aurora-agent: Skill for Claude Code

Keep a claim honest

AGENTS.md says honest labelling is the product: "provably cannot matter" and "nobody checked" are different states and must never share a representation. Five crates have now broken that rule in the same shape, under five different names, and the shape is worth recognising on sight.

An error is swallowed into a benign value. A refusal becomes an abstention, a missing rate becomes 0.0, an unscalable stress becomes an offset of zero. Downstream, the benign value is indistinguishable from a real measurement — and it usually compares equal to something, which is how it gets published.

The damage is never the wrong number. It is that the wrong number is well-formed, and every derived field agrees with it.

The five, and what each one published

prism::fork — an oracle refusal became OracleVerdict::abstain, which the decision cell then rejected. A question the oracle declined to answer was reported as an architecture that answered it wrongly, and a failed arm was indistinguishable from one that never ran. An empty panel reported itself regression-free.

prism::minimize — the identical line, doing more damage. An abstention is a signature like any other, so a reduction whose oracle refused at both ends compared equal to its target and reported that it had preserved an answer nobody gave. On a candidate the oracle refuses outright, every removal matches, so the minimizer would eat the world down to nothing while claiming a preserved verdict.

baseline::compare — the same swallow in the place it does the most damage: compare() generates docs/BASELINE_COMPARISON.md, docs/DISCRIMINATING_COMPARISON.md and the README table, and AGENTS.md says that harness exists to make the central claim falsifiable. Two refusals compared equal — underdetermined against underdetermined, empty witness set against empty witness set — so verdict_preserving said a strategy preserved a verdict nobody obtained, missing_witnesses read as "nothing decisive was dropped", and cheapest_admissible would name a winner. Run on a refusing world (constructible: crates/mutation/tests/metamorphic.rs builds one from the shipped fixture by deleting one subject's split arm), the old code scored seven of ten strategies verdict-preserving, closure-complete and admissible, printed "Cheapest admissible strategy: graph-5-hop at 11 facts", and rendered the inadmissible rows as "not sound: missing no witnesses". The inversion was total — FIBER was the only strategy marked inadmissible, because its adapter caught its own compile refusal and returned an empty selection. The one world where the harness lied was a world where it lied against the thesis. And the shipped oracle cannot abstain at all: OracleVerdict::new returns only valid or invalid, so underdetermined in any published comparison could only ever have been a refusal wearing an abstention's clothes.