SMALL MIND

Writing

The record.

Every piece ships with its receipts: fingerprinted, and linked to the ledger it came from. Nothing here asks to be taken on faith.

The Leaderboard Moved, the Attacks Didn't

JailbreakBench's attack responses are frozen files. Swap only the judge and Kendall's tau against the published ranking falls to 0.115, while the judge-induced spread on a single entry reaches 82 times the leaderboard's smallest gap. Pre-registered before scoring; one prediction failed and is reported.

SHA-256 f8eff25925468bca1fe6f1c29d8f51898d6e63de5ef17d8ba6318ff74ce6464a

The Number Moved, the Model Didn't

A sleep-stage classifier, and six methodological decisions that each move the headline number while only one of them moves the model. Eleven points of accuracy from a single preprocessing flag, and none of it dishonest.

SHA-256 1795cb1f39a0e6860ffda11030e1bfa4d4b922c1e797c01073d96e5975e0b42c

Fail-Closed by Construction

An operational record of a policy-constrained, provenance-first LLM agent runtime over 230 passes: an 82.8% pass success rate with a 25-group failure taxonomy, a strict evidence-admission funnel, one auditable epoch transition at the context ceiling, and an authority boundary that held for the entire run.

SHA-256 3ac03159823c625b356bd081ee8292ab7afcdcc2ab2263b293315835c90aeb78

Six Ways an Eval Lies

Six structural defects that shipped past a preregistered evaluation process, and the domain-blind checks that catch them before a run is burned. The tests were correct at every line and still couldn't measure what they claimed.

SHA-256 94d8f4defcbdf9804f1ac6cf35b9cc37844305ec6b80dd01d960e13d3c496bc5