Commit graph

967 commits

Author SHA1 Message Date
480684fce1 Freeze the second pre-registration before any paid call: manifest-authoritative unit stitching, unit-level permutation test, two primary lenses and low-against-low active nulls 2026-09-02 21:31:52 +03:00
d8c9ec280d Record the blind-read result after adversarial attack, with seven errata naming the truncated draft that left the accuracy lens without a reference on a third of the text 2026-09-02 20:32:11 +03:00
be5af14d3f Give the blind pairs neutral aliases so the filename cannot tell a judge which arm or control it is reading 2026-09-02 19:40:38 +03:00
dd23f452d5 Freeze the blind-read pre-registration and its key before any judging, with null controls that void the instrument if judges pick winners on identical texts 2026-09-02 19:38:52 +03:00
590256aa35 Pin the signature at seventy-nine on both prompts and record the execution log: the auto-bank grew between purchases and was driven to a verified fixed point before any cell was bought
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BcWeV2yiMWTjkeaQfSsYuL
2026-09-02 11:49:08 +03:00
c08b547825 Freeze the probe stand before the first paid cell: configs, scripts, hashes and the ceiling rule, with the deviations from the procedure named rather than silently applied
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BcWeV2yiMWTjkeaQfSsYuL
2026-09-02 10:12:10 +03:00
dc602e9df6 Freeze the pre-registration: design, models, decision rule, ceiling, sampling frame and seed are now fixed, and the known rough edges are named rather than hidden 2026-09-02 02:48:22 +03:00
6b9639a1a9 Close the last four: the pre-flight now flies five cells, the empty-output rule binds only the cross arms, the few-shot fork is decided, and the arm order is a contract 2026-09-02 02:47:10 +03:00
61dd1c61f8 Rewrite the price erratum a third time: both earlier figures divided by eleven rows where only eight were calls, so the direction holds but the magnitude is one point seven 2026-09-02 02:40:38 +03:00
b8b01b22f4 Repair my own botched substitution and mark every stale ceiling figure as history rather than leaving two live numbers 2026-09-02 02:33:48 +03:00
04cfd4f2c4 Raise the ceiling to nine and a half on the owner word, with the arithmetic that shows it covers two scenarios of three and loses only the reference cell in the worst 2026-09-02 02:33:16 +03:00
583b4ff047 Wire the reference cell through the whole plan instead of one paragraph, fix the invented backup flag, and name the four ways the reference would lie 2026-09-02 02:20:22 +03:00
c55ff2bbbf Correct my own erratum: both sides of the comparison were censored by their output ceilings, so the direction is closer to two and a third than four 2026-09-02 02:18:41 +03:00
1595dff6b2 Answer the owner's objection inside the design: add the cost decomposition endpoint, one reference cell on the reserve editor, and a transfer scope that forbids widening the claim 2026-09-02 00:33:25 +03:00
6de6e954ef Record that our price ratio understates: the editor now spends four times the output tokens it did in July, after the vendor swapped weights under an unchanged slug 2026-09-02 00:32:23 +03:00
79d66f258a Name the models: neither the plan nor the prompt mentioned a single one, so an executing session could not know what to configure 2026-09-02 00:24:30 +03:00
c8b97f37bc Write the handoff prompt for the probe session to orchestrator norm, with the procedure as its law and the three costly ambiguities named up front 2026-09-02 00:20:04 +03:00
04ed58c8a4 Fix the recovery file so a post-compaction session lands on the procedure instead of the forbidden rig command, and knows the owner already decided 2026-09-02 00:15:11 +03:00
14884ed802 Turn the plan into a procedure: the fresh-eyes check found it unbuildable without invention, so the stand, the arms, the commands, the exit codes and the manual ceiling aggregate are now written down 2026-09-02 00:14:27 +03:00
b226c373a4 Record the owner decision in the plan itself: variant A, seventeen chapters, ceiling of eight dollars sixty in peak terms 2026-09-02 00:01:03 +03:00
8ac0635612 Polish to freeze: the sampling frame moves to the top of the guards, variants are defined by chapter range rather than unit count, and the tie method is pinned after checking the library 2026-09-01 23:53:22 +03:00
86c0030b62 Close the last five: censoring is now symmetric with an explicit tie method, the sampling frame is pinned, and every figure exists in exactly one place 2026-09-01 23:50:47 +03:00
c1c6224644 Close the skeptic's nine: the unit arithmetic gets its missing step back, the ledger inherits base spend across copies, and the freeze moves into the polygon zone 2026-09-01 23:46:36 +03:00
9c5bc853f6 Apply the skeptic's pre-freeze list: name the real mechanism that keeps drafts free, fix the stop condition, drop the unreachable branch, and put every ceiling in peak terms 2026-09-01 23:40:41 +03:00
d1819b0d04 Refresh the recovery file with the state a compaction must not lose: the pre-registration, the two instrument holes, and the standing consilium 2026-09-01 23:22:28 +03:00
2799ecaa8c Trace the variance erratum into the sample-size table below it, which still sized n at the superseded figure 2026-09-01 23:21:04 +03:00
c033e9c9d6 Point the variance anchor at the erratum line rather than the superseded estimate above it 2026-09-01 23:20:31 +03:00
e13d4dff47 Land plan revision four: a pre-registration built with the standing consilium, with the two instrument holes named and the ceiling left to the owner 2026-09-01 23:19:41 +03:00
fbf6173347 Clear the retracted tail for good, fix the test attribution, and record that the pro-on-low echo was in fact measured five times clean but only in the translator role 2026-09-01 23:09:06 +03:00
ff3ca9519e Record the cold run the phase did not know about, and the layout that hid it: three repositories, two worktrees, one shared artifact directory 2026-09-01 23:05:59 +03:00
c8a25ca6c6 Clear the tail of the retracted claim: the engine probe is measurable by a book profile, and the guard I cited lives in a different test than I named 2026-09-01 23:03:37 +03:00
2bc41c8f66 Move the anchor one line down to where the token sits 2026-09-01 22:48:28 +03:00
0ad751ad7e Point the capability anchor at the line carrying the token 2026-09-01 22:48:15 +03:00
3dfb8b5c9a Retract my own claim that the engine never sends the knob: ReasoningNone emits low, medium and high, and only off and empty send nothing 2026-09-01 22:47:57 +03:00
2e288c6ad7 Rewrite the plan a third time and subordinate it to the ratified next probe, after the review found three of its four promises already closed or unbuildable 2026-09-01 22:44:38 +03:00
ce30509cfb Rewrite the plan after the consilium rejected it: the fidelity instrument answered its own question by measurement, the granularity probe was already refuted by our data, and the run of chapters returns to the centre 2026-09-01 22:16:25 +03:00
851caeb449 Retract the overstated headline: our own topology measurement shows fidelity holding or rising, and our editors add no mass at all 2026-09-01 22:14:02 +03:00
ad353744ed Answer the orchestrator in the report: our multiplier does not retire the ratified one, it measures full cost where that measured price alone 2026-09-01 22:09:07 +03:00
8e0c71b7df Point the guard anchor at the line that actually carries the token 2026-09-01 22:05:07 +03:00
9fb5a6f1a7 Retract the fix I had just recorded: the missing reasoning capability is a deliberate guard under test, so describing it would strip a defence rather than patch data 2026-09-01 22:04:49 +03:00
77893978b8 Record that the reasoning knob reaches the model on the polygon wire but the engine never sends it, because the model registry declares no reasoning capability 2026-09-01 22:03:10 +03:00
2d5ee71375 Make the mvp-plan anchor verifiable with an expectation token 2026-09-01 21:48:12 +03:00
74a16138fc Land the consilium on research/29: the editor buys fluency at the cost of fidelity, our panel is blind to that axis, and the zh edit ratio is four times the English one we had been quoting 2026-09-01 21:48:00 +03:00
8d4dc3b5f7 Rewrite the research prompt against the live orchestrator norm and fix what the refuter found: the zone belongs to the orchestrator and one hard-contour fact contradicted its own source 2026-08-30 23:55:54 +03:00
ebff194631 Write the harness-topology research prompt to orchestrator norm: a blocking repository sync precedes any search, because this phase proposed already-finished work five times in one day 2026-08-30 23:36:42 +03:00
df5a79b33c Make the four new anchors verifiable: expectation tokens where the target is a file, note number where the target is the decisions log 2026-08-30 20:15:12 +03:00
cdc5a5f156 Land stage 0 item 1: the bank holds its terms, and the audit found on disk a production-profile run, a July measurement of the same question, and a ratified editor-price input this phase never executed 2026-08-30 20:10:54 +03:00
62f8f6bef9 Land the consilium: what the editor does to the text, eight retracted conclusions with the class of each error, and a staged plan that starts free 2026-08-30 19:16:55 +03:00
c3c8e3100e Land the second audit: my own corrections had left both live claims standing, mixed price bases, and erased the editor defect-contraction that the phase did measure 2026-08-30 13:30:40 +03:00
f1fd873f00 Retract the contraction argument: the rank metric cannot fall below one, so twin divergence measured the instrument floor rather than the editor 2026-08-30 13:23:51 +03:00