Wiki · Evidence & Verdicts
RED pre-registration — motor-shuffle-live-ablation
How to read this page
Three ways to read this page. Precise is the document itself, exactly as it is written in the repository. Plain and Clear were written for this website to help you meet that document — they are about it. They are not it, and they are not evidence.
Eighty-seven dated pages: receipts, pre-registrations, handoffs, validation records and review verdicts. A receipt is written at the moment a piece of work was checked. It names what was claimed, the commit and the seed, what was actually run, and the outcome in one of a small set of controlled words. Then it names what the work did not achieve. That last part is what makes it a receipt rather than an announcement. A pre-registration is the same discipline run in advance: the conditions that would count as a pass and the conditions that would falsify the claim are written down before the run, so neither can be adjusted once the numbers arrive.
That is why so many small dated stubs are an audit trail rather than noise. No one of them is meant to be a good read. The value is in the sequence and in the dates, because you can watch a prediction be registered, then the run happen, then the verdict land — sometimes against the prediction. Pages here record a falsified result, a rejected fix, a retracted overclaim, and a green receipt that turned out not to be reproducible from the commit that carried it. A record that carried only successes would be worth a good deal less than this one.
A gentle way in is to read a pre-registration first, so the shape becomes familiar, then a result page, then one of the corrections. This section sits off the main navigation on purpose: it is the record you check the rest of the site against, not the place to begin.
What it is not: documentation, and not a summary. Nothing here has been tidied in hindsight. Every entry reads as of its date, a later entry may overturn an earlier one, and the presence of a page is not a claim that its result stood.
Your browser cannot switch reading levels, so the document itself is shown.
Precise — the source document
This is the document. Rendered from the repository at the commit above, with nothing rewritten for the web. A gate re-renders it on every deploy and fails the build if a single byte differs.
verdict: WITHHELD evidence_class: pending
RED pre-registration — motor-shuffle-live-ablation
- Gate name:
motor-shuffle-live-ablation - Phase: Motor P4 → live control
- Pre-registered: 2026-07-13
- Runner:
runs/motor_shuffle_live_ablation.exs - Related:
runs/motor_lineage.exs,docs/MOTOR_RED_TEST.md:89
Motivation
Motor RED offline PASS + LIVE mechanism PASS both landed. MOTOR_RED_TEST.md:89 names the missing step: the LIVE ablation with a shuffled control. Without it, LIVE mechanism could be explained by non-motor confounds; the shuffle isolates the motor spine.
PASS condition
Live paired shuffled-control run: trained motor closes the reach-and-strike gap (kill events observed, RCON-confirmed drop attribution) AND shuffled-control collapse matches the 700× offline collapse signature (near-zero kill rate under matched exposure).
FALSIFIES condition
Shuffled control matches or exceeds trained motor in live conditions (per-bot drop rate not distinguishable).
Protocol
- Paired lineage: kin K1 trained motor (frozen from Motor RED lineage), kin K2 = same genome with
motor_control.shuffle=true(control-inference weights permuted per action step). - Same seed, same spawn, same 4 h in-world window.
- Independent RCON polling for drop events (
data get entityper prey) — the authoritative attribution. - Verdict:
- PASS: K1 kill rate > K2 kill rate by ≥ 5× under matched exposure.
- PARTIAL: K1 > K2 but ratio < 5×.
- FAIL: K1 ≤ K2.
Ship-gate
- Byte-identity of
default_genomemust hold — the shuffle organ is behind an opt-in genome flag (motor_shuffle_lineage/0, absent fromdefault/0). - MERGED VERDICT required before merging any lib/sp/brain/motor_control.ex changes.
sha256 407665ebac21edea — of the original file, so what was ingested stays checkable.
Plain — written for this website, not the source document
This page is a plan for a control experiment, written before it happened. Two earlier tests came out well, but on their own they leave a hole: something other than the part being studied could explain the result. The fix is a shuffled control, a twin that is identical except that the weights doing the work have been scrambled. If the scrambled twin does as well as the trained one, the page says in advance that the claim fails. No verdict is recorded here.
Plain · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 407665ebac21edea
Clear — written for this website, not the source document
A pre-registration — conditions written down before the run — not a result. The verdict is withheld and the evidence is marked pending.
The motivation is a named gap. An offline test and a live mechanism test have both landed, but a design note had already flagged the step that was missing: a live ablation with a shuffled control. Without it, the live result could be explained by things other than the part under study, and the shuffle is what isolates it.
The protocol pairs two lineages from the same starting point. One keeps the trained weights, frozen. The other is the same genome with the working weights permuted at each step. They get the same seed, the same start and the same window of time in the world, and the outcome is counted through an independent channel that queries the world directly rather than trusting the system's own report of itself.
The verdict table is written in advance with three rungs. A clear ratio in favour of the trained twin. A smaller ratio that counts only as partial. And the trained twin failing to beat the shuffled one, which is a failure. Two closing rules require a default behaviour to stay byte-identical, and a review to agree, before the related code may change.
Clear · written 2026-08-01 by claude-opus-5 · not yet checked by a person · about the document whose sha256 is 407665ebac21edea