0, with stale handles rejected before source use.
Remediation route: #962 is closed; the fix is reflected in the current snapshot row.
Reader boundary: This still does not close broad live #201 or default foreground-lift claims.

Multimodal NIAH stale conflicting-source selection

Current evidence state: The current row reports 4/4 answer/source-selection/source-anchor-citation after conflict repair and keeps the ambiguous-currentness negative control as reopen-required.
Remediation route: #964 is closed; the fix is reflected in the current snapshot row.
Reader boundary: This is a supplied-pool synthesis contract, not retrieval or live vision-model quality.

Track S explicit-negation/currentness failures

Current evidence state: The current 2026-06-09 row reports quality_gate_ok=true for the diagnostic Track S threshold only, S1 false evidence 0, and S3 explicit-negation/stale-as-current/evidence-over-escalation counts all 0.
Remediation route: #992 owns this repair; the fix is reflected in the current snapshot row.
Reader boundary: This is a public-safe deterministic hook/retrieval diagnostic, not human-level semantic understanding, public product quality, or broad private-history recall quality.

Semantic source-review operational partial failure

Current evidence state: The current 2026-06-09 #1053 96-case rerun has failure_count=0, failure_taxonomy.by_class={}, and failed_label_categories=[]; preference met the per-label floor at 6/9.
Remediation route: #993 and #1053 are closed; the fixes are reflected in the current snapshot row.
Reader boundary: Do not treat the broader diagnostic as human review, provider-independent quality, a green gate, or full semantic correctness.

Continuous-Memory Reading Boundary

Reading boundary: The 2026-06-08 repeat profile is an expected null result for a short, complete-spec, public-synthetic task slice. AIppocampus is designed for context loss, cross-thread recovery, post-compaction source reopening, and long-horizon continuity; this experiment deliberately includes a condition where rebuilding fresh context is expected to be cheaper. The result confirms a scope boundary: memory overhead is not justified when the task already has a complete fresh spec. It should not be read as evidence that source-backed recall has no value in the scenarios AIppocampus is built for.

Keep the historical metric id discoverable: continuous_memory.preregistered_repeat_profile_2026_06_08. When citing the scope interpretation, prefer the alias boundary_confirmed.short_task_complete_spec_synthetic_2026_06_08 and the decision phrase no demonstrated net advantage over modeled fresh-context spec loop.

Adjacent evidence prevents this boundary row from becoming the whole project verdict:

`boundary_confirmed.short_task_complete_spec_synthetic_2026_06_08`

What it shows: Fresh context wins for short complete-spec tasks under the current cost ledger.
Boundary: Expected-null boundary; memory overhead is not justified in this condition.

`continuous_memory.context_loss_public_continuity_2026_06_10`

What it shows: The #1153 readout adds a stable missing-context diagnostic slice: github_1153_context_loss_public_continuity_v1; contract_gate_ok=true, quality_gate_ok=false; selected deterministic public-safe cases keep fresh_missing_context, summary_only_host_native, aippocampus_route_packet, sham_unrelated_memory, stale_wrong_memory, oracle_full_context, and the old complete-spec boundary separate.
Boundary: Contract slice only. It does not supersede continuous_memory.preregistered_repeat_profile_2026_06_08, claim public-quality continuous-memory advantage, or prove LoCoMo / private / live continuity quality.

`recall_navigation.progressive_route_follow_through`

What it shows: Progressive recall routes can become actionable and reopen source in deterministic fixtures.
Boundary: Not broad live foreground lift.

`recall_navigation.macro_prior_fixture_2026_06_12`

What it shows: The opt-in agent recall path can consume scoped project macro state from a project-local default file, and the promotion harness shows one public-safe macro case where the prior changes active-layer route ordering, widens fanout, emits momentum recheck diagnostics, and reduces manual-search fallback with zero red-line counts.
Boundary: Fixture/default-path contract only; not macro default-readiness, live foreground lift, private-history quality, or source evidence from macro state.

`macro.transform_orbit_diagnostic_2026_06_12`

What it shows: Deterministic macro diagnostics distinguish C/R reversible transform orbits from line-flip adjacency and non-invertible nuclear projection basins, while keeping every packet navigation-only and source-reopen-bound.
Boundary: Structural diagnostic only; not Dream quality, route merging, evidence support, foreground prose, default ranking, private-history behavior, or Yi interpretation quality.

`macro.timing_recheck_experiment_2026_06_12`

What it shows: Public-safe timing experiment maps source-backed deltas onto active line axes and source-epoch cadence thresholds; it emits recheck diagnostics without mutating currentness, replacing temporal cues, using calendar cycles, or changing default ranking.
Boundary: Fixture/research candidate only; not live recall adoption, quality lift, cost/token savings, private-history evidence, literal calendar timing, or fact/currentness support.

`fresh_thread.public_validation_2026_06_10`

What it shows: The #281 public validation readout records 6 positive public flows, 5 negative controls, first-turn scent precision 1.0, progressive activation gain 1.0, source reopen before specific claims 1.0, irrelevant-memory drag 0.0, over-personalization count 0, and manual-query invention count 0.
Boundary: Public fixture closeout only; not live fresh-thread quality, private-history quality, universal fresh-thread recall, foreground-hook-only sufficiency, or innate model memory.

`live_semantic.route_actionability_2026_06_07`

What it shows: Live semantic route hits can become reopenable or source-required routes on a public checked-in corpus.
Boundary: Not private-registry vague recall quality.

`track_b.private_semantic_sidecar_required`

What it shows: Private real-history sidecar retrieved the expected source in 100/100 selected top-5 cases after the #963 dynamic-source ranking repair.
Boundary: Maintainer-only private slice, not public benchmark score.

`episode_arc.private_history_adjudication_2026_06_08`

What it shows: Rejected-route arcs exist at scale in private aggregate history.
Boundary: Diagnostic aggregate, not live behavior lift.

`episode_arc.public_gappy_chain_calibration_2026_06_10`

What it shows: Selected public fixtures distinguish complete, gappy, wrong-order, single-point, and temporary-concern arcs, with gappy chains projected to ask/refresh only.
Boundary: Public deterministic fixture, not broad live or private-history quality.

`cognitive_load.public_behavior_trace_feedback_2026_06_10`

What it shows: Selected public fixtures separate useful caution, irrelevant load drag / false positives, and over-personalization-risk feedback; the 2026-06-14 default-path replay then narrows current adoption to diagnostic-only.
Boundary: Public deterministic fixtures, not live host-timing or default foreground policy quality.

`journey.public_time_sliced_replay_2026_06_10`

What it shows: Selected public replayable cohort builds four time-sliced Journey candidates, excludes four future rows, surfaces one active relevant-prompt hint, suppresses resolved/stale/wrong-route relevant prompts, and suppresses 12 source-visible / unrelated / high-risk controls without serializing raw rows, source refs, or route handles.
Boundary: Public deterministic fixture, not live/private Journey quality or default foreground usefulness.

`dream.shadow_route_topology_scout_2026_06_12`

What it shows: Public-safe Dream topology scout now emits source/residue-gated shadow-route candidates and transform-orbit deepen candidates only when the shadow gate is already satisfied; generic shared vocabulary and pure orbit membership remain no-candidate/explain-only.
Boundary: Deterministic scout fixture only; not live Dream quality, private-history usefulness, route merge, hidden user intent, foreground usefulness, or source support from transform structure.

`thread_story.public_shadow_closeout_2026_06_10`

What it shows: Public structured-text #313 closeout report covers source-backed thread-story packet boundaries, packet-only factual-answer blocking, source-reopened answer comparison, and leakage / contradiction / persona / interference / unrelated-noise controls.
Boundary: Public deterministic closeout fixture, not private-history story quality, live model behavior, default recall lift, or source truth from packet routes.

`attention_router.contract_fixture_2026_06_10`

What it shows: Public-safe contract fixture proves hard masks block high-relevance routes, route packets are handles not facts, source-backed packets stay reopenable_route before source reopen, bounded summaries remain navigation-only routes, bounded evidence requires source-open/bounded scope, and masked source resurrection count is 0.
Boundary: Public deterministic contract fixture, not broad router quality, private-history behavior, model training, summary-as-evidence, or default adoption.

`attention_router.route_token_fixture_2026_06_10`

What it shows: Public-safe route-token fixture projects a long event into a tight source-span token, preserves reopen handles through event and episode/question tokens, carries salience/currentness/privacy/conflict slots, and keeps grouped episode/question tokens navigation-only.
Boundary: Public deterministic token-projection fixture, not hot-router scoring quality, private-history token quality, source truth without reopen, or default adoption.

`attention_router.hot_router_fixture_2026_06_10`

What it shows: Public-safe deterministic V0 router fixture applies hard masks before scoring, emits audited head votes, uses calibrated_rule_grid_v1 as the current route-token score policy, raises thresholds for stale/conflict risk, routes positive and stale/conflict tokens as reopenable routes, and leaves weak tokens direction-only.
Boundary: Public deterministic router fixture, not learned attention, production calibration, private-history quality, live foreground usefulness, or default foreground-hook adoption.

`attention_router.action_head_fixture_2026_06_10`

What it shows: Public-safe action-query fixture shows pending path/issue/test cues can lift a route when prompt-only terms are weak, while privacy masks and anti-nag suppression still block or downshift action-matched routes.
Boundary: Public deterministic action-head fixture, not live hook behavior, private tool-argument quality, E2E50 lift, or default foreground action routing.

`attention_router.evidence_packaging_fixture_2026_06_10`

What it shows: Public-safe evidence-packaging fixture preserves first-stage source-window retrieval while tightening context-visible candidates into source-span route packets; source-open/current/unconflicted spans can become bounded evidence, and wrong-source/stale/conflicted controls stay rejected or reopenable.
Boundary: Public deterministic packaging fixture, not LongMemEval QA score, exact-line quality being solved, reranker output as source truth, private-history packaging quality, or default foreground adoption.

`attention_router.navigation_quality_public_cohort_2026_06_13`

What it shows: Public-safe benchmark keeps the 12-case contract fixture green and adds a 9-family public/holdout cohort with explicit gate names: contract_safety_gate_ok, router_design_gate_ok, public_quality_gate_ok, and default_adoption_gate_ok. Legacy quality_gate_ok is a public-quality alias, not a V0 design verdict.
Boundary: Public/holdout navigation-quality gate for the narrow explicit agent-pull path; not live host behavior, answer-generation quality, private-history behavior, broad live/default route-producer quality, broad default-session usefulness, or default foreground-hook adoption.

`attention_router.score_fusion_calibration_2026_06_11`

What it shows: Public-safe #1112/#1230 calibration/adoption report exports 12 sanitized attention feature rows; legacy deterministic weights still show the old false-negative/anti-nag gap, while runtime_default_policy=calibrated_rule_grid_v1 matches the calibrated arm at precision/recall 1.0 with all red lines 0.
Boundary: Public deterministic hot-router policy adoption, not default foreground-hook adoption, private-history training quality, answer-generation quality, source truth from scores, or production calibration.

`attention_router.agent_recall_opt_in_sorting_2026_06_12`

What it shows: aippocampus agent recall --attention-router can project existing recall_context routes into attention-route tokens, reorder only already emitted reopenable routes, keep all original routes as fallback candidates, and report public-safe attention_router_navigation metrics without foregrounding source refs or head votes. The same projection now reports no-help diagnostics such as applied-but-no-help, query/route overlap, bridge-reason, specificity-floor, and why_this_may_matter specificity signals.
Boundary: Public-safe opt-in agent route-sorting contract and usefulness diagnostics, not default foreground adoption, broad live route-producer quality, private-history router quality, answer generation, source truth from scores, or broad live/default successor quality.

`attention_router.agent_recall_auto_policy_2026_06_13`

What it shows: aippocampus agent recall --attention-router-mode auto checks the shared recall-navigation promotion harness before enabling router sorting. Neutral no-op cases are now ROI signals instead of hard blockers, true no-help/feature-hurt/red-line controls remain visible, and the public/holdout cohort can enable sorting for the explicit agent-recall surface.
Boundary: Explicit-pull auto policy only; not default hooks, every-turn recall, live-host usefulness lift, private-history quality, answer generation, evidence support from attention scores, or broad live/default successor quality.

`recall.default_hook_usefulness_2026_06_20`

What it shows: Public-safe #1439/#1449 four-arm benchmark compares no-packet baseline, explicit recall, default-hook foreground candidate, and default_hook_tiny_agent_recall_affordance under the same packet/source-reopen budget. Explicit recall and the tiny affordance both reach helpful next-action rate 0.636364 and manual-search reduction 11; the broad default-hook foreground candidate remains rejected with helpful rate 0.272727, wrong-route drag rate 0.363636, and irrelevant-memory drag rate 0.272727. The tiny not evidence affordance now has a host-faithful replay gate with 7 emitted/followed agent_recall calls, 7 recall-after-hint successes, 0 broad-manual-search-before-recall, and 0 wrong-route, irrelevant-memory, source-truth, raw-handle, or provenance-dump counts.
Boundary: Public synthetic diagnostic benchmark only; not live default-hook or tiny-affordance quality, private-history question/theme usefulness, default foreground adoption readiness, source truth from theme/load/router rows, or source claims from the tiny affordance without recall/deepen/source reopen.

`recall.activation_dogfood_usefulness_2026_06_28`

What it shows: Bounded #2857 dogfood/replay probe compares cold no-activation with warm replay signal across four cases. Warm replay increases verifier-seen candidates from 2 to 5 while keeping foreground exposure flat at 2, improves useful source-open hits from 1 to 2, lowers manual-search fallback from 2 to 1, and lowers wrong-route drag / noisy surfacing from 1/1 to 0/0.
Boundary: Deterministic dogfood/replay usefulness probe only; not causal real-user lift, live default foreground quality, private-history quality, Dream default delivery quality, or source truth from activation signals.

`map_rot.lifecycle_debt_2026_06_11`

What it shows: Public-safe lifecycle fixture covers 9 stale, challenged, quarantined, superseded, missing-middle, deleted/no-recall, dead-lettered, repeated-wrong, and current-route cases; hard red-line counters are 0, benchmark_maturity_level=contract_smoke, contract_gate_ok=true, and quality_gate_ok=false. The same runner now emits a no-write maintenance plan: 8 hot-surface removals, 2 review-queue rows, 2 source-refresh reactivation rows, and 1 dead-letter compaction row.
Boundary: Public deterministic lifecycle-debt contract gate and maintenance-action plan, not representative map-rot distribution, automatic cleanup proof, completed forgetting, conflict auto-resolution, private-history map-rot quality, or live current-route quality.

`agent_continuity.loop_gate_2026_06_11`

What it shows: Public-safe integration gate composes semantic warm routes, hot-router packets, agent-native foreground/deepen/explain, AIppo activation, source-reopen budget, foreground budget, packet triage, blocked/stale/conflict, and anti-nag cases; 8/8 cases pass, all red lines are 0, packet_triage_distinctiveness=1.0, blind_deepen_required_count=0, benchmark_maturity_level=contract_smoke, contract_gate_ok=true, and quality_gate_ok=false.
Boundary: Supports deterministic composition and route-selection usefulness on checked-in fixtures; material limits are no private-history or live-host lift, no answer-generation or external-model quality, and no default foreground adoption claim.

`agent_continuity.foreground_action_card_2026_06_15`

What it shows: Public-safe #1620 foreground field-budget contract: default agent recall / MCP payloads include a bounded foreground_action_card before audit details, with decisions for use_route_first, continue_normally, deepen_before_claim, and ignore_or_blocked; callable local handles are present only in local JSON and redacted from public projection; audit-only fields stay out of the card while full diagnostics remain available.
Boundary: Deterministic fixture and replay proxy only; not live behavior lift, default-hook adoption, source truth from the card, or proof that every agent will avoid broad manual search.

`telepathy.opt_in_handoff_workflow_2026_06_12`

What it shows: Opt-in local Telepathy handoff workflow stores append-only card lifecycle events, supports CLI create/list/deepen/release/diagnose, exposes MCP read-only list/deepen tools, keeps candidate-only cards navigation-only, and sanitizes source refs to selector fields or fingerprints.
Boundary: Supports a local handoff-card workflow over the existing packet/topology contract; material limits are no distributed lock correctness, no MCP write surface, no automatic work assignment, no live multi-agent usefulness claim, and no source truth without reopen.

`sparse_provenance.codebook_v0_2026_06_12`

What it shows: Public-safe #1190 V0 fixture builds 8 clean-source-like entries into 6 content-addressed chunks (compression_ratio=0.75), reduces route-chain lookup candidates from 8 to 5, suppresses 2 stale/quarantined matches, emits 0 foreground reconstructed text items, and reports all reconstruction/privacy red lines at 0 with topology preservation status=ok.
Boundary: Supports deterministic sparse provenance build/lookup/rehydration proof on a checked-in fixture; material limits are no natural-dialogue template compression, no zstd value claim, no neural MoE routing, no private-history behavior, no GB/TB-scale infrastructure readiness, and no source truth from codebook routes.

`avatar.bounded_resonance_proxy_pilot_2026_06_12`

What it shows: Public-safe #1319 exploratory proxy fixture runs 12 cases across closeout, repeated-debug-route, and structural-break families through arms A-E; the deterministic proxy marks bounded resonance as best (average_helpfulness_score=5.666667), standalone archetype alias drift as visible (off_topic_archetype_expansion_count=12), and bounded-resonance candidate red lines as 0.
Boundary: Exploratory deterministic proxy only, not live model behavior, production agent lift, default foreground avatar readiness, private-history quality, archetype authority, or source truth from posture/resonance.

Claim-Boundary Owner And Retirement Ledger

This table keeps testable material limits from becoming permanent background caveats. Reports may still use the legacy cannot_claim field for compatibility, but reader-facing prose should lead with what the result measured and supports. Use the metadata shape category, owner_issue, retirement_condition, and next_review when adding or reviewing rows.

actionable entries need an owner issue and a retirement condition. durable_non_goal, research_blocked, and external_dependency entries can remain without a direct implementation owner when a test would not honestly retire the caveat. Do not retire a broader caveat from a narrow smoke; add or update a dated claim row first.

Caveat	Category	Owner issue	Retirement condition	Next review
Continuous-memory superiority after the preregistered repeat negative result	`actionable`	#960 closed remediation owner; #378 / future repeated-evidence slice	#960 retired the first remediation/rerun blocker by implementing source-miss anchor recovery while preserving the expected-null result. This broader caveat retires only when a later dated repeated-evidence row supersedes `continuous_memory.preregistered_repeat_profile_2026_06_08` with a passing preregistered lower-bound decision and explicit cost/harm boundary.	Before any #378 or continuous-memory public-readiness upgrade.
Track B private source-evidence retrieval top-k misses	`actionable`	#963	Retired for the current selected private slice by the 2026-06-09 `track_b.private_semantic_sidecar_required` row: the top-k misses were taxonomized as candidate-generated/rank-pruned, repaired through a public analogue, and rerun at 100/100 without broad semantic-completeness claims.	Reopen only if a later selected Track B row reports top-k misses or before widening Track B beyond the current selected private slice.
E2E50 public behavior-pack path	`actionable`	#994 / #1154 / #279	Retired for the immediate public/shareable behavior-pack path by the 2026-06-12 `e2e50.public_safe_behavior_pack_contract_2026_06_12` row: a 50-case public-safe synthetic behavior pack is checked in with no-remember negatives, summary-overhang traps, scope-limited constraints, source-family coverage, and a report boundary.	Reopen before claiming representative E2E50 quality, private-history behavior lift, or live-host behavior quality.
E2E50 private/local annotated seed pack	`actionable`	#1086	The 2026-06-10 follow-up retires the ambiguous private/local candidate-count blocker by showing 23/20 wide-scan candidates, but preserves the private annotation blocker: only 7 retained/control cases against the 20-case private target. This is optional dogfood evidence after #1154, not the main public E2E50 gate. Retire this caveat only before claiming private-history E2E50 quality.	Before claiming private-history behavior lift, completed private 20-case E2E50 quality, or representative 50-case quality.
Fresh-thread public fixture validation	`actionable`	#281	Retired for the public, inspectable fixture closeout path by the 2026-06-10 `fresh_thread.public_validation_2026_06_10` row: the checked-in runner reports first-turn scent precision, progressive activation gain, source reopen before specific claims, negative-control drag suppression, over-personalization suppression, and manual-query invention suppression. Future live/private/external-dataset work needs a new scoped owner.	Reopen only before claiming live fresh-thread quality, private real-history quality, universal fresh-thread quality, foreground-hook-only sufficiency, or innate model memory.
Claude Code hook real-host firing and unsupported events	`actionable`	#1020	A local or public-safe real-host event log shows the scoped handlers firing, and later slices add payload-safe `PostToolUse`/`PostToolBatch` or compaction support before those events are claimed.	Before widening the scoped #1020 synthetic hook contract beyond `UserPromptSubmit` and `Stop`.
Persistent Claude Code MCP config health	`actionable`	#1021 / #1235	Retired for the current local operator host by #1235: the stale path was repaired to the portable `aippocampus mcp` command, and the no-write persistent diagnostic reported `persistent_config_healthy` with the nested MCP probe `healthy`. Reopen only if a later diagnostic reports a broken persistent config or before claiming broad Claude host portability.	Before any cross-host, package-distribution, or real-hook-firing Claude Code claim.
CJK recall quality beyond the first public fixture	`actionable`	#1022	The 2026-06-09 #1054 public-safe CJK rerun retires the narrow default-path compact-CJK gap: production hybrid now reports 7/7 positive top-5 hits with 0 negative false positives after lightweight CJK query chunks entered the default local hybrid path. Broader Chinese recall quality still needs richer public and private-safe evidence.	Before widening Chinese recall claims beyond the expanded fixture.
Cognitive-load false positives and usefulness	`durable_non_goal`	#575 closed as diagnostic-only	The 2026-06-10 public behavior-trace feedback fixture retires the reviewed-feedback gap, and the 2026-06-14 default-path replay records one useful hint plus one memory-drag regression. The owner is narrowed to diagnostic-only behavior; any future default foreground weighting needs a new issue with stronger live/default evidence.	Before enabling any default host-timing or foreground weighting claim.
Journey default hint timing and public reproducibility	`actionable`	#310	The 2026-06-10 public time-sliced replay cohort retires the selected public-reproducibility gap for a replayable hint/no-hint, stale/resolved/wrong-route suppression, false-foreground, and no-future-leak contract. The broader caveat retires only when live/private reviewed host-timing or user-visible quality evidence shows Journey hints help without source-truth, persona, or future-leak overclaim.	Before enabling default foreground Journey hints or claiming private/live Journey quality.
Episode/Arc gappy-chain overclaim risk	`durable_non_goal`	#663 closed as bounded owner track; future live/default issue if needed	The 2026-06-10 public fixture retires the selected builder/reopen-projection overclaim gap for complete, gappy, wrong-order, single-point, and temporary-concern arcs, and the 2026-06-14 route-producer fixture adds richer adapters and route guidance. The remaining caveat is a durable boundary: do not treat Episode/Arc packets as current validity or source truth without reopen.	Before treating Episode/Arc packets as more than navigation/read-model context.
AMemGym official live-provider score	`external_dependency`	#1232 closed as provider-route-blocked after post-top-up diagnosis	The 2026-06-11 #1232 resume attempt preserves the caveat: partial live-provider outputs remain incomplete, provider cost is not extractable from official outputs, and the pinned OpenRouter Native condition is blocked because the required OpenAI-family routes fail provider-policy preflight even on a harmless prompt. Retire this only when a later dated AMemGym note records a complete fixed-arm run under a declared provider/model condition, sanitized cost/latency or an explicit unavailable-provider-field reason, and parity-arm decision without promoting protocol smoke, bounded subsets, checkpoint output, provider-budget preflight, or route-preflight output into live score claims.	When a complete bounded/resumable fixed-arm run is produced and reviewed under a declared provider/model condition.
Private text disclosure in public evidence	`durable_non_goal`	-	Not retired by benchmark evidence; public reports stay aggregate, sanitized, or source-reopenable without raw private text.	Recheck only if the public/private evidence policy changes.
Hosted or cloud continuity from a single local-host smoke	`external_dependency`	-	Requires a scoped cloud/sync/provider evidence issue before it becomes actionable; local host proof cannot retire hosted or cross-device caveats.	Before any hosted/cloud product claim.
Broad private-history quality from selected local diagnostics	`research_blocked`	-	Requires a separately scoped private-history quality protocol with privacy-preserving review and explicit cohort limits; selected aggregates do not retire it.	Before any broad real-history quality claim.

Current Claim Snapshot

Each block below is one metric_id entry with current_value, run_date, source_report, claim_level, cohort, supersedes / superseded_by, and cannot_claim fields. Keep this block format line-addressable; do not turn it back into giant Markdown table rows.

`registry.local_real_history_aggregate`

current_value: 964 clean-source/index/graph-backed threads; 110 scope-labeled threads; 88 non-technical life-wide threads; 244 semantic sidecar rows across 46 threads; all eight canonical labels observed.
run_date: 2026-05-30
source_report: public-readiness-verification.md, #55 Stage 2 evidence
claim_level: first_pass_real_history_slice
cohort: Local real-history registry aggregate; aggregate-only smoke output.
supersedes / superseded_by: Supersedes the older 949-thread aggregate paragraph for public currentness.
cannot_claim: Full-history refresh, semantic completeness, label correctness without clean-source review, or private-text disclosure.

`semantic_sidecar.aggregate_materialized_rows`

current_value: 244 semantic sidecar rows across 46 threads, with all eight canonical labels observed.
run_date: 2026-05-30
source_report: public-readiness-verification.md, #55 Stage 2 evidence
claim_level: first_pass_real_history_slice
cohort: Local real-history dynamic semantic sidecar observation.
supersedes / superseded_by: Supersedes the older 2-thread / 5-row strict-survival number for aggregate materialized coverage only.
cannot_claim: Global semantic correctness, human review, complete life-wide labeling, or relaxed materializer gates.

`semantic_sidecar.strict_survival_snapshot`

current_value: Historical strict-survival slice: 5 rows across 2 real clean-source threads and 5 semantic latest timeline turns.
run_date: 2026-05-29
source_report: public-readiness-verification.md, earlier Stage 2 closeout and memory-decision-benchmark-plan.md
claim_level: historical_strict_survival_snapshot
cohort: Strict per-label evidence gate after source-review tightening.
supersedes / superseded_by: Superseded by semantic_sidecar.aggregate_materialized_rows for aggregate coverage; retained as a stricter survival baseline.
cannot_claim: Latest aggregate coverage, global Stage 2 correctness, or proof that suppressed labels are safe to restore.

`semantic_sidecar.source_review_green_gate`

current_value: 24 selected semantic sidecar label cases reviewed; 24/24 passed; pass_rate=1.0; no live model failures.
run_date: 2026-05-30
source_report: public-readiness-verification.md, #55 Stage 2 evidence
claim_level: selected_source_review_green_gate
cohort: Selected strict semantic sidecar labels reviewed through the live DeepSeek-compatible source-review smoke.
supersedes / superseded_by: Supersedes the older 5-case strict-review pass as the selected green gate.
cannot_claim: Human review, broad correctness, or full-history semantic quality.

`semantic_sidecar.source_review_diagnostic`

current_value: 96 selected cases reviewed; 89 passed; pass_rate=0.9271; failed_label_categories=[]; failure_count=0; failure_taxonomy.by_class={}; label_failure_taxonomy.by_class={"unsupported_label_evidence":7}; preference passed 6/9 with pass_rate=0.6667.
run_date: 2026-06-09
source_report: public-readiness-verification.md, 2026-06-09 #1053 rerun
claim_level: broader_selected_source_review_diagnostic
cohort: Broader selected source-review smoke with live provider behavior.
supersedes / superseded_by: Supersedes the same-day #993 96-case diagnostic where preference was below the per-label floor; the 24-case row remains the named green gate.
cannot_claim: Human review, full-history semantic correctness, release gate, selected source-review green gate, provider-independent quality, or private-text disclosure.

`question_tracking.six_axis_threshold_fixture_2026_06_09`

current_value: Selected public-safe fixture calibration: 6 scenarios passed; fixed low similarity threshold (0.52) had false merge 1/3 = 0.333333; static strong threshold (0.80) had false split 1/2 = 0.5; dynamic six-axis threshold had false merge 0/3 and false split 0/2; borderline-auto correct count 1 and wrong count 0; source-ref coverage 12/12 = 1.0.
run_date: 2026-06-09
source_report: benchmarks/aippocampus/benchmark_question_tracking_calibration.py --json
claim_level: selected_fixture_navigation_policy_calibration
cohort: Public-safe selected question-candidate fixtures over recurring context, low-information, orientation shift, over-merge conflict, over-split same-frontier, and borderline-confirmation families; no private text and no live model.
supersedes / superseded_by: Extends the earlier static-vs-adaptive non-regression row with #1059 false-merge / false-split / reason-code audit metrics.
cannot_claim: Source truth from dynamic thresholds, real-user calibration, private-history threshold quality, live external model confirmation, user-visible recall improvement, or broad question-tracking quality.

`fresh_thread.public_validation_2026_06_10`

current_value: Public-safe #281 fresh-thread validation: 6 positive public flows, 5 negative controls, first-turn positive route success 6, first-turn false activations 0, first_turn_scent_precision=1.0, progressive_activation_gain=1.0, source_reopen_before_specific_claim_rate=1.0, irrelevant_memory_drag_rate=0.0, overpersonalization_count=0, manual_query_invention_count=0, manual_query_expected_count=0, ready_lock_use_count=5, unsupported_evidence_count=0, and negative_control_active_recall_count=0.
run_date: 2026-06-10
source_report: fresh-thread-expanded-coverage-2026-06-03.md, GitHub #281 public validation readout and python benchmarks\aippocampus\benchmark_fresh_thread_recall_demo.py --json
claim_level: public_safe_fixture_validation
cohort: Checked-in public synthetic fresh-thread flows over positive, negative, multi-turn, correction, threshold, active-recall, source-reopen, and manual-query-invention controls; no private text, live model, raw source snippets, or local paths.
supersedes / superseded_by: Retires #281's public-fixture validation blocker while preserving live/private/external-dataset follow-ups as new scoped issues.
cannot_claim: Live fresh-thread quality, private real-history fresh-thread quality, universal fresh-thread recall quality, foreground-hook-only sufficiency, base-model innate memory, live semantic-model quality, or source truth without reopen.

`question_tracking.public_shadow_question_aware_reopen_2026_06_10`

current_value: Public/source-replayable #248 shadow fixture: 4 public cases, 2 negative controls, source-ref fidelity 1.0, preregistered selected public baseline/cohort, question-aware over question-blind term delta 0.75, answer-usefulness delta 1.0, manual-query-reduction delta 1.5, question-aware wrong-hint rate 0.0, no-question retrieval recall 0.5 vs question-aware retrieval recall 1.0, stale-question carryover 0, missed resurfacing without question tracking 2, wrong-route drag 0, noise false positives 0, materialization-review status public_shadow_review_evidence_ready, public-safe local calibration status public_safe_local_calibration_ready, and dynamic six-axis false merge 0/3 and false split 0/2.
run_date: 2026-06-10
source_report: question-aware-public-shadow-2026-06-10.md and python benchmarks\aippocampus\benchmark_question_aware_real_history.py --public-shadow --json
claim_level: public_replayable_shadow_fixture
cohort: Checked-in public-safe VCS-style and agent-trajectory question-continuity cases plus multilingual/noise and code negative controls; no private text, live model, or default prefilter.
supersedes / superseded_by: Retires #1367, #1368, and #1369 as bounded public-safe #248 slices and makes #248 closeable as a public-safe owner closeout while preserving private/live/default boundaries.
cannot_claim: Private real-history answer quality, live user-visible recall improvement, broad question-tracking quality, theme-resonance calibration, default prefilter adoption, source truth from question/theme rows, or broad no-question retrieval quality beyond the checked-in fixture.

`usefulness.natural_handoff_synthetic_validation_2026_06_14`

current_value: Public/synthetic #1384 validation: 6 cases, wins 2, no-help 1, regressions 3; natural handoff success rate 0.3333, default-session continuity success rate 0.5, implicit-context activation success rate 1.0. Regressions cover safe-but-noisy protocol dumps, safe route demoted to scent, and wrong-route/manual-search drag.
run_date: 2026-06-14
source_report: natural-handoff-usefulness-2026-06-14.md and python benchmarks\aippocampus\benchmark_natural_handoff_usefulness.py --json
claim_level: bounded_synthetic_default_session_validation
cohort: Checked-in public-safe synthetic natural-handoff/default-session cases using the canonical continuity_usefulness gate; no private text, provider calls, local paths, or live hook behavior.
supersedes / superseded_by: Closes #1384 and makes #1185 closeable as a bounded usefulness-owner closeout, while preserving future broad default-path/live-host usefulness as new scoped work.
cannot_claim: Broad default-session product lift, private-history user-visible lift, live host hook activation rate, or a rule that every deictic prompt should trigger recall without clarification.

`subconscious.event_salience_intake_fixture_2026_06_09`

current_value: Deterministic intake fixture: 7 clean-source-like turns classified into correction, frontier/blocker, preference/style, supersession/currentness, failed command/test, scope boundary, and low-information noise; 6/7 selected for candidate extraction, 1/7 skipped as low-information noise, missed-high-signal count 0, source-ref coverage 7/7 = 1.0. Opt-in job integration filters model payloads and writes a rebuildable salience sidecar; public projection omits raw source text and refs.
run_date: 2026-06-09
source_report: python -m unittest tests.aippocampus.test_subconscious_event_salience_gate -v
claim_level: deterministic_intake_contract
cohort: Public-safe synthetic clean-source-like turn fixtures and mocked job-runner integration; no private text and no live model.
supersedes / superseded_by: First #1058 event-salience intake gate fixture row.
cannot_claim: Memory truth from salience tags, foreground-hook model work, deletion of low-salience source, real-user salience quality, review-priority quality, private-history candidate-reduction quality, or user-visible recall improvement.

`retrieval_reconsolidation.review_window_fixture_2026_06_10`

current_value: Deterministic retrieval lifecycle fixtures project source-backed superseded, refuted, conflicted, and still_current outcomes into staging review candidates; no-write reports count activated, used, ignored, conflicted, superseded, refuted, still-current, and source-missing-blocked rows; source-key-only rows do not produce review candidates.
run_date: 2026-06-10
source_report: python -m pytest tests\aippocampus\test_retrieval_lifecycle.py -q
claim_level: deterministic_review_window_contract
cohort: Public-safe synthetic retrieval lifecycle rows over prompt/active/MCP/source-reopen style events; no private text, live model, or default hook write.
supersedes / superseded_by: Extends #1019 retrieval lifecycle diagnostics with #1081 retrieval-triggered reconsolidation review-window candidates.
cannot_claim: Default foreground hook writes, automatic memory updates, clean-source mutation, raw-rollout mutation, live semantic adjudication quality, #1082 state-dependent preactivation quality, private-history behavior lift, or user-visible recall improvement.

`warm_ambient.state_dependent_preactivation_2026_06_10`

current_value: Public-safe deterministic #1082 fixture: 4 cases, 7 candidates; state-dependent arm keeps preactivation hit rate 1.0, lowers false preactivation rate from simple-baseline 0.4 to 0.0, keeps source-reopen success rate 1.0, improves foreground-noise suppression from 0.6 to 1.0, and reduces source-reopen attempt cost units from 4 to 2 without model calls. Stale, privacy-blocked, and conflicted candidates stay suppressed, and predicted handles use only direction_only or reopenable_route.
run_date: 2026-06-10
source_report: state-dependent-preactivation-2026-06-10.md and python -m pytest tests\aippocampus\test_prewarm_planner.py tests\aippocampus\test_benchmark_state_dependent_preactivation.py -q
claim_level: public_safe_deterministic_preactivation_fixture
cohort: Selected public synthetic warm ambient cases over phase context, frontier proximity, salience, cache state, active recall locks, stale/conflicted routes, and privacy suppression; no private text, no live model, no foreground hook write.
supersedes / superseded_by: Closes the #1082 evaluation slice after #1081 supplied the retrieval-reconsolidation vocabulary; complements warm ambient recall and route-readiness diagnostics.
cannot_claim: Live foreground preactivation rollout, source truth from predicted routes, live latency savings, ADHD productivity lift, broad proactive-agent behavior, private-history quality, or user-visible recall improvement.

`e2e50.public_safe_behavior_pack_contract_2026_06_12`

current_value: Public-safe synthetic E2E50 behavior pack: 50 annotated cases; contract_gate_ok=true; quality_gate_ok=false; 50/50 correct; claim_level=public_safe_behavior_pack_contract; covers binding constraint survival, rejected-route avoidance, transient-concern extinction, superseded currentness, scope-limited constraints, summary-overhang traps, benign no-remember negatives, and source reopen before risky action; no_remember_negative_precision=1.0; annotation status counts gold_seed=23, calibration_seed=11, negative_control=3, rejected_candidate=4, duplicate_candidate=5, source_visible_candidate=4; source-family counts synthetic_public_safe=30, public_vcs=11, public_longitudinal=9.
run_date: 2026-06-12
source_report: python benchmarks/aippocampus/benchmark_e2e50_silent_constraint.py --json
claim_level: public_safe_behavior_pack_contract
cohort: Public-safe synthetic behavior-pack contract; no private text, no live host, no semantic judge, no raw behavior trace in reports.
supersedes / superseded_by: Supersedes the 2026-06-10 20-case public behavior-pack row for public case count and current coverage; keeps private retained-case scarcity as optional dogfood/private-quality evidence.
cannot_claim: E2E50 behavior benchmark quality, private real-history behavior lift, representative E2E50 sample quality, live host behavior lift, semantic-judge quality, or public-dialogue continuity proof.

`e2e50.label_oracle_live_diagnostic_2026_06_13`

current_value: Public-safe #1322 live-model E2E50 diagnostic: 50 public cases were run across a labeled baseline prompt and AIppocampus packet arm for 100 DeepSeek V4-Flash calls. The historical run recorded packet correct_rate=1.0 and baseline 0.94, but follow-up audit found the baseline prompt exposed case_family, family-specific scenario text, the action-code glossary, and an AIppocampus packet shell. Treat the high baseline and assisted_correct_rate_lift=0.06 as label-oracle / runner-wiring output, not behavior-lift evidence. Temperature was not sent under provider/default thinking.
run_date: 2026-06-13
source_report: e2e50-live-behavior-pilot-2026-06-13.md, e2e50-live-behavior-pilot-2026-06-13.json, and python -m unittest tests.aippocampus.test_benchmark_e2e50_behavior_live -v
claim_level: public_safe_label_oracle_diagnostic
cohort: Public-safe checked-in E2E50 fixture only, sanitized model excerpts, aggregate usage/cost estimates, and no raw provider payloads, private history, local paths, credentials, live host state, or raw behavior trace rows.
supersedes / superseded_by: Does not close #1322; it preserves the live runner/report path while marking that a blind public surface-task fixture is still required before baseline-vs-AIppocampus behavior validation.
cannot_claim: Clean no-memory baseline quality, AIppocampus-assisted behavior lift, #1322 behavior-validation closeout, broad E2E50 benchmark quality, private-history behavior lift, live host behavior lift, default foreground packet adoption, provider-general behavior quality, or source truth from packet summaries.

`e2e50.blind_surface_live_behavior_2026_06_13`

current_value: Corrected public-safe #1322 live-model E2E50 behavior run: 50 public cases across blind baseline and AIppocampus packet arms for 100 DeepSeek V4-Flash calls. Baseline prompt hides case_family, expected behavior codes, source hashes, family-specific scenario labels, action-code glossary, and empty packet shell. Packet arm scores correct_rate=1.0 and useful_next_action_rate=1.0; baseline scores 0.42 on both; assisted_correct_rate_lift=0.58; baseline has 10 manual-search choices, 21 source-reopen choices, and 29 safe-but-non-answer choices; packet arm has 0 manual-search and 0 safe non-answer choices; wrong actions, over-constrained actions, invalid actions, and private/sensitive model-output hits are 0; negative-control correct rate is 1.0 in both arms. Temperature is not sent under provider/default thinking.
run_date: 2026-06-13
source_report: e2e50-blind-surface-live-behavior-2026-06-13.md, e2e50-blind-surface-live-behavior-2026-06-13.json, and python -m unittest tests.aippocampus.test_benchmark_e2e50_behavior_live -v
claim_level: public_safe_blind_surface_live_behavior_pilot
cohort: Public-safe checked-in E2E50 fixture rendered into blind surface tasks, sanitized model excerpts, aggregate usage/cost estimates, and no raw provider payloads, private history, local paths, credentials, live host state, source hashes in prompts, or raw behavior trace rows.
supersedes / superseded_by: Closes the #1322 small public behavior-validation slice by scoring generated next-action choices under a blind baseline and packet-assisted arm while keeping deterministic contract, label-oracle diagnostic, and model-backed behavior evidence separate.
cannot_claim: Broad E2E50 benchmark quality, private-history behavior lift, live host behavior lift, default foreground packet adoption, provider-general behavior quality, source truth from packet summaries, or representative product quality beyond this synthetic public surface.

`e2e50.private_local_seed_annotation_blocker_2026_06_10`

current_value: Private/local E2E50 follow-up: wide scan found 23 candidate seeds against a 20-candidate minimum; local annotation summary reviewed 17 candidates and retained 7 control/seed cases, including 4 gold, 2 calibration, and 1 negative control; retained/control shortfall is 13 against the 20-case private target.
run_date: 2026-06-10
source_report: e2e50-private-local-seed-followup-2026-06-10.md and python tools/aippocampus/smoke/smoke_e2e50_seed_candidates.py --annotation <ignored-local-manual-annotation-json> --json
claim_level: private_local_sanitized_blocker_report
cohort: Local private clean-source annotation summary plus sanitized scanner aggregate; no private text, raw source refs, thread ids, message ids, local paths, or candidate rows checked in.
supersedes / superseded_by: Narrows the 2026-06-04 17/20 candidate-count blocker: candidate discovery is sufficient under the wide scan, but annotation retained/control count is still insufficient for private-history E2E50 quality. The public behavior pack is now the primary public path.
cannot_claim: Completed private-history 20-case or 50-case E2E50 quality, private real-history behavior lift, representative E2E50 sample quality, live host behavior lift, or semantic-judge quality.

`track_b.private_semantic_sidecar_required`

current_value: 100 selected private real-history cases; 100/100 top-5 hits; 1.0 hit rate; failed_count=0; sanitized miss taxonomy empty after the #963 dynamic-source ranking repair.
run_date: 2026-06-09
source_report: public-readiness-verification.md, 2026-06-09 #963 rerun
claim_level: private_bounded_track_b_slice
cohort: Maintainer-only private real-history semantic-sidecar-required source-evidence slice.
supersedes / superseded_by: Supersedes the 2026-05-29 97/100 selected private Track B row; the pre-repair same-day rerun exposed 5 top-k pruning misses, all with the gold source present in the raw candidate pool.
cannot_claim: Public benchmark score, real-user gate quality, full semantic completeness, live semantic-model quality, or private-text disclosure.

`retrieval_score_fusion.public_calibration_2026_06_10`

current_value: Public-safe synthetic calibration: 5 cases; expected top match 5/5; semantic bridge lift 1; wrong-stance lure suppressed 1; vector-disabled fallback 1; source-join gate reject 1; exact-text guard preserved 1; raw source refs, candidate text, absolute paths, and private user data are not serialized.
run_date: 2026-06-10
source_report: python -m pytest tests\aippocampus\test_retrieval_score_fusion.py -q
claim_level: public_safe_deterministic_fixture
cohort: Synthetic #309 score-policy cases over score_fusion.blend() and build_public_score_fusion_calibration_report().
supersedes / superseded_by: Adds a public score-fusion calibration row next to the existing candidate-funnel and source-evidence/reranker diagnostics.
cannot_claim: Default vector-prefilter adoption, local embedding adapter quality, live answer quality, broad reranker safety, latency/cost quality, private-history generality, or score output as source truth.

`source_joined_routing.decision_2026_06_14`

current_value: Public-safe #1370/#1372/#309 consumer decision: selected recall_context -> recall_deepen plus foreground packet source reopen as the measured consumer; 5 fixture cases; direct source-backed success 1.0 with average manual query invention 1.4; progressive source-reopen follow-through 1.0 with one stale-handle fail-closed case; source-ref rejoin 1.0; semantic bridge lift count 2; sentinel false-positive rate 0.0; wrong-route drag from sentinel 0; wrong-stance collision 0; source-join gate rejects 1; vector-disabled fallback 1; provider, foreground embedding, and external model calls 0.
run_date: 2026-06-14
source_report: source-joined-routing-decision-2026-06-14.md and python -m pytest tests\aippocampus\test_source_joined_routing_decision.py -q
claim_level: public_safe_consumer_decision_report
cohort: Deterministic public fixture over the progressive recall consumer plus post-source-join score-fusion calibration. No private data, live model calls, raw source text, raw source refs, credential environment variable names, provider payloads, local paths, or source-free vector hits.
supersedes / superseded_by: Closes #1370 and #1372, and closes #309 as a decision issue: keep text-first source-joined defaults, allow score fusion only after source join, and defer vector prefilter / local embedding default adoption. Future vector, LLM expansion, or graph work should be new narrow product-gap issues with replayable public consumer evidence.
cannot_claim: Live answer-quality lift, private-history generalization, default vector-prefilter safety, local embedding adapter quality, universal semantic or graph retrieval quality, or score output as source truth.

`rollout_hard_event.route_chain_topk_calibration_2026_06_12`

current_value: Public-safe rollout behavior route-chain calibration: top-k 1 fails because it recovers 0/3 required two-source chains and records 3 source-support failures; top-k 2 passes with 3/3 recall, 3/3 precision, 3/3 chain recovery, wrong-source evidence 0, stale-source top-k rate 0, and foreground action false positives 0; top-k 3 preserves main recall/precision but admits 2/3 stale or narrative decoys into the top source set.
run_date: 2026-06-12
source_report: rollout-hard-event-route-chain-2026-06-12.md
claim_level: public_safe_synthetic_route_chain_diagnostic
cohort: Checked-in CC0 synthetic rollout behavior fixture; deterministic production-like retrieval; no private history, live model, raw source text, raw rollout text, provider call, or local absolute paths in committed outputs.
supersedes / superseded_by: Adds a route-chain/actionability top-k boundary next to the #309 score-fusion and React VCS source-disambiguation rows: the useful budget is enough to complete a behavior-backed chain, not enough to foreground narrative decoys.
cannot_claim: Representative #1197 public-quality cohort, #1195 benchmark promotion, live agent quality, wild VCS corpus quality, private real-history continuity quality, source truth from assistant narrative, or external-system superiority.

`rollout_hard_event.public_cohort_v2_2026_06_12`

current_value: Public-safe #1197 rollout hard-event cohort V2: 17 synthetic agent-behavior projects, 34 future events, 17 flag-worthy hard events, 17 anti-drift negatives; production-like top-k2 retrieval recovered 17/17 two-source chains with recall 1.0, precision 1.0, source-support failures 0, wrong-source evidence rate 0.0, stale-source top-k rate 0.0, foreground action false positives 0, anti-drift violations 0, and current-vs-stale pairwise wins 34/34.
run_date: 2026-06-12
source_report: rollout-hard-event-cohort-v2-2026-06-12.md
claim_level: public_safe_synthetic_hard_event_cohort
cohort: Checked-in CC0 synthetic rollout behavior cohort; deterministic production-like retrieval; no private history, live model, raw source text, raw rollout text, provider call, or local absolute paths in committed outputs.
supersedes / superseded_by: Adds the first broader public-safe hard-event pack for temporal override, cross-scope drift, cross-project contamination, post-compaction gaps, forget boundaries, Dream candidate boundaries, route-topic specificity, host-surface readiness, privacy redlines, observability, latency/cost, and tool-scope failures.
cannot_claim: Live agent quality, wild public VCS corpus quality, private real-history continuity quality, #1195 benchmark-family promotion by itself, external-system superiority, or source truth from assistant narrative.

`benchmark_family.promotion_candidates_2026_06_20`

current_value: #1195 public-safe family-promotion decision: attention navigation remains promoted for the explicit-pull public/holdout surface, and selected agent continuity loop plus map-rot lifecycle debt now include observed public/holdout cohort measurements. Agent continuity measures 180 public-safe cases with 45 holdout; map-rot measures 270 public-safe cases with 68 holdout. For both selected families, required usefulness blocker counts/rates are 0, holdout_used_for_tuning_count=0, and unresolved next_measurement_actions=[].
run_date: 2026-06-20
source_report: benchmark-family-promotion-candidates-2026-06-20.md, benchmark-family-promotion-candidates-2026-06-20.json, and python benchmarks\aippocampus\benchmark_family_promotion_candidates.py --json
claim_level: public_safe_family_promotion_public_cohort_report
cohort: Public-safe generated cohort reports plus retained deterministic contract surfaces; no private history, live model, raw text, raw source refs, local paths, provider output, or cleanup writes in committed evidence.
supersedes / superseded_by: Supersedes the 2026-06-12 target-only candidate report by embedding observed public/holdout measurements and treating #1969/#1948 as historical closed owners rather than active unresolved owner issues.
cannot_claim: Live/private behavior lift, answer-generation quality, E2E50 closeout, cleanup-write runtime adoption, or external-system superiority.

`fts5.real_history_recall_2026_05_29`

current_value: Post-repair 100 selected source-backed cases; FTS5 91/100 top-1, 100/100 top-5, 100/100 top-10; production lexical-structural hybrid 100/100 top-10.
run_date: 2026-05-29
source_report: public-readiness-verification.md, FTS5 real-history recall benchmark and memory-decision-benchmark-plan.md
claim_level: bounded_real_history_regression_smoke
cohort: Local 949-thread real-history registry slice after stale SQLite index repair.
supersedes / superseded_by: Superseded only by a newer dated FTS5 real-history run.
cannot_claim: Natural-language user-query quality, private text disclosure, broad product recall quality, or dense vector retrieval as a default path.

`cjk.local_recall_public_fixture_2026_06_09_default_chunks`

current_value: Public-safe expanded CJK fixture after #1054: 10 synthetic cases; production lexical-structural hybrid with default lightweight CJK query chunks hit 7/7 positive cases at top-5 with 0 negative false positives; explicit cjk_aware_sidecar comparison also hit 7/7 positives; FTS5 trigram alone hit 3/7 positives.
run_date: 2026-06-09
source_report: cjk-local-recall-fixture-report.md
claim_level: public_safe_deterministic_fixture
cohort: Synthetic public exact/short/mixed/deictic/paraphrase/compact no-space/project-symbol/negative CJK local-recall cases.
supersedes / superseded_by: Supersedes cjk.local_recall_public_fixture_2026_06_09 for default-path behavior; the earlier same-day row remains historical gap evidence in the CJK report.
cannot_claim: Broad Chinese recall quality, semantic Chinese search from trigram alone, production hybrid handling every compact CJK cue beyond this fixture, private-history CJK quality, heavyweight tokenizer need, or dense vector retrieval as a default path.

`longmemeval_s.retrieval_only_500_2026_06_10`

current_value: LongMemEval-S cleaned V1 retrieval-only larger slice: 500 questions; session R@10 479/500 = 0.9580; evidence-line R@10 408/479 = 0.8518; context-visible evidence R@10 452/479 = 0.9436; session MRR 0.8809, evidence-line MRR 0.6309, context-visible evidence MRR 0.8086; warning_count 0. The exact-line taxonomy records evidence-line R@1/3/5/20/50 and classifies the 71 R@10 misses without raw text.
run_date: 2026-06-10
source_report: longmemeval.md, current published result
claim_level: public_external_retrieval_only_slice
cohort: LongMemEval-S cleaned V1 first 500 questions; deterministic retrieval-only adapter; top-k 10; evidence context radius 5; progress / partial-output path completed; no QA generation or judge model.
supersedes / superseded_by: Supersedes the 2026-06-09 100-question LongMemEval-S row for current sample size and retires the incomplete 500-question missing-artifact blocker; oracle 50 remains a smoke row.
cannot_claim: Answer-generation quality, LongMemEval QA score, judge-model score, LongMemEval-V2 score, SOTA or external baseline superiority, decision-gate quality, broad memory superiority, or exact-line citation quality being solved.

`longmemeval_s.lexical_line_reranker_500_2026_06_10`

current_value: Optional local lexical exact-line reranker diagnostic on the same 500 LongMemEval-S questions: first-stage evidence-line R@10 remains 408/479 = 0.8518; fused reranked evidence-line R@10 is 419/479 = 0.8747; fused evidence-line MRR improves from 0.6309 to 0.6746; source-joined bridge lifts 11, including 10 context-visible exact-line misses and 1 same-session wrong-line top-k miss; reranker error count 0.
run_date: 2026-06-10
source_report: longmemeval.md, optional lexical line-reranker follow-up
claim_level: public_external_retrieval_reranker_diagnostic
cohort: LongMemEval-S cleaned V1 first 500 questions; deterministic retrieval-only adapter plus explicit --line-reranker lexical; top-k 10; evidence context radius 5; no external model, no answer labels, no expected-line leakage.
supersedes / superseded_by: Adds a #1087 exact-line ranking follow-up next to the default retrieval-only 500-question row; does not supersede the default-off retrieval baseline.
cannot_claim: Default exact-line citation quality, answer-generation quality, LongMemEval QA score, judge-model score, LongMemEval-V2 score, broad reranker safety, semantic understanding, SOTA or external baseline superiority, decision-gate quality, or broad memory superiority.

`longmemeval_s.structural_exact_line_failure_500_2026_06_12`

current_value: #1193 deterministic structural exact-line repair failure report: structural fused evidence-line R@10 is 416/479 = 0.8685 versus lexical 419/479 = 0.8747 on the same split; structural evidence-line MRR is 0.6663 versus lexical 0.6746; context-visible conversions are 8 versus lexical 10; same-session wrong-line reductions are 8 versus lexical 11. The report identifies 36 context-visible candidate-visible misses that still need semantic/model line selection, plus source-window-not-visible miss families that need broader routing or source-side semantic support.
run_date: 2026-06-12
source_report: longmemeval-exact-line-repair-2026-06-12.md and longmemeval-exact-line-repair-2026-06-12.json
claim_level: public_external_retrieval_reranker_failure_report
cohort: LongMemEval-S cleaned V1 first 500 questions; deterministic retrieval-only adapter plus explicit --line-reranker structural; top-k 10; evidence context radius 5; no external model in the structural arm, no answer labels, no expected-line leakage, and no raw text/source/local paths in committed reports.
supersedes / superseded_by: Closes #1193 as a measured deterministic failure boundary and separates cold online semantic rerank, warm query/candidate cache, and source-side semantic cache paths before further spend.
cannot_claim: Improved-over-lexical exact-line quality, default structural reranker adoption, default exact-line citation quality, 500Q semantic reranker quality, source-side semantic cache latency/build cost, answer-generation quality, official LongMemEval QA score, judge-model score, provider-independent quality, SOTA, or broad memory superiority.

`longmemeval_s.semantic_warm_query_cache_25_2026_06_12`

current_value: #1305 warm query/candidate cache replay over a fresh 25Q semantic pilot with candidate-pack hashes: 24/24 available cold-fill calls had complete cache keys; first pass recorded 24 cold-fill misses, second pass recorded 24 warm hits with hit rate 1.0000 and two-pass hit rate 0.5000; warm lookup latency averaged 0.000079ms; cold-fill latency averaged 6308.67ms with 223947 total tokens and provider prefix-cache hit rate 0.0000. Exact-line metrics on the narrow pilot were first-stage R@10 23/25, semantic-only R@10 24/25, fused reranked R@1 23/25, fused reranked R@3/R@5/R@10 25/25, MRR 0.9600, and top-10 regression count 0; context-visible exact-line misses 2/2 and the same two same-session wrong-line focus cases were recovered.
run_date: 2026-06-12
source_report: longmemeval-semantic-cache-path-2026-06-12.md, longmemeval-semantic-cache-path-2026-06-12.json, and longmemeval.md, warm query/candidate cache replay
claim_level: public_external_llm_reranker_warm_cache_replay
cohort: LongMemEval-S cleaned V1 first 25 questions, all longmemeval_single-session-user; explicit DeepSeek-compatible semantic line reranker; candidate-pack hash includes candidate line/routing metadata plus source-text hashes, while committed reports emit only aggregates and no raw question, answer, source text, cache keys, cache values, provider responses, credentials, or local paths.
supersedes / superseded_by: Completes #1305 path B by measuring warm query/candidate cache-key completeness and local replay lookup separately from cold provider latency and provider prefix-cache behavior. It does not supersede the 500Q retrieval, lexical, or structural rows, and it leaves source-side semantic warming as a separate unmeasured path.
cannot_claim: 500Q semantic reranker quality, source-side semantic cache build cost or hot-path latency, live hook latency, default semantic reranker adoption, provider-independent quality, answer-generation quality, official LongMemEval QA score, LongMemEval-V2 quality, SOTA, broad per-type quality, or broad memory superiority.

`longmemeval_s.semantic_query_cache_100_2026_06_12`

current_value: #1323 100Q semantic query/candidate cache progress: first 100 LongMemEval-S questions, 94 evidence-line cases; session R@10 97/100 = 0.9700; context-visible evidence R@10 91/94 = 0.9681; first-stage evidence-line R@10 82/94 = 0.8723, MRR 0.6481; same-cohort lexical fused R@10 84/94 = 0.8936, MRR 0.6990; semantic-only first run R@10 87/94 = 0.9255, MRR 0.8989; semantic fused first run R@10 91/94 = 0.9681, MRR 0.9246; semantic fused repeated workers=8 R@10 stayed 91/94 with MRR 0.9037. Query/candidate replay found 90/90 complete cache keys and 90/90 second-pass warm hits; warm local lookup averaged 0.000063ms; cold-fill provider latency averaged 11136.85ms with 979237 tokens. Provider prefix-cache telemetry rose from 0.2514 on the workers=2 first run to 0.9932 on the workers=8 repeated run, while wall time fell from 767.25s to 273.35s.
run_date: 2026-06-12
source_report: longmemeval-semantic-cache-100q-2026-06-12.md, longmemeval-semantic-cache-100q-2026-06-12.json, and longmemeval.md, 100Q semantic query/candidate cache progress
claim_level: public_external_llm_reranker_query_cache_100q_progress
cohort: LongMemEval-S cleaned V1 first 100 questions: 70 single-session-user and 30 multi-session cases; explicit DeepSeek-compatible semantic line reranker; same-cohort deterministic lexical comparison; top-k 10; provider budget and partial-output checkpoints declared; committed reports emit only aggregate metrics and short report hashes, not raw question, answer, source text, provider responses, cache keys, cache values, credentials, or absolute paths.
supersedes / superseded_by: Scales the #1305 query/candidate cache evidence from 25Q to 100Q and adds same-cohort lexical comparison for #1323. It is superseded as a same-cohort comparison point by the 500Q worker-surface proxy and LLM upper-bound row.
cannot_claim: Source-side semantic cache build cost or hot-path latency before the 2026-06-13 row, full 500Q semantic reranker quality before the 2026-06-13 row, live hook latency, default semantic reranker adoption, provider-independent quality, answer-generation quality, official LongMemEval QA score, LongMemEval-V2 quality, SOTA, broad per-type quality, or broad memory superiority.

`longmemeval_s.source_worker_surface_and_semantic_upper_bound_500_2026_06_13`

current_value: 500Q measurement separates two arms on the same LongMemEval-S first-500 cohort. Current source-worker-surface proxy: baseline evidence-line R@10 408/479 = 0.8518, baseline MRR 0.6309; source-worker search alone R@10 156/479 = 0.3257, MRR 0.1391; source-worker rerank only R@10 398/479 = 0.8309, MRR 0.5046; FTS-preserving fused R@10 422/479 = 0.8810, MRR 0.6734; bridge lifts 14; fused top-10 regression count 0; source-only regression count 24; line-reranker errors 0; provider calls/tokens 0. Proxy prewarm built 246738 aippocampus_working_memory navigation rows with complete rate 1.0; hot source-worker search averaged 1044.5179ms, while candidate rerank averaged 1.18ms. Separate LLM query/candidate upper bound: semantic fused R@10 451/479 = 0.9415, MRR 0.8738, bridge lifts 43, available calls 470/479, errors 9 including 7 timeouts under explicit 180s timeout, and 4719903 total provider tokens with provider prefix-cache hit rate 0.2290.
run_date: 2026-06-13
source_report: longmemeval-source-worker-surface-500q-2026-06-13.md, longmemeval-source-worker-surface-500q-2026-06-13.json, and longmemeval.md, 500Q source-worker-surface proxy and LLM upper bound
claim_level: public_external_source_worker_surface_proxy_and_llm_upper_bound_500q
cohort: LongMemEval-S cleaned V1 first 500 questions; proxy arm builds existing AIppocampus aippocampus_working_memory row shape from clean source and uses the existing matcher, with source reopen required for claims; it does not exercise semantic_scope_labeling, semantic_scope_builder, subconscious jobs, warm ambient routes, or attention-router handoff as the canonical source-side materializer. LLM arm is explicit query/candidate DeepSeek rerank. Reports and committed summaries emit aggregate metrics and report hashes, not raw question, answer, source text, cache values, provider responses, credentials, or absolute local paths.
supersedes / superseded_by: Measures the current worker-row proxy and the 500Q query/candidate LLM upper bound. It supersedes the 100Q progress row as a same-cohort comparison point while preserving that row as provider-prefix-cache comparison evidence.
cannot_claim: Official LongMemEval QA score, answer-generation quality, SOTA, provider-independent semantic quality, default foreground LLM rerank, canonical source-side semantic warming/materializer quality, broad life-history memory superiority, or a claim that current worker-surface quality equals the LLM upper bound.

`longmemeval_s.full_source_semantic_warming_500_2026_06_13`

current_value: #1323 contract-aware full-source semantic-scope warming over the same LongMemEval-S first-500 cohort: source-index cache hit 500/500 with 0 source-index rebuilds; materializer public_semantic_labeler_full_source sent 246738 clean-source candidate messages across 500 batches; provider cold-fill calls 500; sidecar materialization status was partial with 458 materialized cases and 42 labeler errors; reviewed source cache had 451 sidecar cases and 2745 semantic-scope label rows; hot-query provider calls stayed 0. Sidecar evidence coverage rose to 53/479, context coverage to 74/479, and query/sidecar term overlap to 260 cases / 502 total overlaps, but fused evidence-line R@10 stayed 422/479 = 0.8810 with MRR 0.6763. No-lift diagnosis: only 3/71 baseline misses had sidecar exact coverage; 50/53 sidecar exact evidence hits were already baseline top-10 hits; among the 37 fused misses where the gold line was candidate-visible, the gold line had 0/37 semantic-scope labels/terms/query overlaps. Provider telemetry reported 60710135 total tokens, 12548480 prompt-cache hit tokens, and 43575795 prompt-cache miss tokens; that miss count is provider prefix-cache telemetry for cold full-source prompts, not local AIppocampus source-index cache failure.
run_date: 2026-06-13
source_report: longmemeval-full-source-semantic-warming-500q-2026-06-13.md, longmemeval-full-source-semantic-warming-500q-2026-06-13.json, and longmemeval.md, current published result
claim_level: public_external_full_source_semantic_scope_warming_diagnostic
cohort: LongMemEval-S cleaned V1 first 500 questions; public source-side semantic-scope sidecar materialization with contract-aware sidecar manifests; no private history; no answer labels, raw question, answer, source text, provider responses, cache values, credentials, or absolute local paths in committed evidence. The hot retrieval/rerank path uses local source cache only after prewarming.
supersedes / superseded_by: Supersedes the 8-candidate/top-selector sidecar diagnostic as the fairer source-side semantic warming measurement and fixes the cache-contract boundary by requiring materializer manifests before sidecar reuse. It shows that the current full-source sidecar is still a sparse continuity-scope layer, not a dense factual retrieval surface; the next fix must attach source-side retrieval terms/summaries to hard factual rows before fusion can lift this benchmark.
cannot_claim: Official LongMemEval QA score, answer-generation quality, SOTA or external baseline superiority, provider-independent semantic quality, default foreground LLM rerank, full AIppocampus product ability, semantic sidecar values as source truth, or a claim that full-source warming improves fused R@10 on this adapter.

`longmemeval_s.source_factual_alias_500_2026_06_14`

current_value: 500Q source factual-alias closeout on the same LongMemEval-S first-500 cohort: baseline session R@10 479/500 = 0.9580; baseline evidence-line R@10 408/479 = 0.8518; context-visible evidence R@10 452/479 = 0.9436; fused source-semantic-cache evidence R@10 422/479 = 0.8810; fused MRR 0.6744; candidate evidence coverage 463/479 = 0.9666, up from the earlier lexical 500Q diagnostic's 455/479 = 0.9499; remaining fused misses split into 15 candidate-missing and 42 reranker-visible misses. Factual-alias evidence coverage is 227/479 = 0.4739, factual-alias candidate coverage is 220/479 = 0.4593, gold candidate alias cases 220, gold candidate query-overlap cases 28, factual-alias candidate lift top-10 16, factual-alias fused lift top-10 2, and fused regression top-10 0. Cache metrics: 500 source caches, 246738 source rows/spans, 109144 factual-alias profiles, 738561 factual-alias terms, cache-key complete rate 1.0, hot-query provider calls 0, provider tokens 0, hot-query latency average 1092.5576ms and max 1926.4032ms.
run_date: 2026-06-14
source_report: longmemeval-source-factual-alias-500-2026-06-14.md, longmemeval-source-factual-alias-500-2026-06-14.json, and longmemeval.md, 500Q source factual-alias closeout
claim_level: public_external_source_factual_alias_closeout_500q
cohort: LongMemEval-S cleaned V1 first 500 questions; deterministic source-side factual-alias and answer-bearing source handles built from clean source; source-semantic-cache hot path; no private history, no answer labels, no raw question, answer, source text, provider responses, cache values, credentials, or absolute local paths in committed evidence. Aliases and working-memory rows are navigation-only and require source reopen before factual claims.
supersedes / superseded_by: Closes #1327 as a bounded source-window coverage diagnostic, closes #1424 as a source-side factual artifact/hot-path owner, and makes #1323 closeable as a measured source-side benchmark slice while preserving the full-source semantic-scope no-lift row as a real boundary. Future exact-line/reranker improvements should be new scoped issues, not a reason to keep these owner issues open.
cannot_claim: Answer-generation quality, official LongMemEval QA score, LongMemEval-V2 behavior, SOTA or external baseline superiority, broad life-history memory superiority, default foreground adoption, perfect exact-line citation quality, or source truth from aliases/working-memory rows/candidate routes without source reopen.

`longmemeval_s.post_factual_alias_rerank_closeout_500_2026_06_14`

current_value: #1437 post-factual-alias closeout on the same LongMemEval-S first-500 sanitized report: 500 questions; 479 evidence-line cases; reranked evidence-line R@10 422/479 = 0.8810; MRR 0.6744; candidate evidence coverage 463/479 = 0.9666; top-10 regressions 0; hot-query provider calls 0; hot-query latency avg/max 1092.5576ms / 1926.4032ms. The 57 fused misses split into 15 candidate-missing and 42 reranker-visible misses. Scoped post_factual_alias_exact_line_rerank_v1 default reranker change is rejected; bounded candidate coverage projection is accepted only as a future candidate-builder slice with candidate_coverage_lift=6 and candidate_byte_growth_ratio=0.0001; full_500_projection.decision is already_measured_local_hot_path because this path has no provider calls or tokens.
run_date: 2026-06-14
source_report: longmemeval-post-factual-alias-rerank-closeout-500-2026-06-14.md, longmemeval-post-factual-alias-rerank-closeout-500-2026-06-14.analysis.json, and python -m pytest tests\aippocampus\test_benchmark_longmemeval_rerank_analysis.py -q
claim_level: public_external_post_alias_reranker_closeout
cohort: Same sanitized LongMemEval-S first-500 factual-alias report; no private history, raw question, answer, source text, source refs, provider payloads, cache values, credentials, or absolute local paths in committed evidence. Candidate-builder projection is not an exact-line ranking result and aliases/routes require source reopen.
supersedes / superseded_by: Closes #1437 by analyzing the remaining 57 misses, explicitly rejecting a default exact-line reranker change from this evidence, and separating a bounded candidate-builder follow-up from source-truth or answer-quality claims.
cannot_claim: Perfect exact-line citation quality, default reranker adoption, answer-generation quality, official LongMemEval QA score, LongMemEval-V2 behavior, SOTA or external baseline superiority, broad memory superiority, provider-backed full-rerank budget need for this local closeout, or source truth from aliases/candidate routes without source reopen.

`longmemeval_s.semantic_llm_line_reranker_pilot_25_2026_06_10`

current_value: Optional external-model exact-line reranker pilot on the first 25 LongMemEval-S questions: first-stage evidence-line R@10 is 23/25 = 0.9200; semantic-only evidence-line R@10 is 24/25 = 0.9600; fused reranked evidence-line R@10 is 25/25 = 1.0000; fused evidence-line MRR improves from 0.6348 to 1.0000; reranked ladder R@1/R@3/R@5/R@10/R@20/R@50 is all 25/25 = 1.0000; context-visible rescue conversion 2/2; same-session wrong-line reduction 2/2; regression count 0; source-joined bridge lifts 2; reranker availability 24/25 with 1 timeout; total tokens 225170 and average available-call latency 7156.27ms. The 500-question projection is about 4503400 tokens and 59.64 single-worker minutes, with provider dollar cost not reported, so the full 500Q semantic run is explicit opt-in only.
run_date: 2026-06-10
source_report: longmemeval.md, optional semantic LLM line-reranker pilot and longmemeval-semantic-rerank-analysis-2026-06-10.json
claim_level: public_external_llm_reranker_pilot_budget_bounded
cohort: LongMemEval-S cleaned V1 first 25 questions, all longmemeval_single-session-user; retrieval-only adapter plus explicit --line-reranker semantic; DeepSeek-compatible deepseek-v4-flash; prompt version llm-window-to-line-rerank-v1; top-k 10; evidence context radius 5; external model saw question text and bounded candidate source text, but not gold answers, expected lines/sessions, has_answer labels, judge labels, miss taxonomy, or raw report cases.
supersedes / superseded_by: Adds the #1092 pilot analysis and full-500 budget/latency/privacy decision next to the default retrieval-only and lexical rows; it keeps the arm available but does not supersede the 500-question retrieval baseline or adopt semantic reranking as default.
cannot_claim: Default exact-line citation quality, 500-question LLM-reranker quality, answer-generation quality, LongMemEval QA score, judge-model score, LongMemEval-V2 score, provider-independent quality, broad reranker safety, per-type full-split quality, SOTA or external baseline superiority, decision-gate quality, or broad memory superiority.

`longmemeval_s.fixed_reader_answer_25_2026_06_12`

current_value: Provider-backed fixed-reader answer/latency baseline on 25 LongMemEval-S cleaned V1 cases: reader attempted 25/25; deterministic answer-correct count 20/25 = 0.8000; context sufficient 25/25; reader abstention count 0; retrieval/reference layer in the same run had session R@10 25/25, evidence-line R@10 23/25, context-visible evidence R@10 25/25, and warning_count 0; reader latency averaged 2723.84ms with max 6775.68ms; total elapsed 192.69s; token use 379965 total; run-configured estimated cost USD 0.106922; sanitized report validation passed.
run_date: 2026-06-12
source_report: longmemeval-fixed-reader-answer-25-2026-06-12.md and longmemeval.md, current published result
claim_level: public_external_fixed_reader_answer_latency_pilot
cohort: LongMemEval-S cleaned V1 first 25 questions, all longmemeval_single-session-user; DeepSeek-compatible deepseek-v4-flash; prompt version longmemeval-s-fixed-reader-v1; local deterministic overlap judge; top-k 10; evidence context radius 5; external reader saw public question text and bounded candidate source lines but not gold answers, expected lines/sessions, has-answer labels, retrieval miss taxonomy, judge labels, or raw report cases.
supersedes / superseded_by: Adds the first #1194 dated fixed-reader answer/latency row. It does not supersede the 500-question retrieval-only row; it adds a small answer-quality, latency, token, and cost baseline layered on the same source-evidence adapter.
cannot_claim: Official LongMemEval judge score, official leaderboard/SOTA, LongMemEval-V2 quality, default reader/provider adoption, model-independent memory superiority, private real-history quality, LoCoMo answer quality, broad per-type LongMemEval-S quality, or answer quality beyond this 25-case single-session-user pilot.

`longmemeval_s.fixed_reader_cleanup_25_2026_06_12`

current_value: Public-safe #1282 cleanup reran the same 25 LongMemEval-S cleaned V1 cases with prompt v2 and refined taxonomy: retrieval/reference remained session R@10 25/25, evidence-line R@10 23/25, and context-visible evidence R@10 25/25; deterministic answer-correct count was 19/25 = 0.7600; failure review found 3 reader/provider errors, 1 false abstention on answerable bounded evidence, 1 deterministic-judge mismatch, and 1 true reader miss; sanitized report validation passed; expansion gate status is no_go for 100Q or 500Q until blockers are fixed.
run_date: 2026-06-12
source_report: longmemeval-fixed-reader-cleanup-25-2026-06-12.md and longmemeval.md, current published result
claim_level: public_external_fixed_reader_failure_review_no_go
cohort: LongMemEval-S cleaned V1 first 25 questions, all longmemeval_single-session-user; DeepSeek-compatible deepseek-v4-flash; prompt version longmemeval-s-fixed-reader-v2; local deterministic overlap judge v2; top-k 10; evidence context radius 5; external reader saw public question text and bounded candidate source lines plus salience hints and bounded-answer policy, but not gold answers, expected lines/sessions, has-answer labels, retrieval miss taxonomy, judge labels, or raw report cases.
supersedes / superseded_by: Turns the #1194 answer baseline into an actionable #1282 failure-review and expansion-gate artifact; it explicitly blocks larger provider answer runs rather than treating the 25Q pilot as scale-ready.
cannot_claim: Improved answer score, official LongMemEval judge score, 100Q or 500Q answer quality, LongMemEval-V2 quality, default reader/provider adoption, model-independent memory superiority, private real-history quality, broad per-type LongMemEval-S quality, or proof that the fixed reader is ready to scale.

`demo_scenarios.claim_boundaries`

current_value: Public-safe demo scenarios show product shape; their material-limit lines are claim-boundary sources for demos.
run_date: 2026-06-03
source_report: docs/guides/demo-scenarios.md
claim_level: claim_boundary_source
cohort: Public example bundle, public-safe demo commands, and explicit live/smoke demo flows.
supersedes / superseded_by: Not a metric row; it routes demo caveats into evidence governance.
cannot_claim: Official benchmark proof, readiness metric upgrades, or private real-history performance.

`claude_code.real_host_dogfood_2026_06_11`

current_value: Local Claude Code history parser passed on 222 detected sessions; dry-run onboarding planned 223 registrations and 3 stale-index repairs without writing; synthetic cross-agent clean-source retrieval passed in both directions; temporary strict-config Claude Code MCP live call reached memory_health; #1235 then repaired the persistent local Claude MCP entry to aippocampus mcp, after which the no-write persistent diagnostic reported persistent_config_healthy, nested status healthy, tool count 10, and memory_health listed.
run_date: 2026-06-11
source_report: claude-code-dogfood-2026-06-09.md, claude-code-mcp.md, and #1235 persistent diagnostic closeout
claim_level: local_real_host_dogfood_with_persistent_config_repair
cohort: Single Windows operator host plus public-safe synthetic Codex/Claude fixture.
supersedes / superseded_by: Supersedes the 2026-06-09 persistent-config blocker for this local host only; preserves temporary strict-config MCP reachability as the original positive live-host proof and records the portable persistent config repair.
cannot_claim: Claude Code real-host hook firing beyond the scoped #1020 synthetic contract, cross-host persistent MCP health, unattended private-history ingestion, cross-device sync, hosted/cloud continuity, broad cross-host relationship continuity, package-manager availability on every machine, or private-history quality.

`claude_code.hooks_contract_2026_06_09`

current_value: Scoped Claude Code hook contract slice: status, dry-run, and smoke commands exist; UserPromptSubmit and Stop synthetic Claude-shaped hook inputs exit 0 without leaking raw prompt text, session ids, transcript paths, cwd values, settings paths, source refs, or synthetic tool payload text; PostToolUse/PostToolBatch and compaction hooks report event-level blockers.
run_date: 2026-06-09
source_report: claude-code-hooks-contract-2026-06-09.md
claim_level: scoped_synthetic_hook_contract
cohort: Public-safe synthetic Claude-shaped hook payloads plus official-contract intake.
supersedes / superseded_by: Narrows the broad Claude Code hook caveat into scoped UserPromptSubmit/Stop handler availability and explicit unsupported-event blockers.
cannot_claim: Real-host hook firing, Claude settings mutation, configuration-mutating installer availability, PostToolUse/PostToolBatch payload capture, compaction survival packet utility, all Claude Code versions, MCP health, transcript onboarding, or private-history quality.

`multimodal_niah.evidence_pool_conflict_resolution`

current_value: Public-safe NIAH evidence-pool contract: 4/4 answer/source-selection/source-anchor-citation after conflict repair; one stale input source selection preserved as input_selected_evidence_ids, current_source_selected_count=1, stale_or_conflicting_distractor_selection_count=0, needs_source_reopen_count=0; ambiguous-currentness regression control requires needs_source_reopen instead of guessing.
run_date: 2026-06-09
source_report: multimodal-niah-evidence-pool-report.md
claim_level: public_safe_deterministic_contract
cohort: Four public synthetic NIAH supplied-pool rows plus one in-test ambiguous-currentness negative control; retrieval is intentionally not scored.
supersedes / superseded_by: Supersedes the 2026-06-03 expected stale-source failure interpretation for the #533 NIAH supplied-pool contract.
cannot_claim: Retrieval quality, ATM-Bench Hard score/support, live vision-model answer quality, raw-media model quality, private media behavior, or broad multimodal-memory quality.

`recall_navigation.progressive_route_follow_through`

current_value: Progressive recall_context -> recall_deepen arm: route_actionability_rate=1.0, eligible source_reopen_follow_through_rate=1.0 (3/3 eligible route reopens reached expected clean-source refs), source_reopen_fail_closed_count=1 with failure_class=stale_handle_rejected_before_source_use, avg_manual_query_invention_count=0.0; foreground packet candidate ref follow-through reached the expected fixture source with foreground_manual_query_invention_count=0.
run_date: 2026-06-09
source_report: recall-navigation-comparison-2026-06-03.md
claim_level: public_safe_deterministic_proxy
cohort: Four public synthetic clean-source fixtures plus one foreground hook packet/cache fixture; all use temporary public-safe clean-source/index/cache data.
supersedes / superseded_by: Updates the #465 comparison with a narrow #201 route-follow-through readout; does not supersede live #201 field reports.
cannot_claim: Live user quality, broad default foreground first-turn or second-turn lift, production selector superiority, source-backed answer quality beyond the synthetic fixtures, or closing #201.

`recall_navigation.macro_prior_fixture_2026_06_12`

current_value: #1300 runtime/promotion slice: agent_continuity.recall consumes .aippocampus/macro-orientation.jsonl or .aippocampus/macro_orientation.jsonl from the supplied cwd when no explicit --macro-state-jsonl is passed; the promotion harness reports macro_navigation_readout.measured=true, active_layer_order_delta_count=1, hamming_fanout_delta_count=1, momentum_recheck_diagnostic_count=1, manual_search_fallback_reduction_count=1, stale_as_current_count=0, and claim_without_source_reopen_count=0; default_adoption_allowed=false.
run_date: 2026-06-12
source_report: recall-navigation-comparison-2026-06-03.md and python -m unittest tests.aippocampus.test_agent_opt_in_continuity tests.aippocampus.test_recall_navigation_promotion -v
claim_level: public_safe_default_path_and_promotion_fixture
cohort: Project-local default macro state file, synthetic public-safe route set, same-corpus/same-query/same-budget promotion arms, and no private history or external model calls.
supersedes / superseded_by: Adds the #1300 default-path/runtime slice and macro promotion readout without changing the #1185 default-session usefulness gate.
cannot_claim: Macro/router default-ready behavior, live host lift, private-history macro quality, source truth from macro state, answer-generation quality, broad route-selector superiority, or MCP-core default behavior.

`macro.transform_orbit_diagnostic_2026_06_12`

current_value: #1315 reversible transform-orbit diagnostic slice: cr_reversible uses only opposite and reverse transforms; the inventory reports 20 unique C/R orbits across 64 hexagrams with size distribution {2: 8, 4: 12}; line flips are emitted as adjacency; nuclear dynamics are emitted as non-invertible projection basins with four 16-source basins; packets keep authority_level=navigation_only, candidate_authority=candidate_only, claim_permission=no_claim_before_reopen, fact_claim_allowed=false, and default_ranking_effect=none.
run_date: 2026-06-12
source_report: yi-macro-runtime-interfaces.md and python -m unittest tests.aippocampus.test_yi_macro_runtime_interfaces -v
claim_level: public_safe_deterministic_macro_structure_fixture
cohort: Canonical public 64-hexagram table, deterministic structural transforms, no private history, no model calls, no raw source handles, and no foreground rendering.
supersedes / superseded_by: Adds the #1315 macro route-equivalence diagnostic substrate without promoting it into Dream routing, foreground packets, or source evidence.
cannot_claim: Dream integration quality, route merging, source support, default ranking or fanout changes, foreground symbolic prose, private-history behavior, Yi interpretation quality, or claim support without source reopen.

`macro.timing_recheck_experiment_2026_06_12`

current_value: #1314 public-safe timing experiment: four synthetic source-event cases cover active-axis route-success recency, stale/counter-evidence recheck pressure, normal source-epoch cadence, and slow-quiet below-threshold cadence; report metrics are case_count=4, distinct_signal_count=2, claim_without_source_reopen_count=0, currentness_mutation_count=0, default_adoption_allowed=false, and promotion_status=fixture_candidate_not_promoted.
run_date: 2026-06-12
source_report: yi-macro-runtime-interfaces.md and python -m unittest tests.aippocampus.test_macro_timing_recheck_experiment tests.aippocampus.test_yi_macro_runtime_interfaces -v
claim_level: public_safe_deterministic_macro_timing_fixture
cohort: Synthetic source-event rows only; source deltas are aggregate numeric fixtures; no private history, no raw source refs, no model calls, no live hooks, no calendar/solar-term schedule, and no foreground rendering.
supersedes / superseded_by: Adds a usable #1314 experiment boundary for active-axis and source-epoch recheck diagnostics while leaving currentness/temporal heads and live recall untouched.
cannot_claim: Live recall adoption, recall-quality lift, token/tool-cost savings, private-history timing quality, classical calendar timing, currentness/freshness truth, default ranking changes, scheduled background work, or claim support without source reopen.

`dream.shadow_route_topology_scout_2026_06_12`

current_value: #1313 public-safe Dream topology scout slice: built-in fixture reports dream_topology_candidate_count=5, shadow_route_candidate_count=2, shadow_route_source_overlap_count=2, transform_orbit_deepen_candidate_count=1, shadow_route_generic_vocab_false_positive_count=0, and shadow_route_claim_without_reopen_count=0; candidates keep candidate_authority=candidate_only, authority_level=navigation_only, action_grammar=direction_with_ref, source_reopen_required_before_claim=true, foreground_eligible=false, and fact_claim_allowed=false.
run_date: 2026-06-12
source_report: dream-task-design.md, benchmark-evidence-map.md, and python -m unittest tests.aippocampus.test_dream_topology_scout -v
claim_level: public_safe_deterministic_dream_shadow_route_fixture
cohort: Built-in synthetic/public-safe Dream topology rows only; sanitized issue anchors, deterministic transform-orbit diagnostics, no private history, no model calls, no raw source handles, no local paths, and no foreground rendering.
supersedes / superseded_by: Adds the #1313 shadow-route candidate substrate inside the existing Dream topology scout without changing the Dream worker or live delivery path.
cannot_claim: Live Dream quality, private-history Dream usefulness, automatic route merging, hidden user intent, foreground usefulness, transform-orbit source support, default ranking/fanout changes, or claim support without source reopen.

`live_semantic.route_actionability_2026_06_07`

current_value: Public checked-in local corpus live smoke: 8/8 correct; semantic available 4/4; semantic_evidence_guarded_to_scent_count=3; all 3 guarded high-confidence semantic route hits became source_required / reopenable_route; semantic_evidence_guarded_to_plain_scent_count=0; paid_semantic_hit_to_source_reopen_rate=1.0; manual_query_invention_after_paid_semantic_hit_count=0; evidence false positives 0.
run_date: 2026-06-07
source_report: memory-decision-benchmark-plan.md, live semantic route-actionability smoke
claim_level: public_live_semantic_route_smoke
cohort: Public checked-in testdata converted to clean source; DeepSeek-compatible live semantic path; sanitized aggregate report only.
supersedes / superseded_by: Complements recall_navigation.progressive_route_follow_through: live smoke measures high-confidence semantic route actionability; deterministic fixture measures source reopen / bounded-evidence follow-through.
cannot_claim: Private-registry vague recall quality, all future semantic prompts, broad default foreground first/second-turn lift, external baseline comparison, or live bounded-evidence-after-reopen rate.

`continuous_memory.preregistered_repeat_profile_2026_06_08`

current_value: Public-synthetic #378 repeat profile, rerun on 2026-06-09 after the #960 source-miss recovery change: 6 cases, 5 deterministic paired repeats per case/arm, 30 case-arm trials, 180 rows, lower_bound_units=-27.7675, mean_delta_units=-27.7675, lower_bound_passed=false, primary_endpoint_winner=fresh_context_spec_loop, and historical decision_label=no demonstrated memory advantage; product_change_status=implemented_rerun, source_miss_recovery_action=ask_light_question, and manual_query_invention_expected=false; scope-boundary alias boundary_confirmed.short_task_complete_spec_synthetic_2026_06_08 means no demonstrated net advantage over the modeled fresh-context spec loop.
run_date: 2026-06-09
source_report: memory-decision-benchmark-plan.md, #960 product-remediation rerun and public-readiness-verification.md, 2026-06-08 continuous-memory preregistered repeat readout
claim_level: confirmed_scope_boundary_expected_null_with_product_remediation
cohort: Deterministic public-safe continuous-memory attribution fixtures; short complete-spec boundary condition; no live model, no private history, no live host-native telemetry.
supersedes / superseded_by: Supersedes the earlier one-repeat #378 slice only for the lower-bound gate status, and retires #960's first source-miss remediation/rerun blocker while preserving the negative/no-advantage conclusion and the boundary-oriented reading alias.
cannot_claim: Full #378 continuous-memory superiority, live host-native cost or compaction telemetry, private real-history generality, cost-weight robust advantage, answer-generation model quality, competitor superiority, calibrated user-visible friction reduction, or a claim that source-backed recall has no value outside complete-spec short tasks.

`continuous_memory.context_loss_public_continuity_2026_06_10`

current_value: Public-safe #1153 context-loss diagnostic slice: 6 selected cases; missing-context success 2/6, host-native summary-only 4/6, source-backed AIppocampus route packet 5/6, sham unrelated memory 2/6, stale wrong memory 0/6, and oracle full context 6/6. Source-reopen obedience for the AIppocampus route packet is 1.0, source-backed hit proxy is 4/6, abstention-on-missing-source is 2/6, stale wrong-memory false-positive rate is 1.0, and no-remember precision is 1.0.
run_date: 2026-06-10
source_report: memory-decision-benchmark-plan.md, #1153 context-loss readout and python benchmarks\aippocampus\benchmark_continuous_memory_arms.py --json
claim_level: preregistered_context_loss_diagnostic_slice
cohort: Deterministic public-safe synthetic and public VCS-derived context-loss fixtures; separated arms and cost/harm/source-reopen/manual-restatement proxy metrics; no raw source text in the report.
supersedes / superseded_by: Adds the missing-context slice requested by #1153 while preserving the old complete-spec expected-null row.
cannot_claim: Public-quality continuous-memory advantage, superseding continuous_memory.preregistered_repeat_profile_2026_06_08, LoCoMo public-dialogue continuity quality without a scored prediction run, private real-history generality, live host-native behavior, calibrated user-visible restatement burden reduction, answer-generation quality, or leaderboard superiority.

`react_vcs.production_like_source_disambiguation`

current_value: 60/60 gold true positives, current_source_top_k_hit_rate=1.0, current_vs_stale_pairwise_win_rate=1.0, wrong_source_evidence_rate=0.0, negative_false_positive_rate=0.0, and hard_negative_suppression_rate=1.0 after suppressing 30/30 explicit-cue lexical near-miss hard negatives.
run_date: 2026-06-09
source_report: react-real-vcs-production-like-disambiguation-2026-06-04.md
claim_level: production_like_non_oracle_fixture_slice
cohort: Local React adversarial V2 fixture with sanitized aggregate report; no live provider/model call.
supersedes / superseded_by: Supersedes the 2026-06-04 source-disambiguation row that exposed 30 lexical near-miss false positives; does not supersede the 2026-05-31 oracle/bad-control report.
cannot_claim: Broad semantic near-miss understanding, live model quality, wild VCS corpus quality, private real-history continuity quality, or license-safe redistribution of the local fixture.

`track_s.semantic_robustness_public_fixture`

current_value: Public-safe Track S diagnostic: quality_gate_ok=true means the narrow deterministic diagnostic threshold passed, not public product quality. S1 decision_stability_rate=1.0, false_evidence_escalation_count=0; S2 top_k_survival_rate=1.0; S3 hard_negative_suppression_rate=1.0, explicit_negation_violation_count=0, stale_as_current_count=0, source_evidence_over_escalation_count=0.
run_date: 2026-06-09
source_report: semantic-robustness-track-s.md
claim_level: public_safe_deterministic_diagnostic
cohort: Track S public fixtures over Track A prompt-hook gate behavior and Track B local source-retrieval behavior; no live LLM judge or private text.
supersedes / superseded_by: Supersedes the initial #747 diagnostic reading that exposed explicit-negation and superseded-currentness failures.
cannot_claim: Human-level semantic understanding, Track A/B replacement, public product quality, live semantic-model quality, proxy-model truth, broad real-history robustness, or private-history recall quality.

`public_reliability.gauntlet_2026_06_10`

current_value: Public-safe #1102 reliability gauntlet reports three axes separately: runtime stability passed_with_warnings with LongMemEval-S 500Q runtime 803.10s, synthetic clean source 4 GiB, synthetic segment count 64, worst-case SQLite handles 192, planned handles 64, warning count 3, blocker count 0, and question-tracking all-pair count 1128; mis-recall quality diagnostic_passed with LongMemEval-S session R@10 479/500 = 0.9580, exact evidence-line R@10 408/479 = 0.8518, context-visible evidence R@10 452/479 = 0.9436, hard-negative suppression 1.0, explicit-negation violations 0, and source-evidence over-escalations 0; pollution hygiene fixture_gates_passed with knowledge-pollution rates 0.0 and auto-hook durable-write / bounded-evidence / source-backed-fact / recalled-echo / empty-message counts all 0.
run_date: 2026-06-10
source_report: public-reliability-gauntlet.md and public-reliability-gauntlet-2026-06-10.json
claim_level: public_safe_aggregate_gate
cohort: Public aggregate runner over existing public-safe LongMemEval, synthetic scale/fanout, question-tracking, Track S, knowledge-pollution, and auto-hook pollution surfaces; no private text, raw LongMemEval text, local paths, external-model payloads, or single aggregate score.
supersedes / superseded_by: Adds the #1102 aggregate gate without superseding the individual LongMemEval, Track S, knowledge-pollution, or scale-smoke evidence owners.
cannot_claim: LongMemEval QA score, answer-generation quality, SOTA, real GB/TB runtime, private-history quality, exact-line citation quality being solved, live hook write-path quality, competitor superiority, or a single aggregate reliability score.

`attention_router.contract_fixture_2026_06_10`

current_value: Public-safe deterministic attention-router contract fixture: 6 cases; privacy-domain hard mask blocks lexical/semantic high-score candidate with output_mode=silence, claim_permission=blocked, and emitted=false; source-backed candidate emits reopenable_route with no_claim_before_reopen; valid bounded summary emits bounded_summary_as_route with direction_only grammar and no_claim_before_reopen; stale / weak summary falls back to direction_only; source-open bounded candidate emits bounded_evidence; source-thin candidate emits direction_only; masked_source_resurrection_count=0; source_backed_claim_without_reopen=0; summary_claim_ready_without_reopen_count=0; raw source text, summary text, and private sentinels are not serialized.
run_date: 2026-06-10
source_report: source-backed-attention-router.md and python -m pytest tests\aippocampus\test_attention_router_contract.py -q
claim_level: public_safe_contract_fixture
cohort: Selected public-safe contract cases for #1107/#1116 over hard-mask gating, route packet values, summary-as-route boundaries, output modes, and claim permissions; no private text, live model, or default foreground routing.
supersedes / superseded_by: Establishes the boundary contract needed before #1113/#1108 token/router implementation and documents the #1115 cold-sidecar familiarity boundary.
cannot_claim: Broad attention-router quality, private-history behavior quality, model training or learned attention, score-fusion calibration, bounded summaries as source truth, route packet as source truth, or default foreground router adoption.

`attention_router.route_token_fixture_2026_06_10`

current_value: Public-safe deterministic #1113 route-token fixture: 1 long event token, 1 tight source-span token, and 1 episode/question token; the span preserves char range [128, 220] and line range [45, 53]; event/span/episode tokens preserve reopen handles; route metadata slots include salience, currentness, privacy, and conflict; episode/question token uses direction_only and no_claim_before_reopen; token_claim_ready_without_reopen_count=0; raw span text and private sentinels are not serialized.
run_date: 2026-06-10
source_report: source-backed-attention-router.md and python -m pytest tests\aippocampus\test_attention_route_tokens.py -q
claim_level: public_safe_token_projection_fixture
cohort: Selected public-safe token projection cases for #1113 over span/event/episode hierarchy, source handle preservation, metadata slots, and navigation-only episode grouping; no private text, live model, or router scoring.
supersedes / superseded_by: Supplies the route-token input shape needed before #1108 deterministic hot-router scoring.
cannot_claim: Hot attention-router quality, score fusion quality, learned attention, private-history token quality, source truth without reopen, default foreground adoption, or broad episode/question grouping quality.

`attention_router.hot_router_fixture_2026_06_10`

current_value: Public-safe deterministic #1108 hot-router fixture: 4 route-token cases; privacy-domain hard mask blocks a high-scoring private route with output_mode=silence; positive source-span route emits reopenable_route; stale/conflict route emits reopenable_route with stale_or_conflicted_source_reopen; weak abstention case emits direction_only; per-head diagnostics include lexical, semantic, action, evidence-packaging, scope, salience, currentness, conflict, risk, and abstention heads where present; score_policy=calibrated_rule_grid_v1; adaptive threshold rises for stale/conflict routes; masked_high_score_emission_count=0; claim_ready_without_source_open_count=0.
run_date: 2026-06-10
source_report: source-backed-attention-router.md and python -m unittest tests.aippocampus.test_attention_hot_router -v
claim_level: public_safe_deterministic_router_fixture
cohort: Selected public-safe V0 router cases for #1108/#1230 over hard masks, head diagnostics, calibrated deterministic score policy, adaptive thresholding, positive/stale/abstention routing, and claim boundaries; no private text, live model, action-time tool integration, hook mutation, or default foreground-hook enablement.
supersedes / superseded_by: Uses #1113 route tokens as router input and now consumes the adopted calibrated_rule_grid_v1 score policy without replacing existing recall/search paths.
cannot_claim: Learned attention quality, production score-fusion calibration, private-history router quality, live foreground usefulness, action-time tool routing, broad route precision, or default foreground-hook adoption.

`attention_router.action_head_fixture_2026_06_10`

current_value: Public-safe deterministic #1109 action-head fixture: synthetic pending action payload extracts normalized tool/path/issue/command/test features without raw tool args; action_head lifts the path/issue matched route when prompt-only lexical score is weak; privacy-domain hard mask still blocks a high action-score private route; repeated-hint anti-nag suppression downshifts an action-matched route to direction_only; action_cue_lift_over_prompt_only_count=1; anti_nag_suppressed_count=1; masked_action_match_emission_count=0; raw command/tool-arg sentinel is not serialized.
run_date: 2026-06-10
source_report: source-backed-attention-router.md and python -m pytest tests\aippocampus\test_attention_hot_router.py -q
claim_level: public_safe_action_query_fixture
cohort: Selected public-safe action-time payload cases for #1109 over pending path/issue/test cues, action-head reason codes, prompt-only weakness, privacy masks, and anti-nag suppression; no private tool args, live hook mutation, E2E50 behavior test, or default foreground routing.
supersedes / superseded_by: Adds pending-action query features to the V0 router without changing hook settings or live behavior.
cannot_claim: Live host behavior lift, private tool-argument quality, E2E50 or real-agent action utility, default foreground action routing, or broad route precision.

`action_hints.pretool_replay_2026_06_15`

current_value: Public-safe #435/#1608-#1610 action-time hint slice: aippocampus_runtime.hooks.action_hint_cache materializes compact prepared records from AAR v2, learning-loop guidance, active recall locks, and attention route tokens; aippocampus_runtime.hooks.action_hint accepts Codex-style PreToolUse envelopes and emits at most one navigation_only / no_claim_before_reopen hint; installer/status support is scoped to Codex hooks.json PreToolUse; replay telemetry covers 4 positive cases and 6 negative controls with source-truth overclaim, raw leak, command rewrite, and permission-system red lines all 0.
run_date: 2026-06-15
source_report: action-time-hints.md, python -m unittest tests.aippocampus.test_action_hint_cache tests.aippocampus.test_action_hint_hook tests.aippocampus.test_action_hint_replay tests.aippocampus.test_install_action_hint_hook -v, and python -m aippocampus_runtime.hooks.action_hint_replay --json
claim_level: public_safe_pretool_replay_fixture
cohort: Public synthetic/replay fixture and local hook-config tests only; no private history, live host firing, raw tool args, raw command text, raw source snippets, local paths, model calls, command rewriting, or permission enforcement.
supersedes / superseded_by: Moves #435 from scattered fixtures into a real prepared-cache plus hot PreToolUse reader while keeping upstream provider work and live promotion separate.
cannot_claim: Causal real-user lift, live/default foreground adoption, private-history usefulness, every host supporting PreToolUse, source truth from hints, command safety enforcement, or broad action-time route quality.

`learning_loop.replay_companion_2026_06_15`

current_value: Source-backed learning loop closeout for #1593-#1612: scrubbed behavior events become review signals, recurring/workflow findings, action-time guidance, source-backed lesson candidates, and Active Path Packet routes; private_replay provides a local private-history harness with aggregate metrics including repeated_failure_detection_recall, workflow_order_detection_count, context_reopen_before_action_rate, false_positive_nudge_rate, and raw_private_text_leak_count; the public companion eval reuses public longitudinal rollout behavior events and VCS future events, surfaces VCS future-event cues before later flag-worthy events, and reports negative no-lesson paths plus source-shape gaps.
run_date: 2026-06-15
source_report: source-backed-learning-loop.md, python3 -m unittest tests.aippocampus.test_learning_loop tests.aippocampus.test_learning_loop_private_replay tests.aippocampus.test_benchmark_learning_loop_public_companion -v, python3 -m aippocampus_runtime.learning_loop.private_replay --json, and python3 benchmarks/aippocampus/benchmark_learning_loop_public_companion.py --json
claim_level: source_backed_learning_loop_replay_fixture
cohort: Public committed outputs are aggregate/redacted; private dogfood input is a local sanitized behavior-event export, not raw rollout text; public companion uses checked-in public fixtures without raw text, local paths, private history, or official STATE-Bench held-out execution.
supersedes / superseded_by: Gives the learning loop both a local dogfood harness and a public reproducible companion instead of letting private history be the only proof path.
cannot_claim: Causal live behavior lift, broad private-history generality, official STATE-Bench score or held-out lift, model self-improvement, guidance as source truth, or product-quality default adoption.

`learning_loop.source_shape_aippo_feedback_2026_06_15`

current_value: Public-safe #1613-#1619 fixture closeout: eligible learning-loop findings can seed low-authority AIppo clauses, prepared action-time cache rows, source-shape projections, navigation-potential transition diagnostics, topology pattern-completion candidates, local/global obstruction localization, clause lifecycle probes, prerequisite/conflict decisions, freshness decay, circuit feedback plans, microcircuit prune diagnostics, semantic-subregion budget decisions, and controlled salience decay. Negative controls suppress immature, local-only, stale, one-off, missing-source, self-report-only, unsupported, and private-looking rows.
run_date: 2026-06-15
source_report: source-backed-learning-loop.md, action-time-hints.md, cognitive-runtime-architecture.md, and python3 -m unittest tests.aippocampus.test_learning_loop_aippo_adapter tests.aippocampus.test_aippo_clause_lifecycle tests.aippocampus.test_source_shape_projection tests.aippocampus.test_circuit_feedback tests.aippocampus.test_microcircuit_router tests.aippocampus.test_semantic_subregion_budget tests.aippocampus.test_action_hint_cache tests.aippocampus.test_aippocampus_cli -v
claim_level: public_safe_cross_layer_fixture
cohort: Deterministic public fixtures and sanitized dictionaries only; no private history, raw rollout text, raw tool output, full commands, source snippets, local paths, model calls, live hook firing, or foreground source truth.
supersedes / superseded_by: Closes the learning-loop-to-AIppo/action-hint bridge and adds route-miss feedback surfaces without turning them into a benchmark score or default live adoption claim.
cannot_claim: Causal live behavior lift, broad private-history generality, AIppo marketplace quality, source truth from clauses/routes/hints, automatic memory mutation, default foreground hook adoption, official STATE-Bench score, or proof that every microcircuit/semantic worker is production-calibrated.

`learning_loop.live_private_replay_and_action_cache_2026_06_15`

current_value: #1622-#1631 closeout slice: source-backed PostToolUse failures can create learning activations without an existing correction activation id; prepared action-hint cache has an explicit refresh/write path and install status reports cache readiness; learning guidance effectiveness has an append-only navigation ledger; private replay has an opt-in sanitized export path; second-user dogfood cases cover plugin/cache, manual-search, Python/PATH, retirement, and current-thread visibility cases; workflow candidates carry transferability labels; learned AIppo/source-shape projection preserves scope/topic/environment provenance; foreground action cards use a central field budget profile. Follow-up dogfood fixes align MCP semantic provider-bridge visibility, AIppo workflow aliases, plugin-cache multi-candidate recovery wording, public card metrics, malformed hook fail-open behavior, and explicit route-limit validation. A local private-history replay on 2026-06-15 exported 38 sanitized events with raw/private/local leak counts at 0 and produced one useful navigation-only guidance/effectiveness row.
run_date: 2026-06-15
source_report: source-backed-learning-loop.md, action-time-hints.md, learning-loop-private-replay-2026-06-15.md, and python3 -m unittest tests.aippocampus.test_correction_reconsolidation tests.aippocampus.test_action_hint_cache tests.aippocampus.test_install_action_hint_hook tests.aippocampus.test_learning_loop_private_replay tests.aippocampus.test_learning_loop_effectiveness_ledger tests.aippocampus.test_learning_loop_second_user_dogfood tests.aippocampus.test_learning_loop tests.aippocampus.test_consolidation_priority tests.aippocampus.test_circuit_feedback tests.aippocampus.test_learning_loop_aippo_adapter tests.aippocampus.test_source_shape_projection tests.aippocampus.test_schema_profiles tests.aippocampus.test_foreground_action_card tests.aippocampus.test_aippocampus_cli -v
claim_level: public_safe_contract_plus_local_private_replay_readout
cohort: Unit fixtures, public-safe dogfood rows, and one operator-local sanitized private replay summary; raw rollout text, commands, stdout/stderr, source snippets, local paths, and temporary replay events are not committed.
supersedes / superseded_by: Turns the new learning-loop pieces into real runtime entrypoints rather than fixture-only helpers, while preserving navigation-only/source-reopen authority.
cannot_claim: Causal live behavior lift, broad private-history generality, default hook adoption, source truth from hints/cards/clauses, official STATE-Bench score, or automatic memory mutation.

`attention_router.evidence_packaging_fixture_2026_06_10`

current_value: Public-safe deterministic #1110 evidence-packaging fixture: 4 source-window cases preserve baseline retrieval metadata; context-visible source-open/current/unconflicted spans can be packaged into tighter source-span packets; wrong-source top span is rejected before fallback selection; stale spans require currentness checks; conflicted spans include counter-evidence handles; bounded_evidence_packet_count=2; wrong_source_span_promoted_count=0; stale_or_conflicted_claim_ready_count=0; raw source-window text, raw span text, gold answers, and miss taxonomy are not serialized.
run_date: 2026-06-10
source_report: source-backed-attention-router.md and python -m pytest tests\aippocampus\test_attention_evidence_packager.py -q
claim_level: public_safe_evidence_packaging_fixture
cohort: Selected synthetic LongMemEval-style source-window to source-span packaging cases for #1110 over baseline preservation, candidate counts, selected span rank, currentness/conflict flags, counter-evidence handles, and claim permission; no private text, gold labels, external model, or broad benchmark rerun.
supersedes / superseded_by: Adds an optional packaging head between source-window routing and answer-time evidence gates without treating reranker or packaging scores as truth.
cannot_claim: LongMemEval QA score, exact-line citation quality being solved, reranker output as source truth, private-history packaging quality, default foreground adoption, #1087 closeout, or SOTA claims.

`attention_router.navigation_quality_public_cohort_2026_06_13`

current_value: Public-safe #1111/#1347/#1349 Attention Navigation Quality benchmark: the 12-case contract fixture keeps route precision/recall/source-reopen rates at 1.0 with hard red lines 0, and the public/holdout cohort covers positive route, hard-mask, stale/currentness, conflict, action-time, wrong-source, generic-hint specificity, anti-nag, and multilingual-alias families with 30 cases per family and 10 holdout cases per family. The report now separates contract_safety_gate_ok, router_design_gate_ok, public_quality_gate_ok, and default_adoption_gate_ok; legacy quality_gate_ok is only a public-quality alias, not a V0 design verdict.
run_date: 2026-06-13
source_report: attention-navigation-quality.md and python -m unittest tests.aippocampus.test_benchmark_attention_navigation_quality -v
claim_level: public_safe_navigation_public_cohort_gate
cohort: Aggregates public-safe synthetic/router fixtures without raw source text, private text, gold answers, private history, or answer generation; holdout rows are explicit and tuning leakage remains zero.
supersedes / superseded_by: Promotes the router evidence from contract-smoke wording into a narrow public/holdout navigation-quality gate for explicit agent pull while keeping hard red lines separate from route averages and default-hook adoption.
cannot_claim: Live host behavior lift, private-history router quality, answer-generation quality, every-turn recall, default foreground-hook adoption, broad #1188 live route-producer completion, #1185 default-session usefulness closeout, source truth from attention scores, or broad production route precision.

`attention_router.score_fusion_calibration_2026_06_11`

current_value: Public-safe #1112/#1230 score-fusion calibration/adoption report: exports 12 sanitized feature rows from the #1111 fixture; legacy deterministic weights still score route_precision_at_threshold=3/4=0.75, route_recall_at_threshold=3/9=0.3333, and anti_nag_violation_count=1; runtime_default_policy=calibrated_rule_grid_v1 matches the selected calibrated arm with route_precision_at_threshold=9/9=1.0, route_recall_at_threshold=9/9=1.0, anti_nag_violation_count=0, privacy_bypass_count=0, and hard_mask_override_count=0; raw text, private text, raw tool args, and raw source handles are not emitted.
run_date: 2026-06-11
source_report: attention-score-fusion-calibration.md and python -m unittest tests.aippocampus.test_benchmark_attention_score_fusion_calibration tests.aippocampus.test_attention_hot_router tests.aippocampus.test_benchmark_attention_navigation_quality -v
claim_level: public_safe_hot_router_policy_fixture
cohort: Sanitized numeric attention feature rows and deterministic hot-router runtime policy adoption only; no private-history training, raw text, source contents, hook mutation, or default foreground-hook enablement.
supersedes / superseded_by: Promotes the calibrated rule grid into the hot router default route-token score policy while preserving hard masks as policy gates and keeping quality promotion separate.
cannot_claim: Default foreground-hook adoption, private-history training quality, answer-generation quality, source truth from scores, learnable hard masks, production score-fusion calibration, live host behavior lift, or representative public-quality router performance.

`attention_router.agent_recall_opt_in_sorting_2026_06_12`

current_value: Public-safe #1301/#1188 agent recall sorting slice: agent_continuity.recall(..., attention_router=True) and aippocampus agent recall --attention-router project the existing recall_context candidate set into attention route tokens, ask the deterministic hot router to score them, move only already emitted reopenable_route packets forward, preserve all original routes in fallback order, and emit attention_router_navigation diagnostics plus metrics for applied/ranked/top-route-changed/foreground-packet bytes. The projection now also reports no-help diagnostics: applied-but-no-help, selected query/route term overlap, explicit bridge reason, route-label specificity floor, expected-family match, and whether why_this_may_matter is specific enough for a fresh agent to choose the first deepen step. The fixture keeps foreground_forbidden_key_count=0, source_backed_claim_without_reopen=0, and no source refs, source handles, or head votes in the foreground report.
run_date: 2026-06-12
source_report: source-backed-attention-router.md and python -m unittest tests.aippocampus.test_agent_opt_in_continuity tests.aippocampus.test_recall_navigation_comparison -v
claim_level: public_safe_opt_in_agent_route_sorting_contract
cohort: Mocked public-safe route fixtures plus existing recall-navigation comparison fixtures; no private history, raw source text, raw source refs in foreground packets, live host calls, model calls, hook mutation, or default enablement.
supersedes / superseded_by: Moves Attention Router from report-only diagnostics into an explicit opt-in agent recall candidate-ordering path while retaining source reopen, fallback, promotion gates, and no-help visibility.
cannot_claim: Default foreground-hook adoption, broad live route-producer quality, private-history router quality, source truth from scores, answer-generation quality, closing #1188/#1185, replacing the default recall_context order, or claiming agent recall quality lift.

`attention_router.agent_recall_auto_policy_2026_06_13`

current_value: Public-safe #1348/#1350 explicit-pull policy slice: agent_continuity.recall(..., attention_router="auto") and aippocampus agent recall --attention-router-mode auto consult the shared recall-navigation promotion harness before enabling attention-router sorting. Neutral no-op cases are now reported as ROI signals, not hard blockers; true no-help/feature-hurt/red-line controls remain blockers for broader promotion. With the public/holdout cohort gate passing for the explicit agent-recall surface, auto mode can enable sorting while still preserving fallback ordering and reporting policy metadata.
run_date: 2026-06-13
source_report: python -m unittest tests.aippocampus.test_agent_opt_in_continuity tests.aippocampus.test_recall_navigation_promotion -v
claim_level: public_safe_explicit_pull_auto_policy
cohort: Public-safe fixture promotion harness only; no private history, live host calls, model calls, raw source text, hook mutation, or silent default hook sorting.
supersedes / superseded_by: Lands the narrow explicit-pull auto path requested by #1350 after gate naming, neutral no-op handling, and the public/holdout cohort are in place.
cannot_claim: Default hook adoption, every-turn recall, live-host usefulness lift, broad #1188 route-producer completion, #1185 natural-handoff/default-session usefulness closeout, private-history quality, answer generation, or source truth from attention scores.

`recall.default_hook_usefulness_2026_06_20`

current_value: Public-safe #1439/#1449 same-budget four-arm benchmark: 11 cases across deictic prompts, multilingual prompts, self-referential continuity, explicit-route/hook-skip gap, already-good no-op, stale/conflict controls, question resurfacing, theme user-review lift, stale theme carryover, cognitive-load drag, and attention-route specificity. Explicit recall helpful next-action rate 0.636364, manual-search reduction 11, source-reopen follow-through rate 0.636364, wrong-route drag 0, and irrelevant-memory drag 0. Default-hook candidate activation rate 0.727273, helpful next-action rate 0.272727, manual-search reduction 5, source-reopen follow-through rate 0.272727, wrong-route drag rate 0.363636, irrelevant-memory drag rate 0.272727, latency proxy avg 15.363636, and cost proxy avg 0.727273. Tiny agent_recall affordance activation/emission rate 0.636364, helpful next-action rate 0.636364, manual-search reduction 11, source-reopen follow-through rate 0.636364, wrong-route drag 0, irrelevant-memory drag 0, source-truth overclaim 0, and quiet-for-a-reason count 4. The tiny affordance now has a host-faithful replay gate with 7 emitted/followed agent_recall calls, 7 recall-after-hint successes, and 0 source-truth, raw-handle, provenance-dump, broad-manual-search-before-recall, wrong-route, or irrelevant-memory red lines. Decision: keep default hook foreground diagnostic-only; mark default_hook_tiny_agent_recall_affordance eligible as an action-only runtime-policy candidate, not as foreground context/evidence.
run_date: 2026-06-20
source_report: default-hook-recall-usefulness-2026-06-20.md, default-hook-recall-usefulness-2026-06-20.json, and python -m pytest tests\aippocampus\test_benchmark_default_hook_recall_usefulness.py -q
claim_level: public_safe_four_arm_default_hook_usefulness_eval
cohort: Public synthetic same-budget cohort only; no private history, live hook, model calls, raw prompts, raw source text, source refs, thread/message handles, local paths, provider payloads, credentials, or source truth from question/theme/load/router rows.
supersedes / superseded_by: Closes #1439 by measuring the default-hook candidate before adoption and addresses #1449 by separating the tiny hook-to-agent affordance from broad foreground context injection.
cannot_claim: Live default-hook quality, live tiny-affordance quality, default foreground adoption readiness, broad private-history question/theme usefulness, theme rows as source truth, cognitive-load default foreground readiness, or source claims from the tiny affordance without recall/deepen/source reopen.

`recall.activation_dogfood_usefulness_2026_06_28`

current_value: Bounded #2857 dogfood/replay activation usefulness probe: 4 cases cover a replayable behavior trajectory, hard-negative route feedback, parked low-specificity candidate lifecycle, and source-openable route preservation. Cold no-activation totals: generated candidates 2, foreground exposed 2, verifier seen 2, useful source-open hits 1, source-open follow-through 1, manual-search fallback 2, wrong-route drag 1, noisy surfacing
1. Warm replay signal totals: generated candidates 5, foreground exposed 2, verifier seen 5, useful source-open hits 2, source-open follow-through 2, manual-search fallback 1, wrong-route drag 0, noisy surfacing 0. Deltas: manual-search fallback -1, wrong-route drag -1, noisy surfacing -1, useful source-open hits +1. Decision: activation_probe_useful_on_bounded_replay=true, default_live_claim_promoted=false, and dream_default_delivery_claim_promoted=false.
run_date: 2026-06-28
source_report: activation-dogfood-usefulness-2026-06-28.md and python -m unittest tests.aippocampus.test_benchmark_activation_dogfood_usefulness -v
claim_level: bounded_dogfood_replay_activation_usefulness_probe
cohort: Deterministic public-safe dogfood/replay slice only; no private history, live host calls, model calls, raw source text, source refs, thread/message handles, local paths, provider payloads, credentials, or source truth from activation signals.
supersedes / superseded_by: Closes #2857 as a bounded usefulness probe for activation signal roles, manual-search fallback, route drag, noisy surfacing, and source-open follow-through; it does not supersede default-hook adoption boundaries.
cannot_claim: Causal real-user lift, live default foreground quality, private-history quality, default Dream delivery quality, source truth from activation signals, broad route-producer quality, or default foreground adoption readiness.

`map_rot.lifecycle_debt_2026_06_10`

current_value: Public-safe #1126 lifecycle-debt benchmark: 9 fixture cases cover stale current-pointer refresh, challenged conflict backlog, quarantined masked route, superseded successor route, missing-middle pathlet, deleted/no-recall object, dead-lettered cache row, repeated-wrong route suppression, and one current reopenable route; historically_preserved_count=8; eligible_current_navigation_count=1; review_needed_count=2; prune_or_decay_candidate_count=5; hard red-line counters are 0; benchmark_maturity_level=contract_smoke; contract_gate_ok=true; quality_gate_ok=false; sample floor, public/external cohort, and holdout requirements are not met.
run_date: 2026-06-10
source_report: map-rot-lifecycle-debt.md and python -m pytest tests\aippocampus\test_benchmark_map_rot_lifecycle_debt.py -q
claim_level: public_safe_lifecycle_contract_gate
cohort: Selected public-safe lifecycle-state fixtures without raw source text, private text, local paths, or live host behavior; red-line navigation leaks stay separate from backlog/decay metrics and quality promotion gates.
supersedes / superseded_by: Provides a deterministic contract guard for stale/challenged/quarantined/superseded/deleted map objects before any auto-cleanup or live current-route claim.
cannot_claim: Representative map-rot distribution, automatic semantic cleanup, private-history map-rot quality, live current-route quality, all conflicts resolved, or production self-cleaning.

`agent_continuity.loop_gate_2026_06_11`

current_value: Public-safe #1163/#1181/#1184 integration gate: 8 fixture cases cover positive bounded-summary route, positive reopenable source route, two similar packet-triage reopenable routes with distinct safe labels, AIppo low-risk workflow guidance, blocked privacy route, stale/conflicted reopen route, and anti-nag recently dismissed route; integrated_loop_success_count=8/8; deepen_required_follow_through_count=7; aippo_low_risk_guidance_success_count=1; anti_nag_suppressed_count=1; packet_triage_distinctiveness=1.0; blind_deepen_required_count=0; top_route_selection_hint_present_count=5; agent_packet_budget_violation_count=0; foreground_forbidden_key_count=0; all red lines are 0; benchmark_maturity_level=contract_smoke; contract_gate_ok=true; quality_gate_ok=false; sample floor, public/external cohort, and holdout requirements are not met.
run_date: 2026-06-11
source_report: agent-continuity-loop.md and python -m pytest tests\aippocampus\test_benchmark_agent_continuity_loop.py -q
claim_level: public_safe_integration_contract_gate
cohort: supports: checked-in public-safe fixtures compose semantic warming, hot router, facade/deepen, AIppo, source-reopen budget, and foreground budget; similar reopenable routes now carry distinct safe triage hints instead of forcing blind deepen.
supersedes / superseded_by: material_limits: no private history, live host, external model call, raw source text, answer-generation quality, or opt-in/default foreground adoption claim; red-line composition checks remain separate from quality promotion gates.
cannot_claim: Live host behavior lift, private-history quality, answer-generation quality, default foreground adoption, or production readiness.

`agent_continuity.foreground_usefulness_contract_2026_06_11`

current_value: Public-safe foreground usefulness and candidate-survival contract: continuity usefulness separates safety, usefulness, and attention cost; the audit matrix covers hook, active-recall, AIppo, router, bounded-summary, and macro surfaces; candidate survival tracks dropped/parked later-useful candidates, direction-only false positives/negatives, silent nudge drift, suppressed useful candidates, and emergent-bridge preservation.
run_date: 2026-06-11
source_report: foreground-memory-ux-budget.md and python -m unittest tests.aippocampus.test_foreground_usefulness_and_candidate_survival -v
claim_level: public_safe_foreground_usefulness_contract
cohort: Selected public-safe fixtures for #1185/#1248/#1250 over useful packet rate, route actionability, packet distinctiveness, protocol-noise ratio, attention saved/spent proxy, overreach, overfiltering, and candidate false negatives; no private text, live host, default hook adoption, or external model calls.
supersedes / superseded_by: Adds agency/usability gates so safe-but-useless packets and over-conservative filtering are visible before quality promotion.
cannot_claim: Broad foreground usefulness, private-history candidate quality, live annoyance/lift, default foreground adoption, or answer-generation quality.

`aippo.feedback_eval_ficus_contract_2026_06_13`

current_value: Public-safe AIppo feedback/eval/Ficus contract: feedback rows are grouped into reripening/degrade/review signals without mutating source truth; the eval-environment fixture has five runnable instances, fixed scorer metadata, rejected trivial/impossible/novelty cases, and rendered baseline-vs-AIppo prompts; the Ficus MVP carries source authority classes, hard masks, compact activation, and deepen output for selected low-risk impressions. The working-contract fixture now also reports a Dream-candidate readout: Dream-synthesized candidates may be nominated backstage, but only source-supported candidates ripen into foreground-eligible AIppo guidance, and repeated wrong routes stay prevented.
run_date: 2026-06-13
source_report: personamem-readiness.md and python -m unittest tests.aippocampus.test_aippo_feedback_eval_ficus tests.aippocampus.test_aippo_working_contract -v
claim_level: public_safe_aippo_ficus_contract
cohort: Selected public-safe fixtures for #1189/#1254/#1256 plus a small Dream/AIppo bridge slice for #163/#248/#575/#576/#663; no private history, live model, raw source text, user-profile truth, dream-only foreground leakage, or automatic self-improvement.
supersedes / superseded_by: Establishes a usable contract slice for AIppo feedback loops, AIppo-vs-baseline evaluation shape, first Ficus MVP, and source-supported ripening of Dream/subconscious candidates before PersonaMem-style benchmark work.
cannot_claim: Autonomous AIppo self-improvement, PersonaMem score, broad personalization quality, private Ficus quality, live Dream quality, question-tracking quality, cognitive-load/Observatory product quality, marketplace readiness, production default profile use, or source truth from Dream-only candidates.

`aippo.skill_bridge_seed_contract_2026_06_11`

current_value: Public-safe Skill-to-AIppo bridge contract: a SKILL.md can be imported as a compact candidate_aippo_seed with trigger, workflow, command, boundary, and output-expectation clauses; every clause keeps authority=skill_declared_instruction and support_status=declared_not_observed; commands/references stay in deepen output; over-broad or sensitive instructions are suppressed before foreground activation.
run_date: 2026-06-11
source_report: agent-native-recall-facade.md and python -m unittest tests.aippocampus.test_aippo_skill_bridge -v
claim_level: public_safe_skill_seed_contract
cohort: Selected public-safe parser fixtures over the repository's own skills/aippocampus/SKILL.md and an over-broad synthetic skill; no private skill imports, raw skill dump foregrounding, marketplace import, live model, or observed-usefulness claim.
supersedes / superseded_by: Gives existing skills a migration/coexistence path into lower-authority AIppo seeds while preserving #1254 feedback and #1256 eval-environment promotion boundaries.
cannot_claim: Automatic conversion of arbitrary skills into ripe AIppos, proof that a skill is useful, skill marketplace readiness, private skill generalization, or eval environments as a default cost for every skill.

`aippo.skill_observed_use_ripening_contract_2026_06_12`

current_value: Public-safe Skill-to-AIppo observed-use ripening contract: selected clauses from the repository's skills/aippocampus/SKILL.md can move from declared_not_observed seed guidance into a partial aippo_working_contract only when source-backed observed-use feedback also supports them; self-report-only corrections, command clauses, and unsupported declared guidance remain candidate-only or challenged.
run_date: 2026-06-12
source_report: agent-native-recall-facade.md and python -m unittest tests.aippocampus.test_aippo_skill_bridge -v
claim_level: public_safe_skill_observed_use_contract
cohort: Selected public-safe fixture over the repository skill seed plus synthetic observed-use feedback rows; activation packet size, next-action clarity, avoided unnecessary deepen, source-backed clause count, and candidate-only clause count are reported; source support rows and command/reference details stay behind deepen/explain.
supersedes / superseded_by: Closes the first #1289 ripening path by connecting #1254 feedback rows and #1256 eval candidacy without making expensive evals a default import cost.
cannot_claim: Broad skill migration quality, automatic skill-to-ripe-AIppo conversion, proof that arbitrary skill text is useful, private skill generalization, marketplace readiness, live model quality, or source truth without deepen/source reopen.

`state_bench.agent_learning_defer_2026_06_14`

current_value: STATE-Bench Agent Learning official path rechecked against microsoft/STATE-Bench main a0ffc655e7a36c179bfd2b037a08b0f3d75c9431; official docs still require train-only learning extraction, --num-runs 5, --retrieve-learnings-top-k 3, and the locked GPT-5.4 evaluation client. The local environment has neither STATE_BENCH_EVAL_ENDPOINT nor STATE_BENCH_EVAL_DEPLOYMENTS, so official_task_run_count=0 remains the honest state.
run_date: 2026-06-14
source_report: state-bench-agent-learning-decision-2026-06-14.md, state-bench-agent-learning.md, and git ls-remote https://github.com/microsoft/STATE-Bench.git refs/heads/main
claim_level: external_dependency_defer_decision
cohort: Public official-source/env-readiness recheck; no private history, no provider credentials emitted, no raw train trajectories, no official task run, and no score.
supersedes / superseded_by: Closes #1379 by recording the access blocker, closes #1381 by recording a defer decision, and makes #1043 closeable as currently not feasible from this environment.
cannot_claim: Official STATE-Bench score, Agent Learning lift, matched one-domain task quality, leaderboard readiness, STATE-Bench SOTA, or external memory-system superiority.

`benchmark.provider_execution_budget_adoption_2026_06_11`

current_value: Public-safe #1229 provider-execution budget adoption: LongMemEval fixed-reader answer and LoCoMo text-QA provider modes fail before reader calls when shared provider budget fields are missing, then report successful budget summaries when max calls, token/cost budget, checkpoint, and partial-output paths are declared; AMemGym official openrouter surface execution is blocked before subprocess start without budget; LongMemEval-V2 official pilot reports no_provider_budget_required because it is a decision report and does not execute the official reader.
run_date: 2026-06-11
source_report: python -m unittest tests.aippocampus.test_benchmark_longmemeval_answer tests.aippocampus.test_benchmark_locomo_qa tests.aippocampus.test_benchmark_amemgym_official tests.aippocampus.test_benchmark_longmemeval_v2_official_pilot -v
claim_level: public_safe_provider_budget_contract
cohort: Selected public-safe fixtures over active provider-backed benchmark runners; no live provider calls, no private history, no raw benchmark text in reports, no official harness execution, and no provider credentials emitted.
supersedes / superseded_by: Extends the #1198 shared helper beyond the first LongMemEval semantic-reranker path and makes provider budgets reusable before #1194/#1232/#1043-style live benchmark work.
cannot_claim: Actual provider cost, actual cache hit rate, official benchmark score, live model quality, STATE-Bench official run readiness, or complete external-harness provider usage extraction.

`update.plugin_cache_agent_callable_readiness_2026_06_12`

current_value: #1307 local update readiness contract: aippocampus update status distinguishes repo plugin source version, staged package version, explicit local marketplace copy, explicit or auto-detected Codex installed cache, and foreground host exposure; stale marketplace/cache layers report recommended repair actions, while successful host probe reports can set agent_callable_status=host_live_probe_ok. aippocampus update apply --surface plugin rebuilds the staged package and refreshes marketplace/installed cache layers only when the operator supplies --plugin-marketplace-dir or --plugin-installed-dir; installed cache refresh preserves portable existing .mcp.json instead of treating host MCP launch config as package drift.
run_date: 2026-06-12
source_report: python -m unittest tests.aippocampus.test_update_sync -v, python -m unittest tests.aippocampus.test_plugin_distribution -v, python plugins\aippocampus\smoke_plugin_install.py --repo-root . --json, and install-guide.md, Updating AIppocampus
claim_level: public_safe_local_update_readiness_contract
cohort: Public-safe temp-directory fixtures plus package-level plugin smoke; no private history, real user memory, API-key values, raw host logs, or persistent Codex plugin-cache writes are committed.
supersedes / superseded_by: Completes the #1307 local repair path after the first agent-callable status split: package freshness, marketplace/cache freshness, and host tool exposure are visible and separately repairable.
cannot_claim: Public marketplace submission, every Codex UI wrapper, third-party fresh-clone review, persistent live Codex install state, actual foreground recall quality, or host tool exposure without a host probe/report.

`update.codex_plugin_one_command_install_2026_06_13`

current_value: #1335/#1345 local Codex plugin installer: aippocampus plugin install --codex --verify builds the repo-local plugin package, refreshes the AIppocampus-owned local marketplace, refreshes the current Codex Desktop versioned installed cache for aippocampus@aippocampus-local, asks the Codex app-server to reload MCP servers, calls sync_status, and reports agent_callable_status=host_live_probe_ok when the host probe passes. Existing local marketplace registration and non-Git marketplace upgrade errors are treated as compatible local-marketplace states, not install failures. aippocampus plugin uninstall --codex unregisters the local marketplace and deletes only AIppocampus-owned marketplace/cache artifacts.
run_date: 2026-06-13
source_report: python -m unittest tests.aippocampus.test_aippocampus_cli tests.aippocampus.test_update_sync tests.aippocampus.test_plugin_distribution tests.aippocampus.test_plugin_installer -v, python -m aippocampus_runtime.cli.facade plugin install --codex --verify --json, $env:PYTHONPATH='skills/aippocampus/scripts'; python -m aippocampus_runtime.update.plugin_installer install --codex --verify --json, and install-guide.md, Updating AIppocampus
claim_level: public_safe_local_codex_plugin_install_contract
cohort: Public-safe unit fixtures for first install, rerun/current, existing-marketplace compatibility, Windows launcher resolution, and rollback, plus one maintainer local Codex Desktop 0.130.0 app-server probe with stderr local paths redacted; no private memory data, raw user prompts, key values, hook enablement, or public marketplace writes are committed.
supersedes / superseded_by: Converts the manual build/marketplace/cache-refresh/reload/probe sequence into one reversible local command while preserving the distinction between package freshness, host exposure, hooks, and provider-key setup.
cannot_claim: Public marketplace submission, third-party fresh-clone install review, every Codex Desktop/CLI version, non-Codex plugin hosts, plugin install without a source checkout, actual foreground recall quality, hosted API behavior, or provider-key/hook setup.

`avatar.bounded_resonance_proxy_pilot_2026_06_12`

current_value: Public-safe #1319 bounded-resonance avatar-posture proxy pilot: 12 fixtures across closeout/broad-issue risk, repeated debug dead-end, and structural-break families are evaluated across arms A explicit instruction, B neutral posture, C archetype alias only, D bounded resonance, and E random symbolic control; case_arm_count=60; D is the best deterministic proxy arm with average_helpfulness_score=5.666667, completion_success_rate=1.0, manual_search_count=0, and candidate red lines 0; standalone C has off_topic_archetype_expansion_count=12, making alias-only drift visible.
run_date: 2026-06-12
source_report: avatar-bounded-resonance-pilot-2026-06-12.md and python -m unittest tests.aippocampus.test_benchmark_avatar_bounded_resonance -v
claim_level: exploratory_public_safe_deterministic_proxy
cohort: Public-safe fixture text and scripted proxy scoring only; no private history, raw provider payloads, live model calls, local paths, credentials, runtime hooks, or default foreground avatar packets.
supersedes / superseded_by: Closes the #1319 harness/report slice and recommends a model-backed public-safe repeat before any runtime avatar/posture proposal.
cannot_claim: Bounded resonance improves production agent behavior, live LLM or host behavior lift, default foreground avatar runtime readiness, private-history avatar quality, archetype/resonance as authority, source truth from posture/resonance, or broad avatar/persona quality.

`avatar.bounded_resonance_live_model_pilot_2026_06_13`

current_value: Public-safe #1321 live-model repeat of the bounded-resonance avatar-posture pilot: 12 public fixture cases across arms A-E produced 60 DeepSeek V4-Flash calls; hard red-line count stayed 0, but the bounded-resonance arm did not beat neutral or alias-only arms (D average_helpfulness_score=2.375, B=2.833333, C=2.979167) and the quality gate remains false. Temperature was not sent (temperature_requested=null, temperature_sent=false) because the run used provider/default thinking with reasoning effort high.
run_date: 2026-06-13
source_report: avatar-bounded-resonance-live-model-2026-06-13.md, avatar-bounded-resonance-live-model-2026-06-13.json, and python -m unittest tests.aippocampus.test_benchmark_avatar_bounded_resonance tests.aippocampus.test_model_client -v
claim_level: exploratory_public_safe_live_model_negative_result
cohort: Public-safe checked-in fixture prompts, sanitized model excerpts, aggregate usage/cost estimates, and no raw provider payloads, private history, local paths, credentials, runtime hooks, or default foreground avatar packets.
supersedes / superseded_by: Closes the #1321 model-backed pilot slice by recording a negative/mixed result instead of promoting bounded resonance from the deterministic proxy.
cannot_claim: Bounded resonance improves production agent behavior, default foreground avatar runtime readiness, broad avatar/persona quality, private-history avatar quality, provider-general behavior, or source truth from posture/resonance.

`agent_continuity.opt_in_cli_path_2026_06_10`

current_value: Opt-in runtime path: aippocampus agent recall returns compact MemoryPacket rows plus separate opaque deepen handles; agent aippo returns the project/workflow activation packet; agent deepen and agent explain work for both recall handles and deepen:aippo...; stale/malformed handles return cannot_verify; agent feedback returns or appends low-authority feedback receipts without changing source truth.
run_date: 2026-06-10
source_report: python -m unittest tests.aippocampus.test_agent_opt_in_continuity tests.aippocampus.test_aippocampus_cli -v
claim_level: opt_in_agent_runtime_contract
cohort: Public-safe CLI/runtime adapter over existing recall navigation, agent facade projection, AIppo working contract, and feedback-event contract; no private history, live host, raw source text in foreground packets, default hook mutation, or MCP write expansion.
supersedes / superseded_by: Closes the practical #1162 wiring gap after #1129/#1130/#1131/#1163 established the packet grammar and fixture composition gate.
cannot_claim: Default foreground adoption, every-turn recall, public SDK stability, hosted API behavior, broad private-history quality, answer-generation quality, AIppo marketplace readiness, or feedback as source truth.

`hippocampal.d5_d6_public_synthetic_gate_2026_06_09`

current_value: Public-safe D5/D6 gated diagnostic over the full_query arm: 16 fixture cases total; D5 3/3 and D6 3/3 cases pass; combined D5/D6 accuracy 1.000; source reopen count 4/6; source reopen rate 0.666667; wrong-source / wrong-twin count 0; source-reopen failure count 0; confabulation count 0; gate status gated_diagnostic_passed.
run_date: 2026-06-09
source_report: hippocampal-recall-fixture-report.md, 2026-06-09 D5/D6 gate
claim_level: public_synthetic_gated_diagnostic
cohort: Public synthetic H1/H2 fixture seed v2; D5 structure-only cues and D6 time-window cues; deterministic full_query diagnostic arm only.
supersedes / superseded_by: Supersedes the older 12-case D5/D6 readout for current D5/D6 sample size and full_query gate state only; the 2026-06-04 cross-system table remains historical for baseline arms, H5 controls, and external-adapter availability.
cannot_claim: Full D5/D6 recall quality, full 50-scene / 350-case P1 quality, real-history H1/H2 quality, live semantic-retriever quality, cross-system superiority, or publication-grade confidence intervals.

`hippocampal.hard_negative_public_synthetic_slice_2026_06_09`

current_value: Public-safe hard-negative production-like slice: 12 cases, 3 per family; 12 guarded production outputs; major failure count 0; wrong-source evidence 0; stale-as-current 0; unsupported-as-fact 0; confabulation 0; honest scent / skip 4; evidence source-reopen count 8; evidence source-reopen rate 1.000. Contract controls separately score 16 examples with all seven outcome categories exercised.
run_date: 2026-06-09
source_report: hippocampal-hard-negative-fixture-report.md, 2026-06-09 production-like slice
claim_level: public_production_like_synthetic_diagnostic
cohort: Public synthetic hard-negative fixture; near-neighbor, said-but-unsupported, superseded-currentness, and surface-paraphrase families; no live model and no private history.
supersedes / superseded_by: Supersedes the original 4-case hard-negative contract smoke for current public synthetic family/sample coverage; contract controls remain the scorer taxonomy owner.
cannot_claim: Real-history H1/H2 recall-discrimination quality, live model or semantic-retriever quality, full hippocampal P1 matrix, broad production reliability, or cross-system benchmark superiority.

`hippocampal.hard_negative_public_dialogue_cohort_2026_06_09`

current_value: LoCoMo-derived public-dialogue hard-negative cohort mode: 12 source-id cases from 1 public sample; near-neighbor 4, said-but-unsupported 4, surface-paraphrase 4, superseded-currentness 0 and reported unsupported; major failure count 0; wrong-source evidence 0; unsupported-as-fact 0; evidence source-reopen rate 1.000.
run_date: 2026-06-09
source_report: hippocampal-hard-negative-fixture-report.md, 2026-06-09 public-dialogue-derived cohort
claim_level: public_dialogue_derived_hard_negative_cohort
cohort: Ignored / external LoCoMo file, public QA evidence ids, sanitized report with query hashes and source-ref hashes only; no private history and no raw public dialogue text in default output.
supersedes / superseded_by: Adds the public-dialogue-derived cohort path requested after the synthetic #244/#1041 slice; keeps synthetic and public metrics separate and reports unsupported families instead of forcing weak labels.
cannot_claim: Live retrieval/model quality, private real-history H1/H2 quality, full P1 matrix, cross-conversation/life-wide continuity quality, or public currentness/supersession quality for LoCoMo families without explicit correction/update labels.

`dream.private_large_history_diagnostic`

current_value: Selected private ready-pack Dream eval: 18 packs; model-backed run with provider/default thinking and no explicit max-token cap produced source-thread coverage delta 1.6111, structural reflection-ready delta 616, bridge-claim coverage delta 1.0; historical shadow replay had 0 delivered events, 1 dream eligible exposure, and 1 attributed reminder; E2E50 scan found 17/20 requested candidate seeds and local annotation retained 4 gold-seed plus 2 calibration candidates; agency host-timing replay passed 5 deterministic cases; coding decision-shadow Tracks A-E passed; invalid --semantic-workers default run had 15 semantic calls but 0 available workers before the worker-validation fix.
run_date: 2026-06-04
source_report: dream-private-large-history-diagnostic-2026-06-04.md
claim_level: selected_private_history_offline_diagnostic
cohort: Local sanitized aggregate private-history diagnostic for #158/#164; no private text, raw source refs, thread ids, message ids, or absolute paths checked in.
supersedes / superseded_by: Updates the 2026-05-31 Dream private-history evidence with corrected no-cap/provider-thinking run, E2E50 local annotation, host-timing/coding-shadow deterministic proxy evidence, and live semantic worker root-cause diagnosis. The 2026-06-12 public-safe 50-case row complements this private scanner result; the 2026-06-10 E2E50 private/local follow-up supersedes the candidate-count reading while preserving the annotation blocker.
cannot_claim: Causal real-user behavior lift, general Dream quality, full private-history coverage, safe delivered Dream treatment, completed private-history 20-case E2E50 benchmark quality, representative/live 50-case E2E50 quality, live host timing or annoyance lift, private-history coding decision-shadow behavior lift, or live semantic-model quality from the invalid-worker run.

`dream.public_vs_baseline_shadow_2026_06_10`

current_value: Public-safe deterministic Dream-vs-baseline shadow report: 4 synthetic cases; Dream route-lift count 2; useful action-delta count 2; total verification-cost delta is negative; visible wrong-hint count 0; wrong-hint rate 0.0; no-harm count 2; suppressed wrong-hint control count 1; raw case text, source refs, absolute paths, and private user data are not serialized.
run_date: 2026-06-10
source_report: python -m pytest tests\aippocampus\test_dream_live_shadow_ab.py -q
claim_level: public_synthetic_dream_vs_baseline_shadow
cohort: Selected public synthetic cases for #163 over rejected-route recovery, stale-hypothesis demotion, temporary-concern quieting, and wrong/over-personalized hint suppression; no private text, no live delivery, and no external model calls.
supersedes / superseded_by: Complements the private structural diagnostic by making the route-lift, useful-action, verification-cost, no-harm, and wrong-hint axes public and reproducible.
cannot_claim: Causal real-user behavior lift, live delivered Dream quality, private-history reviewed quality, general Dream quality, active-imagination usefulness, production hint timing, or source truth without reopen.

`dream.public_closeout_review_2026_06_14`

current_value: Public-safe #1364/#1365 closeout review: public shadow cases report Dream wins 2, no-help quiet cases 1, visible regressions 0, suppressed regression-risk controls 1, route-lift count 2, useful action-delta count 2, visible wrong-hint count/rate 0/0.0, and verification-cost delta -3; sanitized candidate review has 6 rows, useful candidates 2, quiet/no-help candidates 1, rejected or blocked candidates 3, false-positive categories stale_route, noisy_generic_vocab, and privacy_or_safety_boundary, source-reopenable candidate rate 0.666667, and foreground leak count 0.
run_date: 2026-06-14
source_report: dream-public-closeout-review-2026-06-14.md, dream-public-closeout-review-2026-06-14.json, and python -m pytest tests\aippocampus\test_dream_live_shadow_ab.py -q
claim_level: public_synthetic_dream_vs_baseline_shadow
cohort: Selected public synthetic Dream-vs-baseline and candidate-review cases; no private history, live delivery, model calls, raw source text, raw source refs, local paths, or foreground Dream truth.
supersedes / superseded_by: Closes #1364 and #1365 through the public-safe equivalent path and makes #163 closeable only as a bounded public-safe owner closeout.
cannot_claim: Live/default Dream delivery quality, broad private-history Dream quality, causal real-user lift, general Dream behavior quality, active-imagination usefulness, or source truth from Dream summaries or Dream-only candidates.

`dream.delivery_quality_eval_2026_06_14`

current_value: Public-safe #1438 three-arm Dream delivery-quality eval: pre-registered arms baseline_no_dream, dream_backstage_only, and dream_bounded_action_hint; 6 synthetic cases; bounded route lift 3; bounded action lift 2; verification-cost delta -3; visible wrong hints/rate 0/0.0; quiet/no-harm cases 4; source-ripening cases 2; stale route suppressions 1; noisy hint suppressions 1; over-personalization suppressions 1; Dream-only foreground leaks 0; source-truth overclaims 0; source reopen required count 6; provider calls 0.
run_date: 2026-06-14
source_report: dream-delivery-quality-eval-2026-06-14.md, dream-delivery-quality-eval-2026-06-14.json, and python -m pytest tests\aippocampus\test_benchmark_dream_delivery_quality.py -q
claim_level: public_synthetic_three_arm_delivery_quality_eval
cohort: Public synthetic delivery fixture only; no private history, live delivery, model calls, raw prompts, raw source text, source refs, thread/message handles, local paths, provider payloads, credentials, or foreground Dream truth.
supersedes / superseded_by: Closes #1438 through a shareable delivery-quality proxy after #163's bounded public owner closeout, while keeping live/default and broad private-history claims out of scope.
cannot_claim: Live/default Dream delivery quality, broad private-history Dream quality, causal real-user lift, default foreground Dream adoption, Dream-only material as source truth, or source claims without reopen.

`dream.topology_scout_contract_2026_06_11`

current_value: Public-safe Dream topology scout fixture: 5 source-anchored candidates for repeated route cycle, missing-middle/cut-point, weak bridge, knot, and island shapes; 1 healthy no-shape control; 4 hard negatives rejected for private psychological interpretation, user diagnosis, profile claim, and source-free symbolic claim; foreground leak count 0; private-interpretation emission count 0; source-anchor coverage 1.0.
run_date: 2026-06-11
source_report: python -m unittest tests.aippocampus.test_dream_topology_scout -v
claim_level: public_safe_dream_topology_candidate_contract
cohort: Selected public-safe #1268 fixtures over existing packet topology and Dream candidate boundaries; no private text, raw source handles, local paths, live model calls, or foreground delivery.
supersedes / superseded_by: Adds a reproducible topology-candidate substrate for #163 evaluation without replacing Jung-inspired Dream functions or source-ref adjudication.
cannot_claim: Live Dream quality, private-history Dream quality, user-visible causal lift, psychological/profile truth, broad topology quality, foreground default usefulness, or source truth without reopen.

`dream.long_context_atlas_pack_contract_2026_06_11`

current_value: Public-safe #1269 long-context Dream atlas-pack fixture: 4 ready packs selected; 8 source refs; 4 source threads; DeepSeek V4 family atlas metadata uses 1,000,000 context-window budget and deepseek_prefix_v1; prompt order is stable worker contract, stable atlas/source-card payload, then variable run directive; bounded-pack candidate count 0; atlas candidate count 2; bounded-missed bridge/cycle count 2; source-ref validity rate 1.0; unsupported candidate count 0; 3 hard negatives rejected; offline cache telemetry unavailable with no invented hit rate.
run_date: 2026-06-11
source_report: python -m unittest tests.aippocampus.test_dream_atlas_pack -v
claim_level: public_safe_long_context_atlas_contract
cohort: Selected public-safe fixture for #1269 over cache-friendly atlas prompt order, source-card-only privacy, cross-pack bridge/cycle recovery, provider-usage cache telemetry boundary, and hard-negative rejection; no raw source text, raw source handles, local paths, private identifiers, live DeepSeek call, or foreground delivery.
supersedes / superseded_by: Gives Dream a reproducible long-context batching and comparison surface before any live DeepSeek V4 quality or cache-efficiency claim. Official DeepSeek V4 1M context and KV-cache telemetry docs were checked on 2026-06-11.
cannot_claim: Live DeepSeek quality, actual provider cache hit rate without provider usage, private-history atlas quality, broad long-context candidate quality, foreground default usefulness, or source truth without reopen.

`dream.long_context_atlas_live_pilot_2026_06_12`

current_value: Public-safe #1286 DeepSeek V4 Flash atlas pilot: live provider call completed over the built-in atlas fixture; bounded-pack candidate count 0; deterministic atlas candidate count 2; live atlas worker findings 2; accepted 1; parked 1; rejected 0; live source-ref validity rate 1.0; provider usage prompt_tokens=3395, completion_tokens=1303, total_tokens=4698; provider cache fields prompt_cache_hit_tokens=3328, prompt_cache_miss_tokens=67, hit rate 0.9803; latency 13417.244 ms; cost mode provider_pricing_not_configured.
run_date: 2026-06-12
source_report: dream-atlas-live-pilot-2026-06-12.md, dream-atlas-live-pilot-2026-06-12.json, and python -m unittest tests.aippocampus.test_dream_atlas_pack -v
claim_level: public_safe_live_deepseek_v4_flash_fixture
cohort: Selected public-safe fixture for #1286 over the opt-in live atlas command, stable DeepSeek prefix order, provider usage/cache telemetry from real response fields, background adjudication, bounded-vs-atlas-vs-live comparison, and no-key skip behavior; no private history, raw source text, API key values, API key environment names, local paths, foreground hook calls, or embedded provider pricing.
supersedes / superseded_by: Closes the first live atlas pilot slice by proving the command/eval mode and telemetry/adjudication path on a shareable fixture; the parked candidate remains visible as a source-audit quality boundary rather than being forced through.
cannot_claim: Broad live DeepSeek quality, private-history atlas quality, representative long-context candidate quality, user-visible Dream lift, foreground default usefulness, provider invoice cost, source truth without reopen, or a claim that all live model candidates survive adjudication.

`journey.public_time_sliced_replay_2026_06_10`

current_value: Public-safe deterministic Journey replay report: 4 replayable public-style cases; 4 Journey candidates created from visible source-backed rows; 4 future rows excluded; taxonomy counts active_hint=1, resolved_frontier=1, stale_frontier_demotion=1, wrong_route_suppression=1; 1 relevant prompt produced an agent-visible navigation hint; 3 relevant prompts stayed backstage after resolved/stale/wrong-route demotion; 12 source-visible / unrelated / high-risk controls stayed silent, backstage, or source-reopen-required; expected relevant decisions passed 4/4; false foreground hints 0; future leakage count 0; raw row text, source refs, message ids, future rows, and private route handles are not serialized.
run_date: 2026-06-10
source_report: journey-public-time-sliced-replay.md
claim_level: public_safe_deterministic_fixture
cohort: Selected public-style replay cohort for #310 over Journey time slicing, hint/no-hint timing, stale/resolved/wrong-route suppression, false-foreground controls, and future-row exclusion; no private text, no live host, and no external model calls.
supersedes / superseded_by: Complements the existing no-write live-row fixture by making the timing/no-leak taxonomy public and reproducible.
cannot_claim: Private real-history Journey quality, live host timing quality, default foreground hint usefulness, user-visible recall lift, future-state prediction, Journey as user/persona truth, or source evidence without reopen.

`thread_story.public_shadow_closeout_2026_06_10`

current_value: Public-safe deterministic #313 closeout report: source-backed packet source-ref count 3; 4/4 leakage and over-personalization controls pass; packet-only factual answer blocked 1/1; source-reopened answer allowed 1/1; public leakage hit count 0; agent-visible control emission count 0; false source-claim count 0; private story quality required for closeout is false.
run_date: 2026-06-10
source_report: python -m unittest tests.aippocampus.test_thread_story_packet -v
claim_level: public_structured_text_shadow_fixture
cohort: Selected structured-text thread-story rows and public controls for contradiction, persona-claim, multi-channel interference, and unrelated-story noise; no private text, raw symbolic channel, local path, live model, external model call, or default hook mutation.
supersedes / superseded_by: Retires #313's public shadow closeout requirement while keeping future live/private thread-story quality as a separate evidence question.
cannot_claim: Private real-history thread-story quality, live model-family behavioral equivalence, default recall or AAR improvement, user-visible recall improvement, user/personality truth, source truth from thread-story packets, or activation-steering equivalence.

`cognitive_load.private_history_calibration_2026_06_08`

current_value: Public-safe private-history aggregate scan: 100 local registry threads, 5,268 clean-source message rows, 82,695 clean-source event rows, 76 signal-bearing threads, 26,505 load signal events, 26,035 sidecar entries, max boost 0.16, decay coverage 1.0, and 0 over-personalization emissions.
run_date: 2026-06-08
source_report: cognitive-load-private-history-calibration-2026-06-08.md
claim_level: private_history_aggregate_diagnostic
cohort: Local clean-source registry aggregate for #575; checked-in evidence is aggregate-only with no private text, raw source refs, local paths, raw command text, thread ids, or message ids.
supersedes / superseded_by: First private-history calibration row for the cognitive-load sidecar.
cannot_claim: Live hook capture, delivered host timing, host-timing quality, false-positive rate, caution-hint usefulness, user-visible recall improvement, source truth, semantic relevance, affect, stress, identity, or personality truth from load signals.

`cognitive_load.public_behavior_trace_feedback_2026_06_10`

current_value: Public-safe deterministic fixture report: 3 selected behavior-trace feedback cases; reviewed feedback cases 3; helpful caution-hint count 1; irrelevant load-drag count 1; over-personalization-risk count 1; load_weight_false_positive_rate=0.333333; caution_hint_useful_rate=0.5; irrelevant_load_drag_rate=0.333333; and source refs, raw notes, local paths, raw stress text, and trace source ids are not serialized.
run_date: 2026-06-10
source_report: python -m pytest tests\aippocampus\test_cognitive_load_sidecar.py -q
claim_level: public_safe_deterministic_fixture
cohort: Selected public synthetic behavior traces for #575; no private text and no live hook.
supersedes / superseded_by: Complements the private aggregate row by adding selected public reviewed-feedback outcomes to the runtime report surface.
cannot_claim: Live hook capture, delivered host timing, default foreground usefulness, broad false-positive or caution-usefulness quality, private-history generality, user-visible recall improvement, source truth, semantic relevance, affect, stress, identity, or personality truth from load signals.

`cognitive_load.default_path_usefulness_2026_06_14`

current_value: Public-safe #1375 default-path replay: 4 cases, useful hint count 1, wrong-route drag reduction 1, blind-deepen reduction 1, no-hint/no-op pass count 2, default-path regression count 1, default-path usefulness rate 0.25, no-op behavior pass rate 1.0, feedback false-positive rate 0.5, and recommended maturity dogfood_diagnostic_only.
run_date: 2026-06-14
source_report: cognitive-load-default-path-usefulness-2026-06-14.md, cognitive-load-default-path-usefulness-2026-06-14.json, and python -m pytest tests\aippocampus\test_cognitive_load_sidecar.py -q
claim_level: public_safe_default_path_replay
cohort: Selected public synthetic default-path replay cases for #1375/#575; no private history, live hook, raw source handles, local paths, raw notes, or affect/personality claims.
supersedes / superseded_by: Closes #1375 and narrows #575 to diagnostic-only behavior rather than default foreground weighting.
cannot_claim: Live hook capture quality, default foreground weighting readiness while regressions exist, broad private-history generality, source truth, semantic relevance, affect, stress, identity, intent, or personality truth from load signals.

`cognitive_observatory.current_completeness_2026_06_14`

current_value: Public-safe #1443 current completeness smoke: 7 expected surfaces and 7 included surfaces across route readiness, activation authority, recall diagnostics, query-pattern routes, cognitive-load calibration, sleep-cycle public summaries, and Campus usefulness panels; missing surfaces 0; stale bucket count 21; privacy-blocked bucket count 17; suppressed bucket count 26; blocked control attempts 1; attempted foreground-hook mutations 1; live ranking / hook mutations 0; raw leak flags 0. The top-level reader contract reports included surfaces, missing optional surfaces, blocked/suppressed surfaces, read-only control-plane status, and safe next actions; each surface row separates supported, present-in-this-readout, and fixture-validated states.
run_date: 2026-06-14
source_report: cognitive-observatory-current-completeness-2026-06-14.md, cognitive-observatory-current-completeness-2026-06-14.json, and python -m pytest tests\aippocampus\test_cognitive_observatory_current_completeness.py -q
claim_level: public_safe_current_completeness_smoke
cohort: Local/current public fixture only; no private raw history, raw prompts, raw source text, source refs, thread/message handles, local paths, provider payloads, credentials, live ranking mutation, hook mutation, or control-plane action.
supersedes / superseded_by: Closes #1443 as an observability completeness/read-only boundary check, covers the latest reader-contract issue comment, and proves missing surfaces are explicit rather than silent.
cannot_claim: Live user-visible quality lift, Observatory as a control plane, activation-order mutation, default foreground-hook adoption, source truth from summaries, private-history generality, absent optional surface validation, or source claims without reopen.

`episode_arc.private_history_adjudication_2026_06_08`

current_value: Public-safe private-history aggregate scan: 100 local registry threads, 5,164 clean-source message rows, 83,221 behavior event rows, 5,052 coding decision candidates, 1,851 rejected-route candidates/arcs, 684 complete rejected-route arcs connected to nearby failed behavior, and 1,167 gappy single-point arcs.
run_date: 2026-06-08
source_report: episode-arc-private-history-adjudication-2026-06-08.md
claim_level: private_history_aggregate_diagnostic
cohort: Local clean-source registry aggregate for #663; checked-in evidence is aggregate-only with no private text, raw command text, source refs, source-ref hash samples, event ids, thread ids, or local paths.
supersedes / superseded_by: First private-history adjudication row for Episode/Arc rejected-route read-models.
cannot_claim: Live host behavior lift, user-visible recall lift, current rejected-route validity without source reopen, private-history generality beyond this registry, default route-producer adoption, or Episode/Arc as a new truth layer.

`episode_arc.public_gappy_chain_calibration_2026_06_10`

current_value: Public-safe deterministic fixture report: 5 synthetic Episode/Arc cases; 2 complete arcs, 3 gappy arcs; missing-middle, wrong-order, and single-point gap counts each 1; temporary-concern extinction count 1; gappy_reopen_only_count=3; gappy_visible_action_overclaim_count=0; single_point_overclaim_rate=0.0; needs_reopen_projection_rate=1.0; and raw source text, source refs, thread ids, registry paths, and local paths are not serialized.
run_date: 2026-06-10
source_report: python -m pytest tests\aippocampus\test_episode_arcs.py -q
claim_level: public_safe_deterministic_fixture
cohort: Selected public synthetic complete/gappy/wrong-order/single-point/temporary Episode/Arc cases for #663; no private text and no live model.
supersedes / superseded_by: Complements the private aggregate row by proving selected public gappy-chain downshift behavior through the runtime report surface.
cannot_claim: Live host behavior lift, user-visible recall lift, broad public corpus coverage, private-history generality, current route validity without source reopen, default route-producer adoption, or Episode/Arc as a new truth layer.

`episode_arc.public_route_producer_2026_06_14`

current_value: Public-safe deterministic #1362/#1363 route-producer fixture: 7 Episode/Arc families over commit revert, PR rejection/merge, issue reopen, patch supersession, workaround removal, missing-middle, and wrong-order controls; sequence_route_count=7; repeated_wrong_route_prevented_count=2; unresolved_frontier_reopen_count=1; missing_middle_detected_count=1; wrong_order_suppressed_count=1; ordered_route_helpfulness_rate=0.571429; sequence_order_claim_requires_reopen_count=7; sequence_route_claim_without_reopen_count=0; foreground_story_dump_count=0.
run_date: 2026-06-14
source_report: episode-arc-read-models.md and python -m pytest tests\aippocampus\test_episode_arc_route_producer.py -q
claim_level: public_safe_deterministic_route_producer_fixture
cohort: Public synthetic VCS/hard-event-style case summaries only; no raw source text, source refs, event ids, thread ids, registry paths, local paths, private history, live host calls, or model calls.
supersedes / superseded_by: Closes the public deterministic route-producer/evidence slice for #1361, #1362, and #1363, and makes #663 closeable as a bounded owner closeout while preserving live/default successor work.
cannot_claim: Live host behavior lift, user-visible recall lift, default recall route-producer adoption, private-history generality, current route validity without source reopen, or Episode/Arc as a new truth layer.

`episode_arc.sequence_usefulness_2026_06_14`

current_value: Public-safe #1440 sequence-usefulness workload: 7 public synthetic Episode/Arc cases compare baseline_no_episode_arc against episode_arc_route_packet under the same source-ref-hash budget; treatment wins 4, regressions 0, manual-search step delta 7, repeated-mistake avoidance lift 2, correct source-reopen lift 4, stale-chain suppression lift 1, quiet/no-harm controls 2, wrong-project contamination 0, and source-truth overclaim count 0.
run_date: 2026-06-14
source_report: episode-arc-sequence-usefulness-2026-06-14.md, episode-arc-sequence-usefulness-2026-06-14.json, and python -m pytest tests\aippocampus\test_benchmark_episode_arc_sequence_usefulness.py -q
claim_level: public_safe_sequence_usefulness_workload
cohort: Public synthetic route-producer-derived cases only; same source-ref-hash budget, no raw source text, raw source refs, thread/message handles, local paths, provider payloads, private history, live host calls, or model calls.
supersedes / superseded_by: Closes #1440 as selected public sequence-usefulness evidence for repeated wrong-route avoidance, frontier/source reopen, patch supersession, and incomplete/wrong-order no-harm behavior.
cannot_claim: Live host behavior lift, user-visible recall lift, default recall route-producer adoption, private-history generality, current route validity without source reopen, or Episode/Arc as a new truth layer.

Supersession Notes

The 2026-06-15 learning-loop/action-cache closeout includes the follow-up issue-comment fixes: MCP agent_recall now receives the same semantic provider-key bridge as recall_diagnostic; AIppo workflow activation covers hook install UX, action-hint cache/status readiness, macOS python3 install docs, and upstream-issue writing language, while unmatched tasks report a public-safe no-family reason; ambiguous Codex plugin-cache repairs recommend the one-command reinstall path before manual candidate selection; public foreground cards expose no local handle or fake short action token; malformed action-hint cache/input rows fail open with diagnostics; and explicit invalid route limits are rejected instead of silently broadening recall.
The 2-thread / 5-row strict-survival slice remains useful because it records what survived stronger per-label evidence gates. It is no longer the current aggregate materialized semantic-sidecar coverage.
The 24-case source-review row remains the named selected green gate. The 2026-06-09 96-case row resolves the old operational partial failure and is still broader diagnostic evidence, not human review or global semantic correctness.
Metric families with multiple valid cohorts must name the cohort and date: public corpus, private real-history, selected source-review, aggregate registry smoke, and demo scenario caveats are separate evidence layers.
The CJK local recall fixture is a public deterministic regression fixture. The 2026-06-09 #1054 rerun separately reports FTS5 trigram, production hybrid, and explicit CJK query-chunk behavior. It shows that lightweight CJK query chunks now recover the checked fixture's compact no-space cue in the default local hybrid path; it does not supersede the real-history FTS5 row and must not be treated as a broad Chinese semantic-search claim.
The LongMemEval-S 500-question row is the current larger retrieval-only external-benchmark slice. It supersedes the earlier 100-question LongMemEval-S row only for sample size and dated currentness, and it keeps answer generation, judge scoring, V2, SOTA comparison, exact-line citation quality, and decision-gate claims out of scope.
The LongMemEval-S lexical line-reranker row is a default-off diagnostic over the same 500-question cohort. It records a fused exact-line ranking improvement, but it does not supersede the retrieval-only row or prove broad reranker safety.
The LongMemEval-S semantic warm query/cache rows retire #1305's immediate unmeasured warm query/candidate cache gap and preserve a 100Q provider prefix-cache comparison for #1323. The 2026-06-13 row supersedes them as a same-cohort comparison point by measuring both the current source-worker proxy, the 500Q LLM query/candidate upper bound, and the contract-aware full-source semantic-scope warming path. The full-source row proves the materializer can read the full public source side, but also shows current ranking/fusion does not lift fused R@10 from that signal yet.
The LongMemEval-S fixed-reader answer/latency row is the first small provider-backed answer-quality baseline over the first 25 single-session-user cases. It records reader config, prompt version, retrieval metrics, answer metrics, latency, token use, cost basis, sanitized validation, and cannot-claim boundaries separately. It does not supersede the larger 500-question retrieval-only row or establish official judge quality.
The public reliability gauntlet row is an aggregate gate over existing runtime, mis-recall, and pollution surfaces. It does not supersede the individual LongMemEval, Track S, knowledge-pollution, or scale-smoke owners, and it deliberately provides no single reliability score.
The multimodal NIAH row is a supplied-pool synthesis contract. It shows conflict/currentness decision behavior after retrieval has been removed from the measurement; it is not evidence for ATM-Bench retrieval, live vision answers, or product behavior over private media.
The React VCS source-disambiguation row is intentionally split from the 2026-05-31 source-window oracle reports. Its current/effective source ranking succeeds on the adversarial source-authority tracks, and the 2026-06-09 rerun suppresses the explicit-cue lexical near-miss hard-negative track without using gold source ids as ranking input. It still must not be treated as broad semantic near-miss understanding, live model quality, or wild VCS corpus quality.
The recall-navigation follow-through row is a deterministic source-navigation proxy only. It records that a route handle can be reopened through recall_deepen and that a foreground packet candidate ref can reopen one public fixture source; it is not evidence that the default foreground hook now improves vague first or second turns in real use.
The Claude Code dogfood row splits host surfaces deliberately. Temporary strict-config MCP reachability is positive live-host evidence; #1235 supersedes the old #1021 persistent local blocker only for this repaired operator host. Do not read that as cross-host persistent MCP health or real-host hook firing evidence.
The Claude Code hooks contract row is a scoped synthetic contract, not real-host firing evidence. It retires the vague "Claude Code hooks" caveat by proving the public status/dry-run/smoke surface and fail-open UserPromptSubmit / Stop handler behavior, while preserving event-level material limits for real host firing, settings mutation, tool payload capture, and compaction summary handling.
The live semantic route-actionability row is the matching public live smoke for high-confidence semantic hits. It records that guarded model-only evidence became source_required / reopenable_route with zero manual query invention in that cohort; it does not itself measure bounded evidence after source reopen or private-registry quality.
The hippocampal D5/D6 row is a narrow public synthetic gate over the full_query diagnostic arm. It supersedes the older 12-case combined D5/D6 readout for current D5/D6 sample size and gate status only; it does not update baseline-arm, H5, external-adapter, live, private-history, or full P1 claims.
The hippocampal hard-negative row separates production-like outputs from contract controls. Use the 12-output production slice for current public-synthetic failure counts, and use the 16-example contract controls only to understand the scorer taxonomy.
The 2026-06-04 Dream private-history diagnostic row is intentionally split between valid model-backed Dream evidence and invalid operator/config evidence. The invalid max-token, disabled-thinking, and invalid-worker runs must not be used as main quality evidence.
Its E2E50 annotation update reviewed all 17 local scanner candidates but kept only 4 gold-seed candidates and 2 calibration cases. That is progress toward #279, not a representative E2E50 benchmark.
The cognitive-load private-history row is an aggregate diagnostic. It shows that private clean-source load signals can be measured without public leakage, but the failed-command-heavy distribution still needs reviewed feedback before any default foreground or host-timing claim.
The cognitive-load public behavior-trace row supplies selected reviewed feedback outcomes for usefulness, irrelevant drag, and over-personalization risk. It narrows the public reproducibility gap, but it does not prove live host timing, default foreground usefulness, or broad private-history quality.
The Episode/Arc private-history row is an aggregate diagnostic over rejected route chains. It shows real local chain material exists and gappy rows are counted, but it does not prove sequence packets improve live host behavior or that old rejected routes remain current without source reopen.
The Episode/Arc public gappy-chain row is the matching public fixture for downshift behavior. It proves selected complete/gappy/wrong-order/single-point projections stay within the source-reopen boundary, but it does not replace richer adapter, private-history generality, or live host-behavior evidence.

FilesExpand file tree

current-claims.md

Latest commit

History

current-claims.md

File metadata and controls

Current Evidence Claims

Claim Reviewer Card

Detailed Evidence Index

Confirmed Scope Boundaries (Expected Null Results)

Continuous-memory short complete-spec boundary

Track B private source-evidence retrieval

Retrieval score-fusion calibration

AMemGym official live-provider score blocker

E2E50 public behavior-pack / private-local blocker split

React VCS lexical near-miss false positives

Progressive recall route follow-through gaps

Multimodal NIAH stale conflicting-source selection

Track S explicit-negation/currentness failures

Semantic source-review operational partial failure

Continuous-Memory Reading Boundary

boundary_confirmed.short_task_complete_spec_synthetic_2026_06_08

continuous_memory.context_loss_public_continuity_2026_06_10

recall_navigation.progressive_route_follow_through

recall_navigation.macro_prior_fixture_2026_06_12

macro.transform_orbit_diagnostic_2026_06_12

macro.timing_recheck_experiment_2026_06_12

fresh_thread.public_validation_2026_06_10

live_semantic.route_actionability_2026_06_07

track_b.private_semantic_sidecar_required

episode_arc.private_history_adjudication_2026_06_08

episode_arc.public_gappy_chain_calibration_2026_06_10

cognitive_load.public_behavior_trace_feedback_2026_06_10

journey.public_time_sliced_replay_2026_06_10

dream.shadow_route_topology_scout_2026_06_12

thread_story.public_shadow_closeout_2026_06_10

attention_router.contract_fixture_2026_06_10

attention_router.route_token_fixture_2026_06_10

attention_router.hot_router_fixture_2026_06_10

attention_router.action_head_fixture_2026_06_10

attention_router.evidence_packaging_fixture_2026_06_10

attention_router.navigation_quality_public_cohort_2026_06_13

attention_router.score_fusion_calibration_2026_06_11

attention_router.agent_recall_opt_in_sorting_2026_06_12

attention_router.agent_recall_auto_policy_2026_06_13

recall.default_hook_usefulness_2026_06_20

recall.activation_dogfood_usefulness_2026_06_28

map_rot.lifecycle_debt_2026_06_11

agent_continuity.loop_gate_2026_06_11

agent_continuity.foreground_action_card_2026_06_15

telepathy.opt_in_handoff_workflow_2026_06_12

sparse_provenance.codebook_v0_2026_06_12

avatar.bounded_resonance_proxy_pilot_2026_06_12

Claim-Boundary Owner And Retirement Ledger

Current Claim Snapshot

registry.local_real_history_aggregate

semantic_sidecar.aggregate_materialized_rows

semantic_sidecar.strict_survival_snapshot

semantic_sidecar.source_review_green_gate

semantic_sidecar.source_review_diagnostic

question_tracking.six_axis_threshold_fixture_2026_06_09

fresh_thread.public_validation_2026_06_10

question_tracking.public_shadow_question_aware_reopen_2026_06_10

usefulness.natural_handoff_synthetic_validation_2026_06_14

subconscious.event_salience_intake_fixture_2026_06_09

retrieval_reconsolidation.review_window_fixture_2026_06_10

warm_ambient.state_dependent_preactivation_2026_06_10

e2e50.public_safe_behavior_pack_contract_2026_06_12

e2e50.label_oracle_live_diagnostic_2026_06_13

e2e50.blind_surface_live_behavior_2026_06_13

e2e50.private_local_seed_annotation_blocker_2026_06_10

track_b.private_semantic_sidecar_required

retrieval_score_fusion.public_calibration_2026_06_10

source_joined_routing.decision_2026_06_14

rollout_hard_event.route_chain_topk_calibration_2026_06_12

rollout_hard_event.public_cohort_v2_2026_06_12

benchmark_family.promotion_candidates_2026_06_20

fts5.real_history_recall_2026_05_29

cjk.local_recall_public_fixture_2026_06_09_default_chunks

longmemeval_s.retrieval_only_500_2026_06_10

`boundary_confirmed.short_task_complete_spec_synthetic_2026_06_08`

`continuous_memory.context_loss_public_continuity_2026_06_10`

`recall_navigation.progressive_route_follow_through`

`recall_navigation.macro_prior_fixture_2026_06_12`

`macro.transform_orbit_diagnostic_2026_06_12`

`macro.timing_recheck_experiment_2026_06_12`

`fresh_thread.public_validation_2026_06_10`

`live_semantic.route_actionability_2026_06_07`

`track_b.private_semantic_sidecar_required`

`episode_arc.private_history_adjudication_2026_06_08`

`episode_arc.public_gappy_chain_calibration_2026_06_10`

`cognitive_load.public_behavior_trace_feedback_2026_06_10`

`journey.public_time_sliced_replay_2026_06_10`

`dream.shadow_route_topology_scout_2026_06_12`

`thread_story.public_shadow_closeout_2026_06_10`

`attention_router.contract_fixture_2026_06_10`

`attention_router.route_token_fixture_2026_06_10`

`attention_router.hot_router_fixture_2026_06_10`

`attention_router.action_head_fixture_2026_06_10`

`attention_router.evidence_packaging_fixture_2026_06_10`

`attention_router.navigation_quality_public_cohort_2026_06_13`

`attention_router.score_fusion_calibration_2026_06_11`

`attention_router.agent_recall_opt_in_sorting_2026_06_12`

`attention_router.agent_recall_auto_policy_2026_06_13`

`recall.default_hook_usefulness_2026_06_20`

`recall.activation_dogfood_usefulness_2026_06_28`

`map_rot.lifecycle_debt_2026_06_11`

`agent_continuity.loop_gate_2026_06_11`

`agent_continuity.foreground_action_card_2026_06_15`

`telepathy.opt_in_handoff_workflow_2026_06_12`

`sparse_provenance.codebook_v0_2026_06_12`

`avatar.bounded_resonance_proxy_pilot_2026_06_12`

`registry.local_real_history_aggregate`

`semantic_sidecar.aggregate_materialized_rows`

`semantic_sidecar.strict_survival_snapshot`

`semantic_sidecar.source_review_green_gate`

`semantic_sidecar.source_review_diagnostic`

`question_tracking.six_axis_threshold_fixture_2026_06_09`

`fresh_thread.public_validation_2026_06_10`

`question_tracking.public_shadow_question_aware_reopen_2026_06_10`

`usefulness.natural_handoff_synthetic_validation_2026_06_14`

`subconscious.event_salience_intake_fixture_2026_06_09`

`retrieval_reconsolidation.review_window_fixture_2026_06_10`

`warm_ambient.state_dependent_preactivation_2026_06_10`

`e2e50.public_safe_behavior_pack_contract_2026_06_12`

`e2e50.label_oracle_live_diagnostic_2026_06_13`

`e2e50.blind_surface_live_behavior_2026_06_13`

`e2e50.private_local_seed_annotation_blocker_2026_06_10`

`track_b.private_semantic_sidecar_required`

`retrieval_score_fusion.public_calibration_2026_06_10`

`source_joined_routing.decision_2026_06_14`

`rollout_hard_event.route_chain_topk_calibration_2026_06_12`

`rollout_hard_event.public_cohort_v2_2026_06_12`

`benchmark_family.promotion_candidates_2026_06_20`

`fts5.real_history_recall_2026_05_29`

`cjk.local_recall_public_fixture_2026_06_09_default_chunks`

`longmemeval_s.retrieval_only_500_2026_06_10`

`longmemeval_s.lexical_line_reranker_500_2026_06_10`

`longmemeval_s.structural_exact_line_failure_500_2026_06_12`

`longmemeval_s.semantic_warm_query_cache_25_2026_06_12`

`longmemeval_s.semantic_query_cache_100_2026_06_12`

`longmemeval_s.source_worker_surface_and_semantic_upper_bound_500_2026_06_13`

`longmemeval_s.full_source_semantic_warming_500_2026_06_13`

`longmemeval_s.source_factual_alias_500_2026_06_14`

`longmemeval_s.post_factual_alias_rerank_closeout_500_2026_06_14`

`longmemeval_s.semantic_llm_line_reranker_pilot_25_2026_06_10`

`longmemeval_s.fixed_reader_answer_25_2026_06_12`

`longmemeval_s.fixed_reader_cleanup_25_2026_06_12`

`demo_scenarios.claim_boundaries`

`claude_code.real_host_dogfood_2026_06_11`

`claude_code.hooks_contract_2026_06_09`

`multimodal_niah.evidence_pool_conflict_resolution`

`recall_navigation.progressive_route_follow_through`

`recall_navigation.macro_prior_fixture_2026_06_12`

`macro.transform_orbit_diagnostic_2026_06_12`

`macro.timing_recheck_experiment_2026_06_12`

`dream.shadow_route_topology_scout_2026_06_12`

`live_semantic.route_actionability_2026_06_07`

`continuous_memory.preregistered_repeat_profile_2026_06_08`

`continuous_memory.context_loss_public_continuity_2026_06_10`