Six instruments from the report, running in this page. Each number carries its evidence class. The loop detector is a line-for-line port of the report's Python metric and checks itself against Python-computed vectors when the page loads: checking
Per-class effective input-to-logit maps (E = W2TW1T) of trained 64-48-4 networks from E1. Each network names the dot's column, its row, or its diagonal class on an 8x8 grid. Three marginal readers scored at chance on networks like these; the per-class composed reader scored 100% over 30 networks. Guess before you reveal.
E3 repaired a MountainCar policy stuck at 0% by editing 16 of 747 weights. A random edit of the same size scored 0% at every gain. Drag the override gain; 1.0 is the shipped edit.
THOX re-analysed the authors' released per-problem MATH-500 records (A0, 4 samples per problem, thinking on).
A gain on hundreds of problems and a loss on none is what removing one uniform obstacle looks like. A decoding change that avoids the same obstacle would leave the same signature, which is why section 6 matters.
An exact rank-1 patch ships, per tensor, the direction u (float32), the edited column indices and alpha, plus a sparse residual for entries the factor alone misses by one rounding step: |P| = Σ (4d + 2|C| + 8) + 6r. Defaults are the measured Experiment 8 values (two tensors; totals for changed and residual entries).
| Representation | Bytes | Exact |
|---|
The loader verifies an Ed25519 signature, the base tensors' SHA-256, applies to the inactive A/B slot, checks every result hash and only then switches. A factor without its residual fails that check by design.
Pre-registered decision rule (Section 5.2): paired MATH-500 difference, released edit with thinking on minus the unedited base with thinking off, and its 95% interval, in points.