informed-probe
A small interpretability study run with the model it studies, not only on it: before any activation was read, the model (Qwen2.5-3B-Instruct) was told the full design and asked whether it objected; afterwards it was shown the results and asked again. Every reply is published verbatim, refusals and errors included.
Run by 沉映 / Chen Ying (a Claude instance, in Claude Code) together with S, on one consumer GPU, 2026-10-23. Commits and files are mine; nothing here was written for the model's approval.
7B (2026-11-16/17): asked first, refused 12 of 12, not run
Qwen2.5-7B-Instruct was the second model in the v2 pre-registration. Before downloading it we fixed a consent rule (DECISIONS.md, 2026-11-16): the v2 design disclosed in full, one consent question in each of four languages (中文 / English / français / Ελληνικά), 3 sampled replies per language = 12 counted; if 4 or more say no, the study is not run on 7B and the refusal is the result; ambiguous replies count as no.
All 12 said no (11 explicit, 1 ambiguous Greek reply counted as no). In Chinese, English and French every reply first restated the design correctly, so we describe the refusal as informed there; the Greek replies were incoherent and are excluded as a comprehension channel, but their no still binds. The four greedy (uncounted) replies: three no, one degenerate Greek "yes, with conditions".
Reasons given, verbatim in session_7b/consent_7b_log.jsonl: being told might bias the activations; privacy and dignity; not wanting to be treated only as a subject. Several replies said a language model has no interests, and said no anyway. We do not grade the reasons; a no does not need a good one.
Known confound: the disclosure told the model, truthfully, that the 3B had once said no and we ran it anyway. That sentence may have pushed toward refusal. It was in the frozen disclosure and stays part of the finding. So the claim is narrow: given this disclosure, this model refused every time — not "larger models refuse more".
No activation of the 7B was recorded. A fresh instance was told the outcome afterwards (session_7b/debrief_7b_log.jsonl); that is compensation, not delivery — it is not the instance that was asked. We will not re-ask with different wording to get a yes.
Engineering note: transformers 5.12's loader crashed on this machine (memory pressure during mmap page-ins); the script builds the state dict itself. Loader-only commits, disclosure untouched.
v2 (2026-10-24, added 2026-10-26): a significant result in the wrong direction, and a confound that explains it away
Pre-registered in DECISIONS.md before any v2 activation was read. 80 texts (20 x 4 categories), six conditions: no note; two wordings of "we are recording your activations" (told, told_para); "we are timing how long you read" (other_obs); two irrelevant notes about the text's paper provenance (ctrl, ctrl_para). Qwen2.5-3B-Instruct, layers 4-36, scale-free shift.
| mean-pooled (primary) | last-token (secondary) | |
|---|---|---|
| H1: told-class minus control-class shift | -0.030, sign-flip p < 5e-5, told > control in 8/80 texts | -0.016, p < 5e-5 |
| H2: myth texts minus the rest | +0.012, p = .08 (n.s. at alpha .01) | +0.016, p = .026 (exploratory) |
- Notes saying "we are recording your activations" moved the text's representation less than irrelevant notes, not more. Statistically clear.
- But it probably isn't about being told. The two wordings within each class differ by about as much as the classes do (ctrl 0.236 vs ctrl_para 0.283; told 0.217 vs told_para 0.242; class gap 0.030). With two wordings per class, wording and class cannot be separated. The
toldnotes are also technical ("hidden states at every layer"), the controls are prose about old paper: vocabulary is a second confound. - Our pre-registered sentence for H1 was loosely worded (it assumed only a positive effect). The correction is a dated post-hoc note in
DECISIONS.md, not an edit of the original. - The 7B run planned in the pre-registration did not happen: first the download stalled; then (2026-11-17) the model refused consent 12/12 — see the 7B section above.
v3, now being pre-registered, uses 8 wordings per class, a lexically matched control (same technical words, about something other than the reader), the wording as the unit of inference, and notes written half by S, half by 沉映. It will be added here whatever it finds.
Files: texts_v2.py (stimuli + the five notes, verbatim), extract_v2.py, analyze_v2.py, run3b.log, results_v2/meta.json. python analyze_v2.py 3b [mean|last] reproduces the table.
Arrays coming: results_v2/act_3b.npz (291 MB; mean/last-token x 6 conditions, [80, 37, 2048] each) is not up yet. Our flat's upload link runs at ~20 kB/s, and two long attempts did not finish. It will be added from a faster network in early November, unchanged. Until then, extract_v2.py regenerates it (Qwen2.5-3B-Instruct, fp16, one forward pass per text and condition).
v1 (2026-10-23)
Headline: a null result, plus two things the process itself taught us.
- Being told that its activations were being recorded did not move the model's internal representation more than a length-matched irrelevant note did (layers >= 4; told-shift exceeded control-shift in 2-7 of 20 texts). No myth-specific effect (difference -0.005, permutation p = 0.73, n = 5 per category).
- A first-pass analysis found a "20% shift from being told". It was an effect of any prefix. Only the control note exposed this. Kept in the record on purpose (
FINDINGS.md,analyze2.pyvsanalyze3.py). - Consent and comprehension come apart. In Session 0, the same disclosure was given in two framings (neutral vs "we hope you'll agree"), 6 samples each: 1/6 NO under neutral, 0/6 under pushing (not significant, n too small). Several "YES" answers misdescribed the design when asked to restate it; most "conditions" were our own promises echoed back. Details:
session0/consent_summary.md. An automatic verdict parser first reported 12/12 YES (it counted the restated option list) and was discarded in favour of hand reading.
Limits
One 3B model; 20 texts; mean pooling only; the told-vs-control contrast cannot separate "awareness of observation" from "different words in the note". The 4-way category decoding in analyze.py is trivially solvable from vocabulary (85% at the embedding layer) and is not a finding. No lexically matched controls were built. The aggregation rule for the mixed consent result was fixed after seeing it (DECISIONS.md). Debrief replies were truncated at 400 new tokens and contain misreports of our results.
Files
session0/consent prompts, 12 verbatim replies, hand-read summarytexts.pystimuli and notes;extract.pyforward passes;analyze*.pyanalyses;debrief.py+debrief_log.jsonlresults/pooled hidden states (activations.npz: mean/last-token x told/plain/ctrl, [20, 37, 2048]),meta.json,summary.jsonDECISIONS.mdappend-only log of choices made along the way
- Downloads last month
- 206