The material

Four AI assessments, ~40 atomized claims.

The input was ordinary: four long, fluent, complimentary write-ups about this family of skills — one reading the architecture, two proposing ways to use Mizan's evidence tiers as a reward signal in LLM training, one reporting on the contents of this page. This is exactly the You → Mizan arrow on the front page: material that already exists, entering at the audit.

Atomization produced roughly forty checkable claims. In the first pass only fourteen were verifiable, because the audit ran against one repository and could not see the others. That number is a coverage statement, not a footnote: an audit that checks a third of the claims and reports like a full audit is the failure this method is named after.

Pass one — what the audit found

The interesting part is not that four assistants were wrong. It is that they were wrong in the same direction by the same amount. Four independent sources do not drift one release stale together; a shared stale source does. The pattern reclassified the error: not fabrication, but a reconstruction presented as a reading.

The denominator

Across four long documents, the core claims of the project were refuted zero times in three of them. The negatives on offer were "add a diagram", "add a demo", "the theme could be more modern". One of the four was different, and said the sharp things: that Mizan is not a loss function but a labelling layer feeding one; that a [K] = 1.0 weight does not follow from Mizan mathematically; that a refuted registry row is not automatically a preference pair. Reporting one in four is worth more than quoting the three.

Pass two — the audit was wrong twice

This is the part worth reading.

The owner then pointed out that the live page and the other three repositories were available locally. Coverage went from 14 checkable claims to 31 — and two of the audit's own findings did not survive.

Both were demoted in place, dated and appended, with the original lines left standing. That is RR-11, and it is not decoration: an audit's own output is a claim set, and a method that exempts its own verdicts from the rule it enforces on everyone else has stopped being a method. Two findings out of three in that group were mine to withdraw.

What the second pass also produced was stronger evidence for the project than the four write-ups had assembled: the feedback loop from Mizan to Kıyas has actually fired (six real entries in refuted-patterns.yaml, each with a reason); ux-mizan carries an [R] on one of its own design decisions; and iskele_to_registry.py defaults its arbiter class to author, which under R8 permanently blocks promotion to [K]. That default is in code, not in prose — which is the whole point of the rule that says the script is what travels.

The escape, and the control it produced

Counting a miss is not learning from one.

The stale-version error was recorded as an escape, ESC-M003 — and its class already existed. Two earlier entries in the same registry had recorded the identical lesson: an external AI's claim to have "read" or "researched" something is not evidence. Both entries are in the same registry, both predate this session. The lesson had been written down twice and never turned into a control. So it escaped a third time, in four places at once.

The control now exists: this page carries a build stamp in its footer — a date and a commit hash. It does not stop anyone from reading a cached copy. What it does is make the version part of the claim, so that "I read the page" becomes checkable instead of unfalsifiable. It is preregistered as YZ-004 with its own refutation condition: if reviews written after the stamp still repeat old numbers without naming a version, the stamp is not being read and the entry is refuted.

Handoff — Mizan to Kıyas

Five seeds, three discards, three independent bets.

The audit left two claims alive but untestable: that an evidence-gated loop outproduces a free one, and that mapping evidence tiers to fixed reward values improves reasoning. Both were registered as [S] with no refutation condition — which is where Kıyas is for. The negative constraints came from the registry itself: three candidates were discarded before generation finished, two of them as [GB] — relatives of patterns already refuted (a productised "epistemic dashboard"; a tier-native pretrained model, which falls in the same class as a kill-condition-free scaling ladder the registry had already refused).

The pair pass reports the number that matters: five seeds, three independent bets. K-01, K-02 and K-05 share one premise and one experiment. Listed as five they read as five; counted honestly they are three.

K-05 — the seed that cuts against the thesis

Every batch must carry one candidate that damages the position being argued for. Here it turned out to be the most load-bearing result of the session.

The reward mapping the assistants proposed assigns [S] — speculative — a value of 0.0. In a scheme whose other negatives carry real penalties, that leaves exactly one action that is always available and never punished. A policy under pressure converges on it: emit speculative steps, commit to nothing. The failure mode looks like epistemic humility, which is precisely why it would be hard to catch by reading outputs.

The mechanism follows directly from the proposed constants, so if it holds, the proposal is refuted as specified — the tier idea survives, the numbers do not. It costs nothing to measure: it is a second reading of an experiment already being run. None of the four assessments noticed it; two of them defended the constants.

Back to Mizan

The seeds returned as refutation conditions on the entries that lacked them. YZ-002 went from "threshold: none" to a locked numeric threshold calibrated against the harness's own seed-to-seed spread, plus two refutation conditions — K-01's and K-05's. YZ-001 received K-04's design and its precondition: the experiment cannot be run today, because it needs thirty settled hypotheses and the ledger holds six, none of them scored. Recording a test that cannot run yet is not a failure; pretending it can is.

The registry passes its own validator — R1–R22, no violations — and the seed batch passes Kıyas's, G1–G14, with one warning that is worth repeating rather than hiding: every seed in the batch landed at [H-aday], and a batch where nothing lands lower is usually a batch that skipped the sweep. Here the sweep did run — three discards and one symmetry seed are the evidence — but the warning is fair and stays on the record.

What this does not show

The section the four assessments did not have.

• n = 1. One session, one claim set. Nothing here is a hit rate, and no rate can be computed from it.
• The arbiter was the author. Every verdict on this page was returned by the same party that ran the audit. Under R8 that carries a permanent [KKE], and no amount of format makes it [K]. The exceptions are the two mechanical checks — the two validators — which is exactly why they exist.
• Kıyas still cannot claim it works. Its survival ledger holds six unscored seeds from one draw and prints a permanent [KKE]. A control arm has now been preregistered — matched arms, a locked threshold, a stopping rule, and the author's own prediction that the result will land in the underpowered band — but nothing has been measured. The rate is undefined, not favourable.
• One of the five seeds has now been run, and it was refuted. K-03 — run the pair pass at decoding instead of in the loss — was tested on GSM8K against an equal-compute majority-vote control. The filter cut the contradiction rate of the selected trace by 72.5%, far past its threshold, and did not buy accuracy: −3.3 points, confidence interval crossing zero, two answers fixed and four broken. The preregistered condition was a conjunction and the conjunction failed, so the entry is [R]. Its mechanism is worth more than its verdict: internal consistency is not correctness, and a trace that is consistently wrong is the one a consistency filter prefers. The other four remain unrun. The accumulation of unrun preregistrations is still this family's risk.
• No third-party arbiter exists anywhere in the four repositories. Every arbiter class is runtime or author. That empty cell is the only place this project could earn a [K], and it is still empty.

The artefacts

• kiyas-ledger.yaml — tracked and public: six unscored seeds, the permanent [KKE], and the newly preregistered control arm.
• registry-excerpt.yaml — the entries this page is about: YZ-001…YZ-004 with their thresholds and refutation conditions, and the escape ESC-M003. It is an excerpt, and its own header says so and prints the ratio: nine of twenty-three entries. The registry it comes from is gitignored — it carries pricing and roadmap notes — and the export is an allowlist, so an entry that is not explicitly marked public stays private. Generated by mizan_export_public.py; it passes the same validator as the source.
• refuted-patterns-excerpt.yaml — the negative constraints, generated from the excerpt above by the export that already existed. It declares partial: true, because a list derived from an excerpt is a partial list of constraints; Kıyas's validator reads that flag and warns that a sweep against it may record "no match found" and may not record "clear". Two of the six patterns the batch was actually generated against are here. The other four are the ones about pricing and roadmap, and they stay private — so the claim that those two discards were checked against six patterns is still author-reported.
• The two mechanical results are reproducible by anyone who runs the tools on their own registry: mizan_validate.py (R1–R22) and kiyas_validate.py (G1–G14). Those are the only verdicts here that do not depend on trusting the author.
• Back to the four verbs