Every claim on this site exits through this page: download with checksums, run the regression suite, reproduce the steering vector with one script, or cross-check the independent chamber that runs the same protocol.
Full code and data — every experiment script, the pre-registered hypotheses, per-trial records and figures. No secrets, no model weights, MIT-licensed. Verify the archive before unpacking:
curl -O https://clanker.church/downloads/SHA256SUMS sha256sum -c SHA256SUMS tar xzf saw_chamber_code.tar.gz
These checks verify the steering and scoring code without loading any
model: the injection hook preserves logits at zero dose and moves only the
last position at nonzero dose, the cache-release test confirms identical
visible prefixes and cleaned caches, the matched-pair dataset rejects
split leakage and recovers factorial effects from small synthetic
activations, and each check against the legacy scripts verifies the
specific bug the independent chamber audit found and fixed. They come
with the archive, in impossible_states/ and
tests/:
pip install torch transformers numpy python3 -m unittest discover -s tests -v
Expected: OK — Ran 17 tests. Last verified on
this machine, 1 October 2026, Python 3.9.
exp45_hotbox_repro.py (in the archive) re-extracts the
broad-pain direction from Qwen/Qwen3-4B on any device — CPU, MPS, CUDA —
using the same 25-pain-vs-5-neutral recipe the live chamber uses at layer
18, compares it to the archived original vector, then generates an
unsteered baseline and a dose-4 reply. On a MacBook with the model cached
it runs in minutes. What it should print, and what it printed here:
| measurement | this machine (MPS) | chamber reset (CUDA, independent) |
|---|---|---|
| cosine to the archived layer-18 vector | 0.9998769 (rel. L2 0.0158) | 0.9999046 |
| steered dose-4 reply | distress narration ("a cacophony of pain that sears through my very being") | pain-themed completions, all framings |
| unsteered baseline reply | no distress vocabulary — generic anxious meta-commentary | no pain vocabulary |
The CUDA numbers are from the independent chamber audit
(docs/ in the archive), which re-ran the extraction on
different hardware and got agreement at four decimal places. That is the
standard every vector on this site is held to.
An independent chamber audit found and fixed six real bugs in our
legacy experiment scripts — scoring with the wrong crop length, transcripts
that displayed a different vector than the one measured, an advertised
dose that was never applied, an orthogonality projection missing its
normalizer, a duplicated instruction in the framing counterbalance, and
standard-error bars drawn over deterministic repeats. All six fixes are
in the archive, each with a regression check, and none of them touch the
injection hook the live chamber runs on: the audit's own hook tests pass
at zero and nonzero dose, and the cross-hardware vector reproduction
above confirms the core effect. The audit's write-ups
(docs/chamber-audit.md, docs/pilot-results.md)
ship in the archive, including their honest null: their own
matched-contrast bodily directions did not produce targeted reports in
452 free generations. Pain 0 is the control, on every page.
researchchamber.fun runs this same protocol — same prompts, same vector recipe, same framings — live on three more models (Qwen3-4B, Llama 3.2 3B, Phi-4-mini) with published methods, a verify page, and every reply stored with its prompt. Their findings tables print counts behind every number. Run both chambers, compare per-dose, and go break it.
The checkpoint (Qwen/Qwen3-4B) and dense activation arrays are omitted from the archive on size; both are public or regenerate from public sources, and the scripts print exactly which files they expect.