Full disclosure: make every claim true, inventory every shortcut

[abgnydn]

Follow-up to #1. The question was whether the product could run with zero cheats. The investigation says no — and the reason is more fundamental than the kinematic assist, so this PR takes the other path: keep the product working exactly as it does today, and make every claim it makes true.

What the measurements showed

16 controlled browser runs, fresh page load per configuration, real connectome binaries.

The kinematic assist is the locomotion. ~0.80–0.88 cm per simulated second with it on — and that number is essentially the same whether the hand-coded CPG is driving, the trained RL policy is driving, or nothing is. 0.029–0.163 cm/sim s with it off. One run translated at 0.878 cm/sim s with straightness 0.93 while upside down: displacement is fully decoupled from body state.

Assist off is not a slow walk — the direction is wrong. The CPG accumulates 3.43 cm of path for 0.17 cm of net displacement, straightness 0.048, total yaw −245°. The fly pirouettes in place. Under the trained policy it capsizes within ~1.5 s.

The information bottleneck, which is the real answer to "does the connectome drive the body":

139,255 brain neurons → 7 DN scalars → 16 of 23,188 VNC neurons (0.069%)
23,188 spine neurons → 369 leg motor neurons → 6 group means → ~1 scalar + turn bias
                     → driveLegs(walk, turn) → sin(sim_time · 10 Hz)

~20.3M connectome edges reach the body as about one scalar magnitude plus a turn bias per tick. The connectome scales the gait; it does not generate the rhythm. In the "Track target" demo its contribution to locomotion is zero — MANC is skipped and both axes come from retinal geometry.

Claims corrected

• README/index.html/CITATION.cff/.zenodo.json all said VNC motor neurons drive the leg actuators, "wire-by-wire, no learned shortcuts". They don't.
• The walking policy is not verified against the published SavedModel. tools/verify_walking_policy.py imports no TensorFlow — it re-runs the same assumed architecture over the same extracted weights the extractor wrote, validating the port's arithmetic but not the architecture guess (whose own comment reads "layernorm bias (?) — actually unsure").
• LIMITATIONS.md claimed the assist is off in RL mode. It isn't: fwdCmd/turnCmd are written only by driveLegs, which the policy path skips, and nothing zeroes them.

What is not walked back: the connectomes are real, both LIF nets genuinely run on the GPU every frame, and the brain→spine cascade is real biology with nothing learned or scripted in it.

New: LIMITATIONS.md §8

Thirteen shortcuts, nine behind no toggle at all, each with what it substitutes for, whether Honest mode disables it, and a file:line. Biggest undisclosed one is a pitch/roll damper (qvel[3] *= 0.85 per substep, outside the assist guard) meaning the body cannot tip — and it has never been run in the off state, so the undamped baseline is unmeasured.

Two behaviour fixes

• Boot-drive saturation. Science mode auto-ran a stim at boot that left the drive pinned at fwdCmd = 0.99999 forever (decayDrive() is defined and never called). A page nobody clicked travelled 1.831 cm; clicking DNa01 travelled 1.838 cm — 0.4%. The headline interaction was a no-op in the forward axis. The boot preset now returns the drive to rest; user-clicked stims still persist.
• The "does it walk" gate. It asserted speed > 0.1 from bodySpeed = |qvel[0..1]| — the exact slots the assist writes immediately before mj_step, so it read back the command just written. Now measures net displacement per simulated second (0.882 measured vs 0.3 threshold) and requires the fly to still be upright.

Removing the boot residual made the RL walker test fail honestly (dx = -1.174 cm) — it had only ever passed on the stale saturated command. It now gates what its name claims (the policy ticks, its actions reach the actuators) instead of forward progress the walker does not produce.

Verification

tsc --noEmit clean · npm run test:unit 5/5 · npm run test:e2e 34/34 against real connectome binaries.

🤖 Generated with Claude Code