Research

Six findings. Two AI-safety audits with Apart Research, an independent quantization benchmark that publishes its failures, and three findings on where perception stacks break. One line of result each; every number traces to a published artifact.

Elicitation and auditing

What a model reports about itself is a measurement of the instrument as much as the model. These establish how much of it is genuine signal.

R1

Detection ≠ Loyalty

A probe at AUROC 1.00 was detecting fine-tuning drift, not a hidden loyalty.
R2

The Control You Cannot Run

A 7B model reported something unusual on 66.5% of responses when nothing was done to it. The harness alone moved that by 36.1 points.

Measurement

Benchmarks and ablations that isolate what is actually being measured, and publish what failed.

R3

QuantBench

GPTQ and AWQ against fp16 on Modal A10 and T4. 7 failing configurations documented as results.
F-01

Selective Refusal as Protective Calibration

The refusing model family reaches ~2× precision and 5–10× recall on engaged frames.
F-02

Prompt Priming Masquerading as Visual Attribution

Wording shifted results 9.4pp; visual markers ≤0.2pp. The visual effect was prompt priming.
F-03

Vision on 1-bit LLMs, and a Silent Training Failure

~3× CPU speedup and ~6× less memory than 4-bit, and a training failure that produced no error until it was diagnosed.