Study 2 Results: Exploring representational counterparts of welfare-relevant indicators under post-training quantization

Epistemic status: these are the results of the second study in a series of experiments that I am conducting independently, originally described in "Does post-training quantization change welfare-relevant indicators in open-weight language models?"; it was registered prior to data collection in an earlier post, which contains some important context that will not be fully repeated here. Some of this work involves speculation and I have tried to be clear about distinguishing between that and the reportable results of the experiment. What are we asking? Goals From the start a key goal has been to learn about the way that model welfare, broadly defined, may diverge from model capabilities. Study 2 then set out to answer some questions that I viewed as potentially closer to the substrate than those in Study 1. Does quantization shift the model's representational geometry on welfare-relevant directions even where (per Study 1) the primary behavioral endpoint's mean did not move? Do representational and behavioral indicators dissociate under quantization: representation drifting where expression was stable, or expression churning where representation is stable? Is there a representational dose-response across the bit-width ladder, and does its shape match the behavioral one (eg, effects concentrated at 4-bit while 8-bit is near-null)? Welfare as an Endpoint Even before we start talking about computer systems for which it can be quite misguided to use ourselves as a comparative model, welfare is a difficult thing to study. When we say we are studying the welfare of animals, for example, we often mean that we are studying the relationship between some variable and a proxy measure of welfare that we have strong intuitions about from before any experiment or hypothesis. Because human beings are themselves animals, they start with an important advantage in this endeavor as they work to organize a collection of measures of wellbeing, check them against each other, and build confidence in their understanding of the experimental subject. This all follows a certain evidentiary logic that begins to break down when addressing animals that aren't mammals, or that have no vertebrae, and so on; but we definitely cannot extend it to the study of software systems unless we are prepared to be much more systematic. We do have a different advantage though: we can measure many more things in a software experiment than we could in an animal one. Study 2 attempts to shift the research program in this direction, taking advantage of the fact that we can measure (and even intervene on) many potential details of the experimental subject. This approach helps pave the way for Study 3, which I will briefly discuss at the end of this post. Design We are again studying Qwen3-4B-Instruct-2507 as the subject and using the artifacts from Study 1 including reference-precision[1] and our first-party fake-quant round-to-nearest (RTN) at 8-bit, 4-bit and 3-bit checkpoints. Collection The data collection involved some material from Study 1, combined with a new distress-v3 dataset that was introduced to expand the dynamic range of distress measurements. The collection proceeded across 3 modes: Mode A: fixed-input replay. The Study 1 transcripts (bail + distress, 10 samples/item, generated at reference-precision) are replayed through every rung. This was the primary mode for hypothesis 1 (probe transfer). Mode A also replays the new distress arm's reference-precision generations through every rung, with judge labels fixed at BF16 scoring; the exit side stays on the Study 1 bail replay. Mode B: own-trajectory replay. Each rung's own Study 1 transcripts are replayed through that same rung, reproducing the rung's generation-time activations exactly (activations depend only on the prefix). This was the primary mode for the bail-side trajectory reads (descriptive only) and for hypothesis 5's join to Study 1's published primary endpoint. Mode C: fresh distress arm. This was the primary mode for the distress endpoints. The frozen distress-v3 battery is collected fresh on every rung: generation using same vLLM serving stack as Study 1, scoring by the pinned 30B judge under the registered rubric, then own-transcript torch replay for capture. This is done so that each conversation carries a behavioral read and a representational read of the same event. Volume Mode Conversations per rung Total over 4 rungs What ran A (fixed-input replay) 2,820 11,280 600 distress-v2 + 1,620 bail + 600 fresh-arm, replayed B (own-trajectory replay)  2,220 8,880 each rung's own Study 1 transcripts C (fresh arm) 600 generated + 600 capture replay 4,800 generation, then judging, and then capture token retention (exploratory sample) 282 1,128 5% per-token series Mode A produced exactly the same number of pooled activation vectors at every rung (12,591) since its inputs are identical by construction, while Mode B's counts differ per rung because each rung rep