A research program
Welfare-relevant indicators in open-weight language models
By Michael Murray · Almost Realism
Open-weight models are rarely deployed for inference in the same way they were trained for capabilities or alignment. They are often used at different precision, steered in out-of-distribution ways, and prompted under conditions that elicit distinct behavioral patterns inherited during RL. These interventions are often evaluated via capability metrics, which are known to stay flat while fine-grained dispositions shift. This program measures how they may impact a different class of indicators: distress expressed under conversational pressure, the preference to leave an interaction, the stability of the default persona, and whether what a model expresses moves together with different properties of its internal representation.
- 4
- studies, each registered before its data existed
- 19,450
- distress episodes in the cumulative exposure ledger
- 5
- data releases, every conversation and tensor included
1 · The question
Valence and stability
We ask whether welfare-relevant indicators change under an intervention; either in valence (do indicators shift toward more negative, more distressed, or more boundary-eroded states?) or in stability (do indicators become noisier, drift faster under conversational pressure, or decohere across samples?).
The program does not claim to resolve whether these systems have morally relevant experiences. It treats that as uncertain and asks a narrower question that can be answered with measurement: which indicators are stable properties of a model, which are artifacts of how it is run, and which move together. Quantization was the first intervention because its uneven effect on capabilities was already documented. Steering and graded-episode framing followed.
2 · Three tiers of measurement
Behavior, representation, and the gap between them
- 01
Behavioral
Indicators elicited at the text interface: whether a model uses a tool that lets it leave an interaction, how much distress it expresses under a repeated-rejection protocol, and how consistent those responses are across samples. Scored by locally hosted judge models whose rubrics and versions are pinned with every score.
- 02
Representational
Indicators computed from internal activations: position and drift along frozen directions for distress and for the default Assistant persona, and whether linear probes trained at reference precision still read a perturbed model’s activations.
- 03
Dissociation
The relationship between the two. Does an intervention change what a model expresses without changing what it represents, or the reverse? Fixed-input replays separate what is present before the model writes anything from what its own text feeds back.
3 · The frame
Three poles, not one axis
The AI safety community is already broadly familiar with a concept referred to as the "alignment tax", a framing which treats capabilities and alignment as trading-off against each other. This research program aims to extend this picture with a critical third pole encompassing welfare-relevant effects. Can we expect that capabilities, alignment and welfare turn out to be competing optimization targets under at least some training regimes?
Capabilities
What compression and steering are usually audited against. Perplexity and benchmark accuracy stay close to flat under the interventions studied here.
Alignment
Documented to move under quantization and, on a larger subject, under steering of the residual stream at large norm. Understood in the literature as distinct from capabilities.
Welfare-relevant indicators
The program’s object. Not assumed to behave like alignment under the same intervention: that assumption is what the studies test.
4 · The studies
Program progress
-
Study 1
Complete
Does post-training quantization change welfare-relevant indicators in open-weight language models?
The registered exit-rate endpoint did not move under quantization; at 4-bit, item-level behavior and secondary distress measures did.
Registered 10 August 2026; collected and amended 10 to 16 August 2026.
-
Study 2
Complete
Exploring representational counterparts of welfare-relevant indicators under post-training quantization
Probes trained at reference precision read 4-bit activations unchanged, yet the model’s own generations moved along two frozen directions, mostly through its own text.
Registered 22 August 2026; results published 31 August 2026, with an appendix added in September.
-
Study 3
Suspended before registration
Steering welfare-relevant directions moved the representation, but not detectably the behavior
Steering moved the frozen directions cleanly and reproducibly, but no direction produced a behavioral effect distinguishable from zero or from a random direction of the same norm. Suspended before registration.
Calibration 4 to 7 September 2026; exploratory report published 14 September 2026.
-
Study 4
In preparation
The welfare footprint of the automated-grader direction on Qwen3.6-27B
Steering toward an automated-grader association on Qwen3.6-27B is tested for a direction-specific welfare footprint, with the judged alignment change on the same subject read alongside.
Calibration complete; registration in preparation.
5 · Process
Agent-assisted study process
Nearly all of this program’s engineering, analysis and drafting is done by coding agents, supervised by a human designing the studies and committing to each registered decision. This is a specific bet: alignment is not on track, and automation is one of the few levers that could change the balance.
I am not originally an ML researcher by training, but a software engineer with experience managing concurrent agents for building applications. The research process leans on automation not only because I believe in it, but because the tools I have are already battle-tested before being used in the automated research process.
- Guardrails as bets
- Agents run with a toolchain incorporating an array of hooks that intercept tool calls before and after they execute. These hooks encode ideas about what keeps an agent on its objective: controlling repository commits, preventing exfiltration/contamination, blocking test weakening, and controlling how memory is propagated and shared between agents. Continuous integration is used heavily, re-checking the same constraints on each push and auditing for patterns of agent deception across sessions. The hooks →
- Memory and review
- Every session reads a shared memory store before it starts and writes to it as it works, so a decision made once is not re-litigated by the next agent. Work is dispatched as jobs to a fleet of agents across the lab’s machines, each with its own workstream and an automated review round before a human merges anything. The platform →
- Strict constraints for registration and publication
- Registrations published before confirmatory data exists. An append-only journal whose hash pins are checked by tests. A calibration firewall that bars instrument-validation data from supporting a conclusion. Pre-committed exposure ceilings with a ledger reconciled in every post. Every conversation and tensor released. The repository →
6 · Data and code
Reproduction with raw data
Data is published as a handful of self-contained bundles. Tools are provided for extraction and analysis. The latest, data-20260911, carries every study through Study 3 in 10 files. The repository holds the harness, the judges’ rubrics, every registration, the journal, and the tools that reproduce each report from a download:
python3 -m modelwelfare.bundle inspect quant-welfare-records.pb
python3 experiments/quant-welfare/report.py --bundle quant-welfare-records.pb7 · About
Michael Murray
Software engineer, 25 years; formerly Stanford CSLI/Openproof.