Retain failed authorized replication and reconcile gap work

Assistant: codex
Assistant-Model: gpt-6.1-sol
Assistant-Session: 01a1045b-933d-7aa1-b2db-104557487a79
This commit is contained in:
tegwick 2026-10-04 11:15:05 +02:00
parent 88826ca598
commit 5da8b2405c
40 changed files with 2816 additions and 10 deletions

View file

@ -26,7 +26,7 @@ Research fixtures exercise bounded mechanisms; they are not implementation recip
| Implementation mappings | Locate scoped mechanisms for transactional and offline recovery | [TransactionalRecovery](technologies/TransactionalRecovery.md), [OfflineRecovery](technologies/OfflineRecovery.md); no product supplies a certified complete configuration |
| Metadata and publication checks | Validate contract examples, candidate inventory/hashes, classification, evidence locators and immutable release artifacts | `make schema-test evolution-check classification-check reference-check`; checks establish structural traceability, not source truth |
| Research bench | Plan and execute bounded source investigations, retain failures, screen structure and explicitly replay compatible cached runs | [Bench adapter](research/bench/README.md); paid inference uses llm-connect and its task authorization/budget, not automatic invocation |
| Learning visibility | Read a selected retrospective ledger, prospective protocols, measured/unknown effort, immutable snapshots and comparison readiness | [ComparisonWindows.json](research/evolution/ComparisonWindows.json); 8 prospective cycles in 7 declared groups, none reaches the 3-cycle review trigger |
| Learning visibility | Read a selected retrospective ledger, prospective protocols, measured/unknown effort, immutable snapshots and comparison readiness | [ComparisonWindows.json](research/evolution/ComparisonWindows.json); 9 prospective attempts/cycles in 8 declared groups (one wholly failed execution), none reaches the 3-cycle review trigger |
| Deterministic scheduled reporting | Observe retained natural daily, weekly and October 1 monthly reports from activity-core | [Receipt reconciliation](docs/evidence/2026-10-04-loop-reconciliation.json); zero enrolled consumers, dated owner health snapshot, no admitted pattern-editing executor |
The frozen preview and live research are separate version surfaces. Published candidate
@ -44,7 +44,9 @@ paths. Vocabulary and catalog names do not constitute additional implemented pat
The consumed development pilot scored 18/24 in each presentation condition after
model-review adjudication. It establishes neither a naming benefit nor equivalence.
Independent human scores, fresh-scenario replication and decision-time evidence remain
missing. M1/M2 remain revise/defer, and canonical admission is open. Full canonical graph
missing. Fresh atomic replication tooling and a preregistered 12-attempt run exist,
but all attempts failed without answers; [retained record](research/evaluations/language-replication-02/README.md).
M1/M2 remain revise/defer, and canonical admission is open. Full canonical graph
export, held-out retrieval evaluation and real-repository applicability studies retain
their recorded prerequisites. Eight disclosed preview navigation cases are development
checks; they are not held-out retrieval results.
@ -52,7 +54,7 @@ checks; they are not held-out retrieval results.
Full-contract recurrence, production fitness, accessibility conformance, native crash
and off-device recovery guarantees are not established by the scoped source studies or
authored runtime fixtures. Seven prospective cycles have positive recorded research
time; LEARN-0017's time is unknown. A matching group label is not proof of comparability,
time; LEARN-0017 and LEARN-0019 have unknown research time. A matching group label is not proof of comparability,
and there is no basis for a diminishing-returns or plateau claim.
The capability entry and hosted-publication preparation exist, but hosted federation