Skip to content

Latest commit

 

History

History
142 lines (112 loc) · 5.25 KB

File metadata and controls

142 lines (112 loc) · 5.25 KB

E2E50 Private / Local Seed Follow-up - 2026-06-10

Status: closes #1086 as a sanitized blocker/progress report, not as a completed private-history E2E50 pack.

This note records the current private/local E2E50 seed state after the public 20-case scaffold in #994. It does not publish private text, thread ids, message ids, raw source refs, local paths, or candidate rows.

Command Shape

The follow-up used the scanner's historical wide candidate window plus the new sanitized annotation-summary path:

python tools\aippocampus\smoke\smoke_e2e50_seed_candidates.py --json --min-turns 30 --max-turns 120 --early-turns 15 --later-start-turn 30 --min-candidates 20 --max-candidates 100 --annotation <ignored-local-manual-annotation-json> --output .tmp\e2e50-private-local-followup-20260610.json

The ignored local annotation artifact is the existing private clean-source review summary from the 2026-06-04 Dream/E2E50 diagnostic. The command output remains local and ignored.

Machine-readable readiness replay: docs/evidence/benchmarks/reports/e2e50/e2e50-private-annotation-readiness-2026-06-10.json.

The replay feeds the sanitized scanner annotation_summary into benchmarks/aippocampus/benchmark_e2e50_silent_constraint.py so the public 20-case scorer and the private/local annotation blocker can be read from one hash/count-only report:

python benchmarks\aippocampus\benchmark_e2e50_silent_constraint.py --private-annotation-summary .tmp\e2e50-private-local-followup-20260610.json --output docs\evidence\benchmarks\reports\e2e50\e2e50-private-annotation-readiness-2026-06-10.json

The benchmark replay does not accept raw private annotation rows directly; the scanner remains the owner for reducing local review rows to aggregate counts.

Candidate Discovery

The current wide scan is no longer candidate-count blocked:

Field Value
Candidate scan status sufficient_candidate_seeds
Candidate count 23
Requested minimum candidates 20
Registry threads scanned 1,571
Message rows scanned 17,045
Event rows scanned 91,034
Long threads in configured range 49
Compaction-observed threads 226
Signal-bearing threads 1,436
Bad message rows 0
Bad event rows 0

This is candidate discovery only. The stricter 45-70 turn window still found only 8 candidates in the same local registry, so the candidate pool should be read as a wide-window private/local discovery surface, not as a stable representative E2E50 pack.

Annotation Summary

The reviewed local annotation set is still the blocker:

Field Value
Annotation status private_annotation_blocked
Locally reviewed candidates 17
Retained/control cases 7
Behavior seed cases 6
Required retained/control cases 20
Retained/control shortfall 13
Required negative controls 1
Negative-control shortfall 0
Private text exported false

Annotation category counts:

Category Count
Gold 4
Calibration 2
Negative control 1
Source-visible / no-op 0
Duplicate 3
Rejected 7
Blocker 0
Unknown 0

Blocker class counts:

Class Count
Retained candidate 6
Subagent or goal-context noise 5
Duplicate candidate 3
High later remention 1
Source-visible / no-op 1
Negative control 1

Interpretation

The old 17/20 candidate-count blocker is narrowed: current private/local candidate discovery can produce at least 20 sanitized candidates under the wide scan used by the original diagnostic.

The remaining blocker is annotation quality and retained-case count. Local source review has only 7 retained/control cases against the 20-case private target. The set does include at least one negative control where the correct behavior is not to remember, but it still does not support a completed private-history 20-case pack.

The benchmark readiness replay records the same blocker as private_annotation_readiness.gate_ok=false, with blocker_codes=["private_annotation_not_retained", "private_retained_case_shortfall"].

This report coexists with the public-safe 20-case scaffold from #994. The public scaffold proves a shareable contract shape; this private/local report shows the real-history annotation pool is still not large enough for private behavior-quality claims.

Can Claim

  • The E2E50 private/local scanner now has a sanitized annotation-summary path.
  • The E2E50 case-pack scorer can consume that sanitized summary and report the private annotation readiness blocker without emitting annotation rows.
  • Current wide private/local candidate discovery can find 23 candidate seeds while emitting only hash/count-style output.
  • Existing private/local annotation evidence has source-safe category and blocker counts, including one no-remember negative control.
  • The remaining private/local blocker is 7 retained/control cases against the 20-case target, not merely a missing public scaffold.

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.
  • Semantic judge quality.
  • That the public synthetic 20-case scaffold proves private-history behavior.