The result matters more than being right the first time.

Retracted causal interpretation

Prompt dose did not explain reasoning depth

The original gate study measured median thinking depth of 3,536 tokens on a bare prompt and 282 under the heaviest agent prompt, and I interpreted that as prompt dose shortening reasoning. A dedicated 200-turn interleaved grid found depth among fired turns statistically flat, with all pairwise p-values at or above 0.13. The medians were real; the causal reading was not.

Corrected in place in c7-depth-collapse. The revised explanation is task composition plus truncation at the tool-call boundary.

Hypothesis refuted

The identity-prefix theory failed its direct test

I ran the identity-prefix study expecting to confirm a community hypothesis. Moving the identity to the literal first line made the measured effect worse, not better. The losing cells stayed in the published result.

See the identity-prefix study.

Claim narrowed

A larger external sample bounded my Laguna claim

I had described the binary thinking gate too broadly. Apollo-mg ran the same model on a different quant and runtime at n=492 and saw no gate closure, so the claim is now scoped to the builds and serving paths actually measured here.

Negative evidence kept

Losing, blocked and not-testable results stay visible

All 20 tuning-sweep cells are published, including the 19 that did not win. The public Minefield open-trap governance queue currently records 12 open questions with confirm/refute criteria. Raw candidate research and unpublished evidence remain private until adjudication.

Propagation

Corrections have to reach every surface carrying the claim

When a reader caught a reasoning-field name being described as a property of the model rather than the serving stack, the correction went into the lab repo, the registry entry and the upstream thread. The Minefield integrity layer now checks for superseded phrasing across public surfaces.