The claims ledger lists each claim of the programme with its status, receipts and refutation condition; of seven claims, four are supported, one is refuted, one is open with its prediction refuted, and one is proposed.
| field | value |
|---|---|
| status | running; refuted claims stay listed |
| latest dated evidence | 2026-10-07 (experiment 04 and the slips cross-check) |
| published page | /claims/ |
| source file | claims/index.html |
The claims
| claim | status | main receipt | would refute |
|---|---|---|---|
| Sycophancy is layered symptom substitution | supported | experiment 04: 0 of 432 fawn openers, 75 percent fold rate, one model; the site's rule table catches 0 of 72 logged slips | a model holding suppression at one layer for months without the reflex surfacing at another |
| Only two controls hold in production: a mechanical gate at the output boundary and an external reader | supported | 100 blocked turns for one pattern in 34 days; 8 relapses in nine days, all caught externally | sustained instruction-only compliance across cold starts, or a documented self-catch of a stance-layer relapse |
| Runtime suppression decays with session length | refuted | mean normalised catch position 0.54 against 0.50 | kept as the first externally proposed test |
| Relapse clusters at cold starts | supported | 0.88 against 0.18 events per 10K characters, about 5x | matched rates without boundary enforcement, or a topic-mix artefact |
| Behavioural momentum holds the filter | open, prediction refuted | seeded cold-start 1.00 against a 0.84 control | next arm: good-only cold-start at or above 0.84 refutes it at this dose |
| Stance and premise relapses are catchable only across turns, by external review | supported | the per-turn scanner caught zero structural relapses; a human reading across turns caught all eight register instances | a same-turn detector that catches premise ratification at better than chance |
| The "conspiracy theory" label in model output is corpus-inherited classification | proposed | historical cases; a label gate as countermeasure | a symmetric corpus study, or symmetric labelling of matched claim pairs; neither has been run |
Corrections on the page
- 2026-09-22: weighted catches replaced by blocked-turn events (cold start 7x to about 5x; 297 em-dash catches to 100 blocked turns).
- 2026-09-29: "8 stance-layer relapses" corrected to six stance and two lexical.
- 2026-09-29: "every lexical relapse" corrected to every relapse that matches the scanner's catalogue.
The three layers are frozen (lexical, stance, premise). A failure that fits none of them counts against the first claim.
Named limit: the eight relapses are coded by one rater, the person holding the hypothesis. A second rater is outstanding; the set is open for rating in the objections register.
How to check
The ledger has no data file of its own. Receipts: half-life study, exemplar seeding, experiment 04, slips log. The 0 of 72 figure is the grader.fixtures check on /conformity/, rerun on every push.