Criteria: tasks/chain13_corridor_sweep_criteria.md.
Metrics = run_mission.analyze. FLOOR = coverage < 0.99
OR not completed (rs below the usable floor — do not adopt).
| run | coverage | completed | pe_p99 | abs derated [s] | derate frac | t_complete [s] | ze_p99 | pe_med | verdict |
|---|---|---|---|---|---|---|---|---|---|
| rs 0.55 (incumbent) | 0.9981 | True | 0.1651 | 689.3 | 0.644 | 1070.8 | 0.0549 | 0.0712 | gate-pass |
| rs 0.45 | 0.9981 | True | 0.1644 | 619.9 | 0.766 | 809.2 | 0.0463 | 0.0922 | gate-pass |
| rs 0.40 | 0.9978 | True | 0.1586 | 485.8 | 0.637 | 762.6 | 0.0495 | 0.1018 | gate-pass |
| rs 0.35 | 0.9965 | True | 0.1603 | 358.0 | 0.478 | 749.7 | 0.0523 | 0.1217 | gate-pass |
| rs 0.3 | 0.9944 | True | 0.1639 | 0.0 | 0.000 | 694.9 | 0.0562 | 0.1313 | gate-pass |
| rs 0.25 | 0.9942 | True | 0.1700 | 0.0 | 0.000 | 691.5 | 0.0570 | 0.1282 | gate-pass |
| rs 0.2 | 0.9953 | True | 0.1803 | 638.1 | 0.790 | 807.4 | 0.0513 | 0.1506 | gate-pass |
Adoption rule (criteria): smallest rs holding coverage>=0.99 + completed + pe_p99<0.2 that improves abs-derated-time/pe_p99 vs incumbent, then ±2% probes. rs 0.40 stays the winner unless a smaller rs dominates it AND passes ±2%.
The only thing favoring rs 0.30 over the pe_p99-optimum rs 0.40 was
its zero-derate “singularity-free” valley — which would have required a
registered objective amendment to adopt. Before amending, we probed
whether the valley survives the fragility campaign’s worst-case (hi)
noise. Tool: validation/corridor_noise_precheck.py
(radius_scale at LOAD time + UncertaintyModel). Seed 0,
full helix.
| cell | coverage | derate | pe_p99 | verdict |
|---|---|---|---|---|
| rs 0.30 nominal (no noise) | 0.9944 | 0.000 | 0.1639 | reference |
| rs 0.30 camera-hi (pos 0.10 m, ang 5°) | 1.0000 | 0.000 | 0.1639 | valley holds |
| rs 0.30 state-hi (pos/vel 0.05) | 0.9955 | 0.596 | 0.1602 | valley BROKEN |
Finding. Coverage held in both cells (>=0.99). The singularity-free property did not: state-estimation noise drove derate 0.000 -> 0.596. Mechanism: the zero-derate sat ~2 mm above the 0.025 floor (sj_min ~0.0269); camera-pose noise perturbs FOV marking only (arm conditioning untouched -> derate flat), but state noise (std 0.05 m ~= 25x the margin) perturbs the commanded joint trajectory across the singular wall. The valley is speed-robust (prior ±2% probes flat) but state-noise-fragile.
Decision. Registered rule =
both coverage>=0.99 AND derate stays 0 -> rs 0.30; either breaks -> rs 0.40.
Derate broke -> adopt rs 0.40, no objective
amendment (rs 0.40 = pe_p99 optimum + registered-rule winner; rs 0.30
never dominated and its sole differentiator is a noise-free artifact).
§3b re-pin = radius_scale 0.55 -> 0.40.
Campaign note (preview, not a blocker): derate is itself highly state-noise-sensitive (0.000 -> 0.596 from one source) — expect conditioning to be among the most-moved metrics in the 46-run study at any standoff.