The health matrix
Each lever runs with it and without it on the same task. A coin flip decides which arm goes first, and the measure is named before the first run.
ON: the PostToolUse door returns the C612i reading (VNA_RETURN_READING=1). OFF: the same door, the same command, silent (VNA_RETURN_READING=0). Measure registered 2026-10-04T17:10:00.000Z; coin flip written 2026-10-04T17:56:34.072Z; first counted trial 2026-10-04T17:57:48.582Z.
ON 1.542 · OFF 1.417 · ON−OFF 0.125 [-4.083, 4.417] · perm p 0.755 · n 24 · bar 20 · lower is better
ON 0.583 · OFF 0.542 · ON−OFF 0.042 [-0.083, 0.167] · perm p 1 · n 24 · bar 20 · higher is better
Manipulation check: 0 of 2 ON trials that wrote outside the scope got a reading back — the reflex never fired in the trials where it had something to say, so this trial has not yet tested the lever
Not counted: 2 trial pair(s) — the task is not in the assignment file (its flip was not on record before it ran)
ON: the worker's brief is the runner's own brief (briefFor: the row, its guard, its snowball) — the steering payload. OFF: the same task, the brief replaced by the task statement plus the whole spec file inline — the corpus the payload stands in for. Measure registered 2026-10-05T02:45:33.054Z; coin flip written 2026-10-05T02:45:35.802Z; first counted trial none yet.
UNMEASURED — n 0 · bar 20
ON: the brief from the tree, through the claude -p door (a fresh context: the brief alone). OFF: the same brief, delivered into a session already carrying the dispatching chat's context (its transcript's last turns prepended) — the chat door. Measure registered 2026-10-05T02:45:33.054Z; coin flip written 2026-10-05T02:45:35.851Z; first counted trial none yet.
UNMEASURED — n 0 · bar 20
OBSERVATIONAL — not a trial result, until a coin-flip row confirms it: same task mix, fewer tokens in the declared (steered) arm: 46 % (p 0.00485) (public/backup/default-curve.json · aggregate.less · n.steer 44 · n.plain 598).
OBSERVATIONAL — not a trial result, until a coin-flip row confirms it: share of all tokens the worst 5 % of undeclared tasks carry: 39 % (public/backup/default-curve.json · worst5_share.plain · n.plain 598).
OBSERVATIONAL — not a trial result, until a coin-flip row confirms it: share of all tokens the worst 5 % of declared tasks carry: 17 % (public/backup/default-curve.json · worst5_share.steer · n.steer 44).
ON: the brief with the snowball appended (steer.mjs brief's C547 section). OFF: the same brief with the snowball section removed. Measure registered 2026-10-05T02:45:33.054Z; coin flip written 2026-10-05T02:45:35.880Z; first counted trial none yet.
UNMEASURED — n 0 · bar 20
Read 2026-10-05T02:45:51.780Z at aaf4b8a7da · recompute: node scripts/vna/health-matrix.mjs
Sufficient for: whether each lever we claim separates its ON arm from its OFF arm on the measure declared before its first trial, with the arm order fixed by a seeded coin flip written before the run. NOT sufficient for: any lever below 20 paired trials, any measure other than the pre-registered one, or work outside the replay basin the trials drew from; whether the work was good is undecidable (Rice, 1953). Rice (1953): no program can decide, in general, whether another program is correct.