Commit graph

89 commits

Author SHA1 Message Date
Claude (backend session)
784e412c24 Record the decoy base arm in the blind key so the mandatory control is computed for any panel, not only the first pack 2026-08-08 22:16:20 +03:00
Claude (backend session)
ed9e049248 Re-freeze phase 3 so its purchases split by model instead of queueing on one lock 2026-08-08 22:13:46 +03:00
Claude (backend session)
56b34d48ad Freeze phase 3 before its purchases: panel rule, isolation check and an estimate that uses the ledger price for the model whose screen cells never landed 2026-08-08 21:12:16 +03:00
Claude (backend session)
f6f04ff9ed Replace the decoy plant rules with book-independent ones after the control silently vanished on a different book 2026-08-08 21:07:59 +03:00
Claude (backend session)
7384c57f1a Re-freeze the phase 2 floor stage so its two end-to-end runs and their units can be split across processes 2026-08-08 20:58:50 +03:00
Claude (backend session)
fd3e5224d1 Re-freeze phase 2: claim each cell with an exclusive lock so parallel workers cannot buy the same call twice with the second payment invisible to the ledger 2026-08-08 20:43:25 +03:00
Claude (backend session)
af6d885047 Re-freeze phase 2 mid-stage: split units across processes to raise throughput, and read the battle editor cell from phase 2 in the phase 3 rig 2026-08-08 20:38:18 +03:00
Claude (backend session)
a725be5191 Re-freeze phase 2 before its bulk purchases: bound the ledger guard by per-model price and allow a tenant subset per process 2026-08-08 20:04:06 +03:00
Claude (backend session)
43381fe766 Freeze phase 2: panel composition, noise-floor design, declared contrast family and criteria, before its first purchase 2026-08-08 19:54:07 +03:00
Claude (backend session)
d123526cd5 Re-freeze phase 1 tail again: retry transient provider 5xx with backoff after four capacity refusals cost the mandatory candidate its screen cells 2026-08-08 18:12:18 +03:00
Claude (backend session)
05e2d57fd7 Re-freeze phase 1 tail: raise its ceiling to the owner-approved figure and align the markdown gate with the production sanitizer 2026-08-08 17:59:20 +03:00
Claude (backend session)
930a3792c3 Re-freeze before the rest of phase 1: point the ledger at the experiment price table after it failed on a model the old table did not know 2026-08-08 15:47:13 +03:00
Claude (backend session)
9ca51da0a8 Re-freeze the phase 1 harness after the first draft returned the Chinese source: add the echo gate with re-generation, and widen canon regexes 2026-08-08 15:23:27 +03:00
Claude (backend session)
3df9f54007 Widen the polygon purchase guard from one package directory to the whole eval zone so a second pack can buy under its own freeze 2026-08-08 15:15:49 +03:00
Claude (backend session)
e67ef1bea8 Freeze the phase 0 and 1 pre-registration of experiment 22 with its harness, ledger and thresholds before any paid call 2026-08-08 15:14:14 +03:00
84bc73baa5 Add the missing new file of the previous commit: the command that recomputes every итог number from raw 2026-08-08 02:42:48 +03:00
6aa7cfa8da Give every carrying number of the experiment a path in the tree: add itog.py, guard those numbers in the gate, print the phase-2b family, and write the missing report body 2026-08-08 02:42:37 +03:00
defe869c28 Restore the drift checks lost in the incident recovery and pin a linter for the polygon zone after a shadowed-name bug silently cost half the gate coverage 2026-08-08 02:06:06 +03:00
4b82fb1799 Fix the third recurrence of the ledger-glob defect that certified a false phase figure, correct the file and refusal counts, and restore the valid half of the drift measurement the retraction over-reached 2026-08-08 01:54:18 +03:00
0d05af3058 Restore the en-pair figures and money totals the incident reverted, and account for the phase 2b spend 2026-08-08 01:41:14 +03:00
f9986d9ddf Emit the narrow fourth judging pass for the re-measured phase 2b arms against their bases 2026-08-08 01:18:23 +03:00
d654036b5d Give the point-repair fixer the same term law block the editor gets, after the first pass showed it rewriting canonical terms blind 2026-08-08 01:13:39 +03:00
8be5653862 Wire the phase 2b arms into the same judging rig so their contrasts stay comparable with phase 2a 2026-08-08 00:53:16 +03:00
caaf36d24f Restore the polygon scripts lost to the 08.08 checkout incident and land the phase 2b arms module: exact-number gate matching, degrees-of-freedom correction, counting drift metric, point-repair arms C and E 2026-08-08 00:39:31 +03:00
399147aa76 Fix the drift plan price table and purchase loop to include the deconfound arm 2026-08-07 21:06:19 +03:00
059c681c9e Amend the drift freeze before reading its numbers: add the same editor with the term bank, because the first arm set confounded topology with bank presence 2026-08-07 21:06:09 +03:00
3bc01c6d69 Freeze the drift rig before paid calls: contiguous ten-chapter run, arms and ceiling declared, ledger glob covering its own tag, plus the guarded and QE routing modules that closed the routing branch 2026-08-07 20:36:02 +03:00
6d79dfbd83 Freeze the absolute judging rig before the second pass: family pinned into the key, salt out of the repo, per-axis justification, floor verdict, position control, judge identity, ledger covering the arm tags it pays 2026-08-07 16:52:21 +03:00
0eefc97df8 Freeze the fourth-edition agent judging rig before judging: opaque packet names, randomised layout, mandatory quoted justification, embedded decoy floor and twin controls 2026-08-07 15:01:58 +03:00
ae311e7470 Land the third-edition tier scripts into the repository so paid runs reproduce from the tree, with rig floor, judge variance and own-draft arms 2026-08-07 13:49:42 +03:00
769a2c8451 Freeze the third-edition role-topology pre-registration before paid calls: declared contrast family, computed power, re-declared ceilings, purchase order 2026-08-07 10:51:30 +03:00
d4725999f2 Freeze the second-edition role-topology harness before paid calls: mirrored judging, arms matching the prompt letter, shared purchase ledger, repaired battery 2026-08-07 02:13:17 +03:00
864038a414 Freeze experiment 21 pre-registration and the role-topology harness before any paid call, with detector, battery and verifier results measured at zero cost 2026-08-06 18:40:37 +03:00
3d5d5d54f3 Land accepted editor-role experiment 20 with review header downgrading unpersisted judge claims, plus its harness and judge scripts 2026-08-06 16:10:26 +03:00
0c3b55f074 Land accepted diff-contract experiment 19 with review header pinning frozen-scorer numbers, premise-review verifiers and the falsification run log 2026-08-06 16:10:26 +03:00
41cde9f54d Land accepted editor-wire probe 18 with acceptance review header, frontier harness and gold, and section renumber fix 2026-08-06 16:10:10 +03:00
4d1d9e525c Freeze polygon pre-registration for four falsification arms, the measured deepseek-v4-pro effort-mapping correction, and the premise-review harness proving the probe C rig batches engine-identically 2026-08-05 04:58:00 +03:00
52e12f311e Freeze polygon pre-registration and rig for the Q4b diff-editor arm before any paid call 2026-08-05 00:34:40 +03:00
080c15d60a Freeze polygon pre-registration for the editor-wire probe and its harness before any paid call 2026-08-04 22:58:35 +03:00
271f8b0adf Land accepted bank arbitration research 24 with review header, harness and gold, archived draft, decision D39.102, backlog rows 128-132, ping arithmetic corrected 2026-08-04 19:05:46 +03:00
f66518a860 Pin the polygon environment: one commented requirements file with the heavy embedding extra split out, interpreter version declared, README rule five corrected 2026-08-03 02:02:26 +03:00
Claude (backend session)
87a0124eae Ratify D39.80: onboarding overhaul with product requirements registry, glossary, single state carrier, archive link repairs, stable backlog IDs, and puml diagrams moved next to backend code 2026-08-01 23:50:04 +03:00
Claude (backend session)
f4c4d458d6 Land softsign decl study and both decisions as D39.76: seed stopgap for two names, finding closed outside the core, engine anchor rule gated behind a cross-pair measure 2026-08-01 18:43:47 +03:00
Claude (backend session)
f15e35202a Land bank quality both stages as D39.69: head-aware series batching, classifier-first typing, data-driven decl stemmer with neuter, golden verdict-neutral, fix list and owner questions queued 2026-08-01 15:59:14 +03:00
Claude (backend session)
036dc78a41 Ratify D39.67: course renamed to bank quality, mini-run dissolved into ideal phase, checker package promoted to soon, debt anchors moved to living architecture reference 2026-08-01 04:51:02 +03:00
Claude (backend session)
0b6fcc9527 Land polygon bank raw-material and fix probes as D39.65: rows 21 and 22 closed, series co-batching and type re-derivation verified live, terminologist cap mine recorded 2026-08-01 03:58:22 +03:00
Claude (backend session)
3649a7feed Rework CLAUDE.md into a stable project map with lessons only, re-sync backend and eval READMEs to code facts, relocate stand and adult-content notes to zone docs 2026-08-01 03:44:25 +03:00
Claude (backend session)
ca654eb1e0 Land the polygon package as D39.61: R6 negative, fidelity gate passed with no cost found, DeepSeek swapped flash weights under the slug so the owner fork blocks paid runs 2026-08-01 00:57:33 +03:00
Claude (backend session)
21e7099133 Accept D39.51 measurements as D39.52: format example fixes the language slip and joins the build, P1 consistency fails to separate 2026-07-26 22:05:20 +03:00
Claude (backend session)
fb5e873157 Accept polygon package eight as D39.50: emission widening closed negatively, banknote discipline and answer-language hole become the design core 2026-07-26 20:07:19 +03:00