Commit graph

169 commits

Author SHA1 Message Date
Claude (backend session)
29338dfbd3 Correct the Gemini thinking-knob and billing notes against vendor docs and ping the orchestrator about the engine losing the glossary on Gemini hops 2026-08-22 01:03:20 +03:00
Claude (backend session)
70966c219f Switch the Gemini thinking knob to the standard compatibility parameter after the vendor rejected the nested form 2026-08-22 00:39:26 +03:00
Claude (backend session)
457bc0854a Add the Gemini thinking variant with its own budget knob 2026-08-22 00:38:51 +03:00
Claude (backend session)
b222ac8a08 Record the Gemini quirk that silently drops extra system messages and merge them for that vendor only 2026-08-22 00:36:26 +03:00
Claude (backend session)
2b665fd5c4 Add the decisive probe case with two system messages, the shape the bought Gemini cells actually used 2026-08-22 00:35:02 +03:00
Claude (backend session)
20783511e0 Freeze a live probe that separates a wrong token counter from a system prompt that never arrives 2026-08-22 00:33:23 +03:00
Claude (backend session)
af0447aa6b Move the unused stopped-phase budget into the English role-prompt phase without touching the pack ceiling 2026-08-22 00:23:16 +03:00
Claude (backend session)
88a79a06f0 Add gates that prove each review fix behaves, surface the multi-instance pair mine, and raise the English role-prompt ceiling for the sanctioned purchases 2026-08-22 00:21:29 +03:00
Claude (backend session)
80aafdeb99 Fix the pair-path leak that sent Chinese editor rules to English one-pass cells, declare each arm reasoning level explicitly, and record the review errata 2026-08-22 00:12:32 +03:00
Claude (backend session)
7c63ec062c Record the vendor-portability verdict and make the tally anchor and controls configurable per panel 2026-08-21 01:19:02 +03:00
Claude (backend session)
22820288b2 Add the panel gate for truncated cells and record the judging-run audit with its two incidents 2026-08-21 00:17:20 +03:00
Claude (backend session)
26bf4f297a Freeze the erratum arm that runs glm-5 with its vendor-default thinking, the confound the bought cells revealed 2026-08-20 23:47:45 +03:00
Claude (backend session)
3433a85d1f Freeze the vendor-portability pre-registration and the panel gate that byte-checks every packet text against its cell 2026-08-20 23:38:59 +03:00
Claude (backend session)
7647579feb Skip chapters with an incomplete panel instead of aborting the whole emission, and print the skips 2026-08-20 22:45:10 +03:00
Claude (backend session)
db02b7ab64 Freeze the reading packet format: three judging axes and one chapter per packet 2026-08-20 20:38:56 +03:00
Claude (backend session)
276c85b4be Make the blind-reading panel and packet name parameters now that reading is the standing instrument 2026-08-20 19:56:14 +03:00
Claude (backend session)
3075566104 Freeze two ablation variants with the verification tail cut from both and opt-in trace capture 2026-08-20 18:52:19 +03:00
Claude (backend session)
eb49b651d4 Add the blind-reading emitter for this panel and record the role-prompt reading verdict 2026-08-20 18:06:48 +03:00
Claude (backend session)
273d2df41f Merge the two overlapping rule lists into one axis list, add the restored-paragraph arm and harden the coverage gate 2026-08-20 16:16:45 +03:00
Claude (backend session)
824bd209eb Restore three rules the first prompt draft dropped and gate every battle-prompt rule by declared disposition 2026-08-20 15:48:06 +03:00
Claude (backend session)
7f15323a8d Freeze the role-built one-pass prompt, its buying code and the sanctioned ceilings before the first cell 2026-08-20 15:39:46 +03:00
Claude (backend session)
650b4fea12 Freeze the buying code and owner-sanctioned ceilings for the critic-input presentation measurement 2026-08-20 13:57:08 +03:00
Claude (backend session)
0ef049f492 Pick blind-reading units only where the panel is complete, then emit the judging pass 2026-08-17 02:05:39 +03:00
Claude (backend session)
6d0759d331 Buy the last two Chinese floor cells the sanctioned five did not cover due to a low per-cell estimate 2026-08-17 02:01:00 +03:00
Claude (backend session)
5d6b3a6489 Raise both ceilings by the owner's word to buy the last five noise-floor cells 2026-08-17 01:53:40 +03:00
Claude (backend session)
e727881bdc Raise the two run ceilings to cover the remaining 24 cells after the balance top-up 2026-08-17 00:51:26 +03:00
Claude (backend session)
3ca4936212 Raise the run ceilings to the re-priced estimate and trim closed phases to actuals 2026-08-16 22:27:32 +03:00
Claude (backend session)
c776ea8911 Re-pin DeepSeek to the post-cutover peak price after the vendor switched to time-based billing 2026-08-16 22:27:32 +03:00
Claude (backend session)
2087237877 Fix eight defects found by the cross-family adversarial review before any judging answer exists 2026-08-16 03:53:13 +03:00
Claude (backend session)
b6bcefc199 Add the project draft-quality gate to the enriched-draft arm after the source-echo erratum 2026-08-16 03:13:32 +03:00
Claude (backend session)
0961a10d5d Freeze the buying code and ceilings for the D17 run before the DeepSeek price cutover 2026-08-16 03:00:10 +03:00
Claude (backend session)
22a8a6b70f Freeze the buying code for the Japanese leg noise floor: a second editor generation instead of the near-deterministic fixer floor 2026-08-15 20:30:05 +03:00
Claude (backend session)
8c68828303 Freeze the buying code for the Japanese leg of H-4: second editor arm, its own cash phase and measured guard expectation 2026-08-15 19:49:38 +03:00
Claude (backend session)
a03ac66b0b Freeze the pre-registration and density gate for re-judging the tier pass before any answer arrives 2026-08-15 19:30:47 +03:00
b751ac6f1a Point the adjudication module at the judging corpus in the books repository 2026-08-15 17:21:16 +03:00
fa051f9ea3 Point the judging harness at the books repository where judge answers now live under version control 2026-08-15 06:40:15 +03:00
629458a60d Freeze the ceiling before rebuying the identical Japanese floor halves that understate the axis threshold 2026-08-15 05:19:58 +03:00
4ff45c54ce Freeze the ceilings before the sanctioned E3 critic rebuy and bring the ja ceiling into git, where spending had already exceeded the committed bound 2026-08-15 03:39:49 +03:00
9847693556 Freeze the self-assessment micro-probe before its purchase, with the no-decisions caveat asserted by selftest 2026-08-14 21:57:46 +03:00
6a34d9b5d6 Record cash-desk errors and timestamps in the gemini attempt journal so the order verdict rests on a complete log 2026-08-14 21:52:42 +03:00
7d0f133247 Let the floor fall back to deterministic places when an axis has no semantic critic so the reconnaissance axis gets a threshold instead of silently buying nothing 2026-08-14 21:08:25 +03:00
ce33a8bcac Freeze the gemini backfill driver before its first purchase: per-attempt journal, exponential backoff, per-cell cap and the order threshold of twelve of sixteen 2026-08-14 20:58:43 +03:00
548d1ce6dd Make the output budget an axis parameter and gate cell validity so truncated empty cells are quarantined and rebought instead of judged as arms 2026-08-14 18:28:25 +03:00
e64d921e48 Raise the ja ceiling after the guard fired on a conservative projection, with the measured cell price recorded 2026-08-14 18:07:37 +03:00
e691a7e4c0 Add the axis floor mode to both pair drivers so each axis buys its own noise threshold 2026-08-14 17:51:04 +03:00
969ad2f8dc Make the noise floor and judge audit belong to the axis so the en panel gets its own threshold and cannot read any axis raw cells 2026-08-14 17:37:53 +03:00
1f66b354fe Freeze the ja reconnaissance axis and its raised ceiling: shared base purchasing, pair-scoped tags and the descriptive status asserted by selftest 2026-08-14 17:36:28 +03:00
9c8a6c0b06 Raise the en axis ceiling to the owner-sanctioned amount and the pack ceiling by the same delta so the phase-sum invariant holds 2026-08-14 17:32:49 +03:00
bd5c4850da Parameterize the judging harness by axis so the en panel gets its own work dir, blind keys, arm list and contrast family 2026-08-14 17:15:23 +03:00
9273698447 Allow the en axis to buy a subset of contour arms so a phase guard cannot truncate an arm mid-purchase 2026-08-14 17:12:51 +03:00