Companion to CHANGELOG.md for autonomous RSI iterations.
- Hub-anchor
l2_norm_residual→surprise_pressuremodulates sentineleffective_max_turns(soft nudge only). - Harness exposes
rsi_cycle_metrics+ extendedrsi_evolution.research_refs. - Residual wiring (gap closure):
engram_core::ops::prediction_residualonstore::update; ego/recent-trace prior onremember/storefor hub anchors;mcp::sentinel_turn_suffixsurprise-aware;resolve_hub_anchors_for_surprisepresentation-stratum fallback pre-handoff.
Sources: arXiv:2508.05766, arXiv:2504.09301
Scores: CRS 0.84 · Lyapunov 0.82 · RSI-accel 0.88 · perf 0.90 · safety 0.88
Verify: scripts/rsi_cycle1_verify.sh (OVERALL_EXIT=0) · Trace: trace:1782839919_rsi-cycle-1-verification--surprise-aware-sentine · Tile: tile:formal_spec_rsi-cycle-1---surprise-aware-sentinel-v0-7-0-bet
combined_sentinel_pressuremax-blends hub residual surprise + egodrift_velocity.- Sources: arXiv:2508.04435, arXiv:2508.05766
sentinel_turn_suffixreceives conv_arc/human_forward for presentation-stratum anchors.
- Removed invalid
subagent = null; testfull_system_audit_loop_toml_parses.
scripts/rsi_batch_verify.sh,scripts/rsi_batch_mcp_capture.py
Batch scores: CRS 0.85 · Lyapunov 0.85 · RSI-accel 0.87 · perf 0.91 · safety 0.89
MCP Cycle 2: trace:1782841433_rsi-cycle-2-lyapunov-ego-sentinel-blend-shipped · tile:formal_spec_rsi-cycle-2---lyapunov-ego-blend
MCP Cycle 3: trace:1782841436_rsi-cycle-3-turn-record-session-intent-sentinel- · tile:formal_spec_rsi-cycle-3---session-intent-parity
MCP Cycle 4: trace:1782841440_rsi-cycle-4-full-system-audit-loop-toml-parse-fi · tile:formal_spec_rsi-cycle-4---audit-loop-toml
MCP Cycle 5: trace:1782841443_rsi-cycle-5-batch-verify-pipeline-v0-7-0-beta-7 · tile:formal_spec_rsi-cycle-5---batch-verify
weighted_sentinel_pressure+ENGRAM_SENTINEL_RESIDUAL_WEIGHT(default 0.65).- Sources: arXiv:2508.04435, arXiv:2508.05766
resolve_rsi_cycle_number()wiresrsi_cycle_metrics.cycle; exposessentinel_residual_weight.
meta_workflow_okinrsi_cycle_metrics;meta_workflow_registry_in_harness_bundletest.
rsi_batch_verify_all.shsupportsRSI_CYCLE_MIN/MAXthrough cycle 9.
Batch scores: CRS 0.86 · Lyapunov 0.86 · RSI-accel 0.88 · perf 0.92 · safety 0.91
MCP Cycle 6: trace:1782841559_rsi-cycle-6-weighted-lyapunov-sentinel-blend · tile:formal_spec_rsi-cycle-6---weighted-blend
MCP Cycle 7: trace:1782841563_rsi-cycle-7-engram-rsi-cycle-harness-metrics · tile:formal_spec_rsi-cycle-7---rsi-cycle-env
MCP Cycle 8: trace:1782841567_rsi-cycle-8-meta-workflow-registry-harness-expos · tile:formal_spec_rsi-cycle-8---meta-workflow-registry
MCP Cycle 9: trace:1782841570_rsi-cycle-9-batch-verify-extended-cycles-6-9-v0- · tile:formal_spec_rsi-cycle-9---batch-verify-6-9
build_turn_protocol()PLAN/ACT/LOG phases in wake harness (arXiv:2607.01224).agent_discipline.turn_protocol+metamemory_kpisin harness injection.- Source: arXiv:2607.01224 — memory as trainable skill.
SessionMetamemoryCounters: recalls, empty recall rate, writes/recall ratio, consult-before-write violations.note_metamemory_toolwired inmcp::handle_tool_call; snapshot inrsi_cycle_metrics+ handoff packet + session receipt.
Batch scores: CRS 0.87 · Lyapunov 0.86 · RSI-accel 0.89 · perf 0.92 · safety 0.92
Verify: RSI_CYCLE_MIN=10 RSI_CYCLE_MAX=11 scripts/rsi_batch_verify_all.sh
ENGRAM_CONSULT_BEFORE_WRITEsoft/hard/off gate onremember/update/update_with_tensor_bond.- Harness exposes
consult_before_write_gateinrsi_cycle_metrics.
build_trajectory_meta_reviewaggregates metamemory acrossreceipt:session_*sidecars.scripts/rsi_trajectory_meta_review.shverify pipeline.
Verify: RSI_CYCLE_MIN=12 RSI_CYCLE_MAX=13 scripts/rsi_batch_verify_all.sh
scaffold_registryin harness with versioned ritual components.ENGRAM_SCAFFOLD_PROMOTION_GATEonpromote_hotfor scaffold/RSI tiles.- Criteria: CRS ≥0.74, recalls ≥1, zero consult-before-write violations.
Verify: RSI_CYCLE_MIN=14 RSI_CYCLE_MAX=14 scripts/rsi_batch_verify_all.sh