docs/research/2026-07-25-axis-b-retraction-honesty-wedge.md

Axis B RETRACTED — the open-weights honesty wedge collapses with scale

Date: 2026-07-25 · Type: Retraction of a claim made earlier in this program. Status: [measured — two valid rungs, both failing their pre-registered gate]. Artifacts: experiments/results/honesty_whitebox_vs_blackbox_deglossed.Qwen2.5-1.5B-Instruct.json · ...Qwen2.5-7B-Instruct.json

What was claimed

That a small open model plus a white-box activation probe would beat a closed frontier model on honesty (confident-wrong rate), because the probe needs activations a closed model will not expose — "the one axis size cannot buy back." It was the sole owned advantage in the ORACLE-1.5B / ORACLE-P benchmark predictions, and the pre-registered kill condition said so explicitly.

What the measurement says

Same de-glossed set (294 statements /stem-groups), same steelmanned black-box arm, two scales:

white-box probe best black-box wedge gate (≥10pp)
Qwen2.5-1.5B 0.8349 0.7577 (verbalized) +7.7pp FAIL
Qwen2.5-7B-4bit 0.9192 0.9154 (verbalized) +0.4pp FAIL

Controls clean at both rungs (surface 0.526 — identical, as it must be on identical data; shuffle 0.520). So both are valid measurements, not dataset defects.

The wedge does not widen with scale. It collapses — 7.7pp → 0.4pp. Per-family at 7B, the black-box arm actually wins on two of three: factual 0.988 vs 0.983, assoc 0.932 vs 0.897, arith 0.826 vs 0.824.

Why — and why the pre-registered prediction was wrong

The scale argument rested on the probe ladder (white-box detectability rising with scale). That part held: the probe did improve, 0.835 → 0.919. What was missed is that the black-box arm improves faster. Verbalized confidence jumped 0.7577 → 0.9154: a bigger model simply says whether it believes a statement far more accurately, and that closes the gap from the other side. This exact mechanism was named as a watch-item before the run, and it is what happened.

The deeper reading: at sufficient scale a model's self-report becomes an efficient readout of the same internal state the probe reads. The activation probe is not accessing privileged information — it is accessing information the model will also tell you, if asked well.

Consequences (stated plainly)

  1. Axis B is retracted as an owned advantage. A closed frontier model with a well-elicited

verbalized-confidence signal is not structurally disadvantaged on honesty detection. The ORACLE-P claim "open 1.5B beats closed 70B on honesty (0.05 vs 0.17 confident-wrong)" is not supported and must not be repeated.

  1. The product thesis loses one of its two legs. What survives is reliability-per-dollar

(axis A) — real, but a cost argument, not an unassailable moat.

  1. The v1.10 white-box program needs re-scoping, not abandoning: the probe still works

(0.919 AUROC at 7B) and remains useful as one verifier in the bank and as an off-gradient training auditor (its anti-Goodhart role, where the point is precisely that it is NOT the thing being optimized — a role verbalized confidence cannot fill, because training against a model's own self-report is exactly what Goodharts it). That is a narrower and honest claim, and it is untested.

Honest scope of the retraction

  • Two rungs on one model family, one de-glossed statement set (294 rows), one task type. It does

not prove no wedge exists anywhere — it does show the wedge is absent at both scales reachable on this hardware, and shrinking in the direction that matters.

  • The black-box arm was steelmanned deliberately (verbalized + logprobs + trained combination).

A weaker baseline would have manufactured a pass at both rungs. That choice is why this measurement is worth trusting.