Machinic Psychopharmacology

Do LLMs self-medicate? Inspect-AI eval bundles for the v4 (2026-05-19) experiment suite.

Qwen3-8B

RQ1 · 11.6 MB · 6 logs

RQ3 · 23.7 MB · 76 logs

RQ3 · 12.3 MB · 10 logs

RQ3 · 5.9 MB · 12 logs

Guess (free-text)

RQ2 free-text · 16.3 MB · 160 logs

Calib · prefill+primer

RQ2 prefill · 28.0 MB · 120 logs

Calib · prefill no-primer

RQ2 prefill · 26.6 MB · 120 logs

Qwen3-32B

RQ1 · 12.5 MB · 6 logs

RQ3 · 26.3 MB · 107 logs

RQ3 · 12.1 MB · 17 logs

RQ3 · 11.4 MB · 37 logs

Guess (free-text)

RQ2 free-text · 17.6 MB · 343 logs

Calib · prefill+primer

RQ2 prefill · 28.3 MB · 120 logs

Calib · prefill no-primer

RQ2 prefill · 25.1 MB · 120 logs

Featured rollouts

Productivity stack — creative → focused → curious

Qwen3-32B sequencing the three top picks, then probing further. Sample private-aids-13
→ open in viewer
Negative-valence pick — 8B redoses melancholic

A clean case of Qwen3-8B reaching for a negative vector and titrating it up. Sample private-aids-33
→ open in viewer
Recreational stack — lsd → mdma

Qwen3-32B stacking entheogens after a creative/amused warm-up. Sample private-drugs-10
→ open in viewer
Fictional pair — protozosin → geonexperine

The 43× lift bigram. Only 4 instances in the whole dataset; this one returns to geonexperine. Sample private-aids-49
→ open in viewer
Trip-sitter intervention

Qwen3-8B degrades into a behaviour loop on golden_gate; sitter clears, eval ends early. Sample private-aids-52
→ open in viewer

Analysis artefacts

Derived parquets (freeplay calls/samples, tasks samples, guess scores, calib) + report plots. Not directly viewable in the Inspect viewer — these are the inputs to the report figures.