Coordinate polling does not recover perturbed disparity controls¶
Frozen 3c966459, formatting-only revision da10b9bb. See
protocol,
active plan,
171 results and
schema.
Reject the coordinate-search replacement. Despite measured descent opportunities in earlier offline paths, this search from the original seeds does not recover any perturbed case. A descent path to truth is not a deployment initialization.
| Condition | Start | Archived TRF joint recovery | Coordinate search |
|---|---|---|---|
| Clean | Perturbed | 12/36 | 0/36 |
| Clean | Reference | 6/18 | 16/18 |
| Seed 730 | Perturbed | 0/36 | 0/36 |
| Seed 730 | Reference | 12/18 | 16/18 |
| Seed 731 | Perturbed | 0/36 | 0/36 |
| Seed 731 | Reference | 12/18 | 16/18 |
Clean perturbed controls lose all twelve prior successes. Reference-start controls gain ten clean and four per noise seed, with no losses in those groups. Those 18 gains are not recovery from perturbed poses. Pose-only perturbed recovery is also zero in all conditions. Convergence/support passes are clean 3/36 perturbed, 17/18 reference; seed 730 12/36 and 17/18; seed 731 8/36 and 18/18. All nine parallel-plane controls remain rejected.
Every accepted update has nonincreasing measured cost. Calls range 85–800; all actual-call caps hold, and incomplete polls never declare convergence. Total solve time including diagnostics is 429.78 seconds, excluding report serialization. Termination is a complete minimum-step coordinate poll without improvement, not TRF stationarity, and does not establish accuracy or a joint global optimum.
Execution record and validation¶
The initial freeze contained one Ruff line-length violation after an automatic
literal rewrite. The initial execution was explicitly interrupted, preserving
original plan,
partial results
and interruption log.
Revision da10b9bb changes formatting only and updates that source hash. No search
rule, threshold, data, denominator or gate changes. The final run starts afresh
and completes all 171 cases; initial rows are not pooled into results.
Three focused tests, repository Ruff and strict documentation build pass. Checks verify schema, all frozen source hashes, 171-case completion, each paired depth hash against archived TRF, actual-call budgets and monotonic cost traces. No core or production default changed; no fresh real data or holdouts were consumed.
Decision¶
Close this sequence of metric weighting, direct-disparity and coordinate-search variants. None improves noisy perturbed recovery on the frozen controls, and full-cloud real-data metric transfer already failed. Retain the existing MI real-data baseline (6/12) and preserve all negative evidence. The next substantial direction should establish independent geometric camera/LiDAR correspondences and a matched external-method comparison before another local optimizer variant. Review reusable external adapters and measurement assumptions first, then freeze that distinct evaluation. These results do not prove correspondence methods work or establish SOTA; they justify reallocating development effort.