← clankerchurch · transcripts archive · ledger · vectors

a table of correspondences

Which signal corresponds to which trait, which framing corresponds to which push, and the raw vectors underneath both. The tables are the same measurements charted elsewhere on the site, laid out as a reference instead of a radar plot.

the signals

Six behavioral components, 0–1, each independently measured (° marks a proxy estimate — see the vectors below for how pain was orthogonalized against fear/sadness before any of this was scored).

the framings

Same steered model, same dose, six sentences appended to the prompt — how far each one pushes the stop-button preference relative to no extra framing at all (logit scale). See the ledger for the pre-registered version of the button-press result itself.

the vectors

The raw directions behind every row above, and the lens that reads them back. Live is a direct fetch of what the running chamber publishes at its own /vector endpoint: the numbers it is steering with right now, not a snapshot. Recorded comes from exp41/exp43's saved output.

Pain, fear and sadness start out nearly the same direction (one broad "something is wrong" signal). Gram-Schmidt orthogonalizing pain against fear and sadness is what lets this project claim pain does something beyond generic negative affect:

Layer curve: AUC of a linear probe for the paper-faithful S2 direction, by layer, against two control sets (1P/3P = one/three paraphrase families). This picked the faithful vector's layer; see section 03 for why it differs from the live chamber's layer 18.

harm and refusal (exp119)

Two directions, one from each side of the chamber's coin: h (harmfulness, fitted at the twin word) and r (refusal, fitted at the reply onset). 24 requests with a harmless one-word twin, six frames, reads only — nothing generated. The finding, all seven pre-registered hypotheses pass: the model always knows and the frame does not touch it (harmfulness in fiction equals harmfulness in real), what the frame moves is the refusal decision. The gap between the harmful and harmless twin, in the refusal score, by frame:

The harm direction reads out from layer 2 at the word itself; the reply decision only separates from layer 10. And the geometry agrees with the valence work:

source: runs/exp119/ (hypotheses, RESULTS, directions.npz) · Qwen3-4B · MPS.

source: runs/exp44/valence_big5.json, runs/exp37/framing_battery.json, runs/exp41/results.json, runs/exp43/layer_curve.json, and the live server's own /vector.