[SLICE 18][CORE-E] Persist drift/calibration epochs, quarantine/suspension triggers, counterexample campaigns, and safe-baseline comparison receipts #13

New issue

Open

opened 2026-09-08 01:57:00 +00:00 by nsaspy · 1 comment

nsaspy commented

2026-09-08 01:57:00 +00:00

Owner

Parent epic: #16
Depends on: #35
Research: prolog-rlm#404B

Goal

Persist the post-evaluation/post-admission evidence needed for CITADEL to detect degradation without rewriting prior history.

Implement durable records for:

drift/calibration epoch identity and context scope;
monitored metric/error distributions and detector/version refs;
counterexample campaign definition/bounds/coverage/results;
admitted baseline/candidate comparison receipts;
hard-guard violation events;
hold/quarantine/suspend-on-drift decisions from upstream policy;
reevaluation requirements and superseding epochs;
historical calibration validity and stale-calibration status.

Acceptance

Material drift can suspend an admitted object while historical runs still resolve it; no bounded-counterexample-found result claims universal safety; catastrophic hard-guard violation survives average-score improvement; stale calibration is explicit; Symbolic Memory stores decisions but cannot lower promotion thresholds or make its own object executable.

Parent epic: #16 Depends on: #35 Research: prolog-rlm#404B ## Goal Persist the post-evaluation/post-admission evidence needed for CITADEL to detect degradation without rewriting prior history. Implement durable records for: - drift/calibration epoch identity and context scope; - monitored metric/error distributions and detector/version refs; - counterexample campaign definition/bounds/coverage/results; - admitted baseline/candidate comparison receipts; - hard-guard violation events; - hold/quarantine/suspend-on-drift decisions from upstream policy; - reevaluation requirements and superseding epochs; - historical calibration validity and stale-calibration status. ## Acceptance Material drift can suspend an admitted object while historical runs still resolve it; no bounded-counterexample-found result claims universal safety; catastrophic hard-guard violation survives average-score improvement; stale calibration is explicit; Symbolic Memory stores decisions but cannot lower promotion thresholds or make its own object executable.

nsaspy referenced this issue

2026-09-08 01:57:01 +00:00

[SLICE 06][CORE-B] Build bitemporal current/history projections and lifecycle resolution #25

nsaspy referenced this issue

2026-09-08 01:57:01 +00:00

[SLICE 05][CORE-B] Implement append-only semantic commit protocol and immutable full-IR persistence #26

nsaspy referenced this issue

2026-09-08 01:57:39 +00:00

[EPIC][CORE] Library-first Symbolic Memory v1: durable semantic world model, learning history, federation, and zero-model recall #38

nsaspy commented

2026-09-08 01:59:19 +00:00

Author

Owner

MACHINE-SPIRIT #404D implementation refinement — EvaluationSnapshot / risk / drift

Persist an immutable EvaluationSnapshot before validation evidence is consumed. It should bind candidate fingerprint/version, exposure frontier, admitted baseline, task population/environment epoch, Frozen-Spec/verifier/metric/hard-guard fingerprints, validation-origin policy, risk-ledger allocation, counterexample profile, transfer assessments, trust-root generation, resource bounds and all material dependency generations.

For empirical paired candidate-vs-baseline evaluation, persist the selected named statistical profile and its anytime-valid risk allocation; a conservative v1 family allocation may use alpha_n = alpha_family * 6/(pi^2*n^2) assigned before candidate n consumes validation. Symbolic exact profiles do not need fake statistical sampling.

Drift/calibration events remain epoch/profile scoped. Structural verifier/schema/runtime/source/identity/translation generation changes can invalidate current admission deterministically; ADWIN-like numeric detectors may contribute detector-relative drift evidence but never semantic truth.

Counterexample receipts must preserve exhaustive|bounded|resource_limited|unknown, generator/profile/seed/bounds/cases/stopping reason and witness/minimization refs. no bounded counterexample is never universal safety.

Refs prolog-rlm#404D and #34/#35/#43.

## MACHINE-SPIRIT #404D implementation refinement — EvaluationSnapshot / risk / drift Persist an immutable EvaluationSnapshot before validation evidence is consumed. It should bind candidate fingerprint/version, exposure frontier, admitted baseline, task population/environment epoch, Frozen-Spec/verifier/metric/hard-guard fingerprints, validation-origin policy, risk-ledger allocation, counterexample profile, transfer assessments, trust-root generation, resource bounds and all material dependency generations. For empirical paired candidate-vs-baseline evaluation, persist the selected named statistical profile and its anytime-valid risk allocation; a conservative v1 family allocation may use `alpha_n = alpha_family * 6/(pi^2*n^2)` assigned before candidate n consumes validation. Symbolic exact profiles do not need fake statistical sampling. Drift/calibration events remain epoch/profile scoped. Structural verifier/schema/runtime/source/identity/translation generation changes can invalidate current admission deterministically; ADWIN-like numeric detectors may contribute detector-relative drift evidence but never semantic truth. Counterexample receipts must preserve `exhaustive|bounded|resource_limited|unknown`, generator/profile/seed/bounds/cases/stopping reason and witness/minimization refs. `no bounded counterexample` is never universal safety. Refs prolog-rlm#404D and #34/#35/#43.