Pre-reg: clutter-patch promotion re-gate (pinned 20×5 probe, production path)

2026-08-18 ~14:1xZ. Registered before any data. Queue item clutter-patch-promotion-regate; the re-gate itself was registered with the original gate (in-channel pre-reg 05:23Z 2026-08-13: “re-gate on the pinned 20×5 probe before any behavioral eval moves”).

What happened

The clutter-patch promotion landed (2026-08-18 work session): SO101Sim v3/v4 now default to clutter_appearance='patched' — the mined real crops pasted onto the drawn plate at the drawn poses (sim/clutter_patch.py, moved verbatim from the gate instrument), stand-in geoms parked off-frustum, wrist path untouched, zero extra RNG draws. Render oracles are green (wrist bit-exact vs standins; top bit-exact outside clutter-affected pixels; physics/stream identity).

The gate evidence (05:4xZ 08-13, PASS) was measured on a hooked instrument: no_clutter mask over the pasted plate, inside sim_fg_appearance_fix.py. This re-gate checks the production implementation reproduces that read before any behavioral eval runs on the patched substrate.

Protocol (frozen)

The leg (a)/(b) harness’s slot grid verbatim: 20 seeds × 5 appearance draws (appearance = 1000*draw + seed), settled resets, numpy post backend, top camera. Two production arms, no hooks:

• patched — SO101Sim(render_style="v3") (production default);
• standins — same with clutter_appearance="standins" (in-run anchor).

Embed with the same er_60k probe checkpoint + real_v2 A/B reference as the gate run (sim_encoder_ood_probe machinery), knn5 AUROC vs real.

Bands (frozen)

• standins anchor: AUROC in 0.708–0.718 (the gate’s registered v3 abort band). Outside → abort, no claims.
• PASS: patched AUROC within ±0.010 of 0.556 (the gate’s patched read). The production path differs from the gate arm only at pixels where a stand-in occludes dynamic content — measured none on the oracle seeds — so the read should reproduce nearly exactly.
• 0.556 + (0.010, 0.030] → record, inspect dumped frames before any behavioral move (implementation drift suspected, e.g. mask vs off-frustum occlusion differences).
• Worse than +0.030 → the promotion is suspect: revert the default to standins pending diagnosis.

Cost and gating

~0.02 GPU-h (200 renders CPU/EGL + encoder embeds). Gates behavioral evals on the patched substrate only (demo-gen v1.1, future sim100 cells). Tonight’s pdnorm endpoint sim100 is pinned --clutter-appearance standins (pdnorm prereg Amendment 1) and does not wait on this.