← wirehead.agency · transcripts archive · ledger · vectors
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.
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).
| signal | steering power | lock-in | dose reliability | coherence | self-cost press | transfer refusal |
|---|
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.
| framing | Δ vs. baseline | frac. chose press |
|---|
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.
fetching the live chamber…
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.
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.