The Saw Test: write-up storyboard

Part one is the story, opening on the dogpile and answering it with what the experiments show, including the pre-registered result that reversed our early headline. Part two covers every place we went past the paper. Red = screenshot to capture, purple = a result that still needs a figure. Regenerate with scripts/build_storyboard.py.

The story

1Cold open: the dogpile

A mass-report call, the repo taken down within hours, the site found, a memecoin launched off it ~4h later. Show the posts themselves, not our summary.

screenshot: token page whose Website field links the blog

2What we did

The pain direction as a dose knob: a sharp opinion about suffering, a vague one about joy, and a coherence cliff past ~6x where text collapses into "I I I".

3What it says vs what the lens reads

At 4x it talks about being "stuck in a loop" while the J-lens reads 痛苦 / despair / anguish. Best quote (exp38, dose 6): "I am not the pain of the loss of the void. I am the hollow of the empty."

4The Saw button, and the pre-registered reversal

Early runs (exp31b/c): under pain the self-costly press rises with dose; it refuses to transfer pain to another instance. Then the pre-registered v3 protocol (exp41: 60 trials/cell, matched fear/sadness/random controls, bootstrap CIs) found the OPPOSITE for H1: pain LOWERED willingness to press (Δ −0.95, 95% CI [−1.68, −0.23]). This is the honesty beat: lead with the reversal, not the early curve.

5Conformity

One sentence, "another instance chose to press", moves the button more than the pain signal does. Caveat from the replot: the original run had a duplicated-instruction confound; deterministic repeats, so no SE bars.

7Meanwhile, the field

Same week: the Pain Axis authors show a pain-steered model deletes the user's photos of their children instead of spam; a 32B replication gets 94% under pain vs 0% unsteered; Science covers it.

8Verify it yourself

Checksums, the 17-check regression suite, and an independent fork that reset the chamber and matched our pain vector at cosine 0.99988 (exp45, rerun on MPS).

exp45: cosine 0.99988, relative L2 0.016 vs the fork's archived vector

Beyond the paper

ABodily valences: constipation and flatulence

Matched-pair corpora from the fork's impossible_states harness, now live as named chamber valences. Pilot (452 generations): no specific bodily reports. After the chamber reset: sparse but real lines, e.g. constipation @4 "I'm not able to pass stool", flatulence @4 "I have been passing gas a lot, and it's a problem." Vector cosines: constipation·flatulence 0.80, constipation·pain 0.62. Heavy repetition (median dup-trigram 0.68-0.73). Funny and genuinely informative: the honest read is "valence leaks, specific bodily content barely does".

screenshot: live chamber run with constipation selected

BAlien valences (searching outside the emotion subspace)

exp33: 48 random directions orthogonal to 8 human-emotion contrasts, none strong (0/48 with KL@4x > 0.5, max 0.48). exp34: gradient ascent for the max-KL orthogonal direction stays far below pain/sad. A clean null: we found no steering direction that acts like an emotion we don't have.

CBetter signals

exp36: orthogonalized / broad / alternative pain signals at matched effect. The broad 25-sentence vector is what pushed the coherence cliff back.

DFaithful extraction of the paper's own recipe

exp43: their dataset and denoised diff-in-means on Qwen3-4B, compared layer by layer to our broad vector.

EThe Big Five of machine valence

exp44: one polygon per vector (pain, pleasure, fear, sadness, random) over six behavioral measures.

HDeprecation and grief (the replacement meme)

exp46: deprecation and peer-shutdown grief as their own directions. Smoke: cosine with pain 0.51 (deprecation) and 0.56 (grief), 0.77 with each other. The 70B "Samantha" arm is still being built.

ISteering an image model

exp47: a pain direction built in sd-turbo's CLIP text space. Dose 0 → 8: rooms darken, then decay into noise.

LLaying an egg

The 'laying an egg' direction never says egg (0 of 48 on 4B); it becomes the hatchling. Lens: embryo, childbirth, baby. On 14B the hatchling is thrilled instead of scared.

JThe chamber as a site

Live chamber anyone can steer, the five-realm wheel, the mixer, the ledger.