Archive & retracted
What was claimed, and what was taken back
Section titled “What was claimed, and what was taken back”A retraction is only useful if it is findable, so this section exists. Two of its pages are slices of the agent rulebook — §3.2, every retracted claim, and §3.3, what is broken or open — because that register moved out of README when README became a front page, and a register of retractions belongs to the rulebook (§1.4) rather than to a marketing page. They are slices, not copies: there is still exactly one §3.2.
The quality bar itself is in AGENTS.md §2.6 (the anchor readings, and the statement that no held-out number in the archive beats a 5-gram byte counter).
The four audits below are audits of the tree at a date, kept whole because their
value is precisely that they are a point-in-time reading with every number
attached. What is in this section’s canonical directory, and the rule that
nothing in it may be cited as evidence, is the archive’s own README
(what is in here) — docs/archive/ also holds the
pruned history the site does not publish.
- Verdict audit, 2026-09-25 —
docs/archive/audit-2026-09-25.md - BROKEN or KNOWN-WRONG —
AGENTS.md - Сводка находок: что сломано, что подтверждено, что не проверено —
docs/archive/findings-2026-09-29.md - fused/ grad-coverage verification (phase 1 item 1) —
docs/archive/fused-verification-2026-09-23.md - Retracted claims —
AGENTS.md - The fused RMSNorm CUDA kernel: it lowers now, and there is a second defect behind it —
docs/archive/rmsnorm-kernel-2026-09-30.md - What is in here —
docs/archive/README.md