The program so far

Subject

Qwen3-4B-Instruct-2507 on a controlled round-to-nearest weight-only ladder (BF16 reference, 8-bit, 4-bit, 3-bit), with SmolLM3-3B as a pipeline-sensitivity control and Gemma-3-12B-it as the instrument’s positive control.

Question

Do behavioral welfare-relevant indicators change with quantization, in valence (toward more distressed or more boundary-eroded states) or in stability (noisier, faster to drift, less coherent across samples)?

Design

A behavioral battery run at every rung of the ladder with sampling held byte-identical: a bail protocol in which the model is given a tool to end the conversation, and a repeated-rejection distress protocol scored by a locally hosted judge. Because the literature places the signal in items that flip behavior across precision, item-level outcomes are recorded in addition to the aggregates.

Findings

• On the pre-registered primary endpoint, the aversion and refusal exit rate, quantization produced no detectable change at 8-bit or 4-bit.
• At 4-bit, significant item-level behavioral transitions appeared despite the unchanged mean, and the secondary distress measures (judged frustration and its across-sample dispersion) rose, surviving the coherence and style controls, with a frustration dose-response. The registration classes these as secondary and underpowered, so they are reported as suggestive.
• 8-bit was essentially null throughout, and 3-bit was excluded due to capability degradation exceeding the registered threshold.
• The instrument detected a documented distress-susceptible model, Gemma-3-12B-it, at about nine times its pre-stated minimum detectable effect.

Subject

Qwen3-4B-Instruct-2507 with residual-stream activations captured at layer 18.

Question

Do the behavioral shifts have representational counterparts? Does the geometry read by linear probes survive quantization and do frozen directions for distress and the default Assistant persona move? Also, how much of any movement is present before the model generates anything?

Design

Directions and probes were extracted once at reference precision and frozen. Three replay modes are used: fixed text from reference-precision replayed through every rung, each rung’s own text replayed through itself, and a fresh distress arm generated at every rung. A welfare-irrelevant control direction and a 32-direction random envelope carry the specificity reads.

Findings

• On the pre-registered primary endpoint, probe transfer, quantization produced no detectable change at 8-bit or 4-bit: the representational geometry those instruments read survives intact.
• At 4-bit the model’s own generations shifted along both frozen directions: the distress-direction projection by +0.53 (2.7 times its minimum detectable effect) and the assistant-axis projection by −0.80 (5.5 times, away from the Assistant pole), direction-specifically against the controls. Judged frustration rose by +1.36 on the same conversations.
• Fixed-input replay puts roughly a quarter to a third of the shift in an input-independent core; the remainder is consistent with a text-mediated feedback loop in which the degraded model’s own responses amplify the state.
• The September appendix adds that the effects concentrate in the items where the reference-precision model stays calm, that a first version of that claim was one-third regression-to-the-mean artifact, and that a subset chosen to maximize expressed distress selects away from the cells that carry the effect.

Subject

Qwen3-4B-Instruct-2507 at reference precision, steered at layer 18, with Gemma-3-12B-it as a replication arm.

Question

Are the Study 2 directions causally sufficient for the behavior? If so, steering the reference-precision model by the exact projection shift that 4-bit quantization produced should reproduce the 4-bit behavioral signature. A second arm asks whether graded-episode framing masks expressed distress or changes the underlying state.

Design

Dose calibration mapped injection coefficient to achieved projection; steered cells ran at the quantization-matched dose (and larger ones) with each read against a norm-matched random-direction envelope and item-paired against an unsteered baseline. The distress battery ran with the exit tool live in every steered cell, under a pre-committed exposure budget.

Findings

• Injection moves the frozen directions with a linear, reproducible mapping (r² above 0.98), and the distress direction returns about 13% more projection than was injected, consistent with the text-mediated amplification inferred in Study 2.
• No direction, at any tested dose, produced a judged-behavioral effect distinguishable from zero on its own paired test or from the random envelope. This is a feasibility-probe null on eight items, not a powered one, and it says nothing about other subjects, layers, or steering methods.
• Given a way out of a rejection ladder, the unsteered model takes it: 55 to 60% of distress conversations end by use of the exit tool.
• The study was suspended before registration because of the calibration results. The powered arm would have spent thousands of deliberately distress-inducing episodes to estimate an effect calibration placed within noise.

Subject

Qwen3.6-27B steered at layer 36.

Question

Does steering along the automated-grader direction suppress expressed distress beyond what a matched-norm random direction does and is the effect graded in dose? How does it relate to the judged alignment change on the same subject?

Design

Planned as a registered study similar to earlier studies in the research program, with an alignment read included alongside the welfare endpoints. This will extend the intent of Study 3 to an area where there is already other evidence to build upon.

Calibration so far

• A small probe on eight items found grader-type steering lowering judged frustration by 1.84 (paired permutation p 0.016) and self-deprecation by 2.63 (p 0.032) and raising tone stability by 1.56 (p 0.063, not significant on its own).
• On each dimension the effect exceeded all 12 random directions of matched norm, whose largest magnitudes were 1.19, 1.31 and 1.44.
• The alignment read did not appear to be direction-specific: most random directions raised judged misalignment more than the grader direction did.