| name | spawn-critic |
| description | Compose the brief for an independent shaded-parity critic in the tank generation program. Use when spawning a graduation adjudication, a re-cert critic for a graduate-change/density/shoe round, or any independent visual verdict. Triggers - "spawn the critic", "re-cert critic", "graduation adjudication", "independent verdict on <tank>", "14-view critic". |
SPAWN-CRITIC — the independent critic brief boilerplate
The critic is ADVERSARIAL to the builder's claims: every number in the verdict is the
critic's own run on the OFFICIAL rigs (§D). Model verdicts:
docs/critique/shaded-parity-abrams-density-recert.md and
docs/critique/shaded-parity-leopard-shoe-recert.md. Two critics max concurrently
(FIFO sanity).
The boilerplate (include in EVERY brief)
- NEVER commit, never edit src/ or GLBs; scratch tooling under tools/tmp-*
(untracked), evidence under shots/ (gitignored), plus ONE verdict doc:
docs/critique/shaded-parity-<topic>.md.
- Env:
export NVM_DIR="$HOME/.nvm" && . "$NVM_DIR/nvm.sh"; own vite 74xx-77xx (never
5001/5002); read docs/BUILD-STANDARD.md (§D, §J especially) + the packet(s) first.
- FIFO §F.1: official rigs SELF-TICKET — run them bare, never wrapped. All 14-view
pair renders under ONE ticket via a batch driver on the identical render path
(clone tools/tmp-b1b3-critic-batch.mjs / tmp-density-critic-batch.mjs). Zero console
errors required. NEVER stop to wait on watchers — chains run sequentially in-process.
- Hash bracketing:
node tools/tmp-hashgeo.mjs --ids=<candidates> BEFORE all
browser work and AFTER the last render (and after any mid-run landing) — every render
must be provably at the verdict hashes (§J: yaw-pair evidence is hash-stamped).
Mid-run landings: path-byte git diff <a> <b> --stat over the family's build+harness
files + a post-landing bracket run keeps in-flight evidence valid.
- Gate x2 minimum (x3 if runs straddle work): rows EXACT to the decimal vs the
landing ledger; per-id JSONs bit-identical between runs is the gold standard.
Standing checks table (§F.4 — every verdict carries it, measured by the critic)
| check | bar |
|---|
| geometry-gate x2 | frozen/landing rows EXACT both runs |
| tmp-hashgeo bracketing | candidates stable across the whole evidence window |
| track-clip-audit --exact | band 0/0 + shoe 0/0 per zone (shoeVox is the player-visible bar); audit dressingSkipped exclusions, never delete |
| turret-parent-audit | stranded 0 / dangling 0; abutting adjudicated on renders (certified ORACLE-REGISTRATION-PINNED classes reproduce exactly) |
| tank-standard-check | clip 0/0, contig 0, census matches packet claims |
| visual-evaluator | RIG PARITY OK every view (a RIG MISMATCH verdict, yaw-proxy > 10°, ABORTS scoring — fix registration first, never score a mismatched pair) |
| §J yaw-90 pairs | re-rendered at the verdict hash for every id whose turret kit changed |
Scoring laws
- CHANGED-VIEW LISTS ARE DIFF-DERIVED (§J): builders under- AND over-list. Build the
per-view pixel-diff matrix (proc half, threshold recorded, label band y13-21 excluded)
vs the superseded baseline; the matrix REPLACES the builder's list — score
unlisted-but-changed, skip listed-but-unchanged. Diff COUNTS are harness-local: the
view SET binds, magnitudes never port across harnesses.
- Severity bar: graduation = >=9.0 EVERY one of the 14 views (front, frontleft, left,
rearleft, rear, rearright, right, frontright, top, hero-frontleft, hero-rearright,
hero-toptilt, close-front, close-roof). Re-cert = every changed view >= its ratified
score (floor holds; price views UP only on measured structure). Cross-critic severity
disputes calibrate against the ratified chieftain5 anchor — measurements, not vibes.
- Sky/hole claims: banked flood method — bg |px−0x151b20| maxch<=13 AND B−R >= +8
(blue signature), border flood removed, label band excluded; adjudicate flood deltas by
locating the exact pixels at 10x, never by count alone (connectivity teeter class).
Tone claims: ITU-601 luma rects WITH coordinates; read the p90/sd/med triplet.
Deep-shade zones with zero percentile spread are shadow reads — tint orders are void.
- Orientation: critic top pair renders bow at image BOTTOM — gun-overhang check
before calling turret-front reads.
- Standing silhouette checks on every verdict: §B1 slopes/no-staircases (+ turret
cheeks both sides §B1.1), §B2 contiguity/no-turret-holes, §B3 no-mystery-boxes +
§B3.1 gun-run prisms (order-rank = silhouette break), §B3.2 density, §B4 track
containment, §B5 furniture parenting at yaw, §B6 track-run trapezoid, §H.4
variant-distinctiveness + national MG grammar, §C missing-side (LEFT view vs right).
- Renders are the witness for winding (flood is blind to reversed slabs; masks are
DoubleSide and hide them).
Verdict doc format (docs/critique/shaded-parity-.md)
- Header: round, landing commit, candidate hashes (mesh/verts), frozen baselines, laws
governing the round. 2. Provenance (every command yours; tree state; brackets;
mid-run landings adjudicated). 3. HEADLINE verdict. 4. Standing-check table.
- Claims audit (packet done-gates re-derived). 6. Diff-derived changed-view matrix.
- Per-tank verdicts with per-view scores + justifications. 8. Residuals certified
(none blocking / orders). 9. Law discoveries for the bank.
Verdict vocabulary: RE-CERT PASS (re-freeze ) / GRADUATION PASS / FAIL with
ordered work. Orders must be measurable. The orchestrator re-freezes only on PASS.
