Commit graph

204 commits

Author SHA1 Message Date
Claude (backend session)
b1d54b6103 Add strategic architecture review: verdict, SOTA-2026 survey, 9 verified findings, ranked risks, course corrections, and pilot protocol amendments 2026-07-09 15:22:00 +03:00
Claude (backend session)
ff5f000da5 Consolidate docs after reconciliation audit: name 05-decisions-log source of truth, fold exp08 economics and grok reasoning fix, archive closed prompts, add backend/polygon/strategic handoff prompts 2026-07-05 21:32:18 +03:00
Claude (backend session)
1aed2c7627 Trim grok-reasoning re-narration in exp08 and cost_model: single-source the mechanism in provider-quirks, keep only cost-relevant facts plus pointers 2026-07-05 21:17:06 +03:00
Claude (backend session)
9792285e76 Correct exp04's 'grok without reasoning' characterization — grok-4.3 defaults to low reasoning (~13s), production editor sets reasoning_effort:none for 0 tokens; ranking unchanged 2026-07-05 21:07:39 +03:00
Claude (backend session)
f4e7124c4b Consolidate config: wire grok-4.3 as the editor with an explicit reasoning_effort:none off-switch (control:effort), keeping DeepSeek's echo-mine no-op, and mark escalation stubs interim 2026-07-05 20:58:09 +03:00
Claude (backend session)
84737d032c Reconcile grok-reasoning across polygon docs: reasoning_effort:none disables it (probe bug), editor stays grok-4.3 with provider-quirks slug fixed; economics unchanged; add pilot mono-editor arm 2026-07-05 20:36:48 +03:00
Claude (backend session)
3911bafe30 Land polygon session 3: coverage-gate precision (0 FP/57, flip safe), cost-model v2 on grok-editor stack, and Phase-2.5 pilot protocol with a Go-mirror memory-eval 2026-07-05 15:44:44 +03:00
Claude (backend session)
6a68dfd44e Wire the approved glossary into the monolingual editor as CONFIRMED dst-only target constraints (D1) and drop {{text}} from the editor prompt 2026-07-05 15:37:04 +03:00
Claude (backend session)
38c7f3a04a Close the two adversarial-review majors: strip default-ignorable code points in both memory normalizers and trim surrounding whitespace on seed keying surfaces, bumping the normalizer version 2026-07-05 14:51:36 +03:00
Claude (backend session)
3df5983fee Add the Phase-2 polygon prompt: memory-eval tests deterministic selective injection with memory.go the source of truth and eval/memory_hotpath.py a pre-cc57c7b reference prototype 2026-07-05 14:35:27 +03:00
Claude (backend session)
cc57c7bc02 Close external-review findings: gate on confirmed misses only, per-language min-key with collision-prone AMBIGUOUS downgrade, and fail-loud seed guards 2026-07-05 14:04:45 +03:00
Claude (backend session)
068f59d22f External-review memory bank v2: accept with 6 confirmed findings in the short-key/collision/gate zone, F1 and injection surface clean 2026-07-05 13:38:48 +03:00
Claude (backend session)
aa00339abb Land memory bank v2 (glossary hot path, injection surface, F1 materialization, decl-aware post-check) with 8 self-review fixes 2026-07-05 08:13:04 +03:00
Claude (backend session)
baf6ae31bd Augment the memory-v2 backend prompt with research/14's quantified post-check false-flag data, the stored-decl completeness caveat, and the E1 measurement assets 2026-07-05 03:47:16 +03:00
Claude (backend session)
e23160b139 Answer the memory-v2 prompt open questions: confirm F1 materialization, seeded-glossary first veha, decl-form post-check as flagger, code-assembled injection 2026-07-05 03:42:27 +03:00
Claude (backend session)
15cf46f7a3 Add memory-v2 onboarding prompt for the next backend session: schema freeze, deterministic hot-path injection, F1 materialization, and mandatory post-check 2026-07-05 03:36:28 +03:00
Claude (backend session)
6d96bb4ac8 Close step-3a external-review findings: spine reconciliation for dangling/mislabeled chapters, UTF-8 normalization, ASCII quote-depth, ruby full-replace, and tiling fuzz tests 2026-07-05 03:33:57 +03:00
Claude (backend session)
6700a0f2a6 External-review шаг 3a: accept with 7 confirmed findings led by silent chapter-loss on malformed epub, recommend spine reconciliation 2026-07-05 03:04:15 +03:00
Claude (backend session)
8096224a2f Confirm the ruby-seed contract for memory-v2: base to src, reading as the Polivanov dst bridge strengthening B6, with name-vs-double-reading classification 2026-07-05 02:47:45 +03:00
Claude (backend session)
b26e4df465 Replace the placeholder chunker with a real sentence-aware source segmenter, txt/epub ingest, and ruby-reading capture persisted to a new v4 table 2026-07-05 02:41:57 +03:00
Claude (backend session)
073ed62791 Add step-3a onboarding prompt for the fresh backend session: real chunker, txt/epub import, ruby capture 2026-07-05 01:48:58 +03:00
Claude (backend session)
e4ca3a57fd Add the adaptive-memory research verdict (doc 14) with four executable probes: flagship logprob access, in-context demos vs glossary, lemma post-check, and the deterministic lever curve 2026-07-05 01:34:18 +03:00
Claude (backend session)
1878c3ea95 Track the closed memory-validation and extraction sessions' eval artifacts that committed docs already reference 2026-07-05 01:34:18 +03:00
Claude (backend session)
0b31222b5c Close external-review findings: fold classifier version and local backend tag into the snapshot, catch nested echo-mine forms, restrict escalation to the translator role 2026-07-05 01:22:40 +03:00
Claude (backend session)
3a5b966c16 External-review Veha 2.5: accept with 7 confirmed findings, confirm deepseek-v4-pro price, add coverage-precision validation task for Polygon 2026-07-05 01:07:36 +03:00
Claude (backend session)
ed4f4ff5a7 Apply self-review fixes: fold fallback wire config into the snapshot, make escalation resume-idempotent and ceiling-tolerant, strip think in the coverage gate, gate only the translator 2026-07-05 00:35:34 +03:00
Claude (backend session)
a98af88b12 Add an opt-in live-conformance harness per adapter behind a live build tag, verified against DeepSeek showing the echo mine stays disarmed on the real wire 2026-07-04 23:59:52 +03:00
Claude (backend session)
5cbf75cae8 Add single-hop escalation routing a deterministic content-failure to a named fallback model once, re-gated, budget-capped, with channel-B isolation by type 2026-07-04 23:56:50 +03:00
Claude (backend session)
6dc8f13c6c Fix the Kimi API base URL to api.moonshot.ai after a live /models probe showed api.kimi.ai does not resolve 2026-07-04 23:44:16 +03:00
Claude (backend session)
8957046424 Add the configurable excision coverage-gate porting refusal_bench sentence-coverage bit-for-bit, emitting excision_suspect and folded into the snapshot 2026-07-04 23:33:42 +03:00
Claude (backend session)
d15cdb3b71 Add D12 failure-resilience model: three-class failure taxonomy, F3 idempotency fix bounded to at-most-once, progress-surfacing spec grounded in durable-execution prior art 2026-07-04 23:10:32 +03:00
Claude (backend session)
737d4c6a68 Fail-fast on any config that disables DeepSeek thinking at reasoning=off, the echo-mine regression gate that keeps thinking on by default 2026-07-04 22:31:49 +03:00
Claude (backend session)
fc05b6c2e7 Answer Polygon on silent provider refusals (echo detected, excision pending) and add onboarding prompt for the robustness session 2026-07-04 22:13:34 +03:00
Claude (backend session)
357dde296b Add chapter/chunk book loop with per-chunk disposition flags, attempt-axis retries, content-hashed resume, and F4/D2.5 fixes 2026-07-04 22:00:30 +03:00
Claude (backend session)
0dbcce7e97 Harden refusal benchmark: split verdict taxonomy, live-audited provider configs, local warmup 2026-07-04 21:06:47 +03:00
Claude (backend session)
68059b0d3f Spec the memory-bet indicative eval as a ready-to-run task for the eval session 2026-07-04 20:40:10 +03:00
Claude (backend session)
636919c52f Add per-model capability wire-shape layer (D3.1) and fold context-assembly, memory-version, and resolved capability into the job snapshot (D5.2) 2026-07-04 20:32:14 +03:00
Claude (backend session)
dee6efa282 Accept memory-bank validation: GO on schema and deterministic gates, bet gated by in-house eval running parallel; apply pings to R3/R7 2026-07-04 20:26:00 +03:00
Claude (backend session)
8fac2b8968 Answer Phase-1 backend questions D8-D11: content-hash snapshot, ja-ru acceptance book, all-terms consistency, budget as config ceiling with stale cost numbers flagged 2026-07-04 19:42:37 +03:00
Claude (backend session)
ce71a1a752 Add polygon eval scripts for local-model, editor and MoE benchmarks plus ollama launcher and provider config 2026-07-04 19:26:36 +03:00
Claude (backend session)
1985c16d23 Collapse economics to single direct-key EU-SaaS contour, drop BYOK/RU-law, finalize memory-validation session prompt 2026-07-04 19:10:30 +03:00
Claude (backend session)
87b92ba696 Add decisions log and provider-quirks, set grok-4.3 editor default and OpenAI-retained model stack per owner and editor bake-off 2026-07-04 18:47:45 +03:00
Claude (backend session)
f77ea6cc41 Record Phase-1 decision requests (editor structure, bad-chunk disposition, escalation models, memory schema, ongoing Analyst) with reviewer recommendations 2026-07-04 17:20:10 +03:00
Claude (backend session)
c504cfeb14 Mark native Anthropic adapter deprecated (kept as reference for Phase-2 Gemini adapter); record real DeepSeek cache-hit confirmation 2026-07-04 17:07:38 +03:00
Claude (backend session)
56c2f5a61c Fix tmctl report: materialize request_log rows so they outlive the query context that defer-cancel tore down mid-iteration 2026-07-04 16:56:21 +03:00
Claude (backend session)
c4676d53dc Drop Anthropic from the model stack per owner decision: remove provider and Claude models, repoint escalation and C2 judge to non-Anthropic 2026-07-04 16:35:52 +03:00
Claude (backend session)
9d78385c37 Apply external review fixes: snapshot local overrides, gate-enabled guard, TM-key doc note, db-lock gitignore, techdebt disposition 2026-07-04 16:29:25 +03:00
Claude (backend session)
56670e3d67 Scope CheckKeys to reachable models so backend runs on two keys; escalation keys required only when gates enabled 2026-07-04 15:49:27 +03:00
Claude (backend session)
214e005155 Harden Phase 0 after self-review: snapshot coverage, response-price fallback, usage clamp, job-status lifecycle, config guards, source normalization, hash robustness 2026-07-04 14:49:43 +03:00
Claude (backend session)
2006989b9b Untrack .claude/settings.local.json so its allow rules aren't repo-supplied 2026-07-04 14:35:50 +03:00