Fixed-budget multistart results¶
Protocol frozen at b221d707. All 15 baseline
fits and 75 restart fits completed. Baseline result objects exactly reproduce
the preceding known-geometry study.
| Condition | Seed | Selected start | Baseline deg / m | Five-start deg / m | Selected converged | Used calls baseline / five-start |
|---|---|---|---|---|---|---|
| exact_z_visible | +1 | 3 | 1.0459 / 0.2160 | 0.3172 / 0.0559 | False | 229 / 800 |
| exact_z_visible | -1 | 2 | 0.8972 / 0.1700 | 0.5651 / 0.1500 | False | 253 / 800 |
| exact_z_visible | +0 | 0 | 0.0349 / 0.0063 | 0.0349 / 0.0063 | True | 121 / 758 |
| exact_z_all | +1 | 3 | 1.0717 / 0.2349 | 0.3661 / 0.0637 | False | 193 / 800 |
| exact_z_all | -1 | 1 | 1.3748 / 0.2359 | 0.7842 / 0.1381 | False | 253 / 800 |
| exact_z_all | +0 | 3 | 0.0271 / 0.0000 | 0.8646 / 0.1420 | False | 109 / 734 |
| exact_range_visible | +1 | 3 | 0.4725 / 0.0718 | 0.2526 / 0.0515 | False | 337 / 800 |
| exact_range_visible | -1 | 2 | 0.8184 / 0.2767 | 0.6125 / 0.0750 | False | 265 / 800 |
| exact_range_visible | +0 | 0 | 0.2500 / 0.0000 | 0.2500 / 0.0000 | True | 73 / 713 |
| biased_z_visible | +1 | 3 | 1.0459 / 0.2160 | 0.3172 / 0.0559 | False | 229 / 800 |
| biased_z_visible | -1 | 2 | 0.8972 / 0.1700 | 0.5651 / 0.1500 | False | 253 / 800 |
| biased_z_visible | +0 | 0 | 0.0349 / 0.0063 | 0.0349 / 0.0063 | True | 121 / 758 |
| holes_z_visible | +1 | 3 | 1.1449 / 0.2278 | 0.3594 / 0.0718 | False | 193 / 800 |
| holes_z_visible | -1 | 2 | 1.2791 / 0.2236 | 0.6131 / 0.1118 | False | 193 / 800 |
| holes_z_visible | +0 | 0 | 0.0884 / 0.0153 | 0.0884 / 0.0153 | True | 133 / 746 |
Outcome¶
Accepted perturbed recovery falls from 1/10 to 0/10. All ten selected multistart runs hit their 160-call limit. Ignoring convergence, five poses satisfy the geometric thresholds versus one baseline pose; these five are explicitly not accepted. Every perturbed five-start trial consumes 800 calls while baseline trials terminate in 193–337. This is equal maximum budget, not equal used work. Measured total times for the ten perturbed trials are 5.23 s baseline versus 17.71 s multistart on this CPU. Reference-start controls take 1.20 s versus 7.69 s in total.
The selected MI improves versus baseline in 7/10 perturbations and worsens in 3/10: the original start also receives only 160 calls, so it cannot reproduce its fully converged baseline. This allocation neither preserves the best baseline result nor gives the promising alternative starts enough finishing budget. The observed geometric improvement is a useful search signal, not validation of the algorithm.
Reference-start controls pass 5/5 baseline versus 4/5 multistart. In the all-return condition, a higher-MI alternative is selected with 0.8646 degrees / 0.1420 m error rather than the near-truth baseline (0.0271 degrees / 0 m). The selected run is unfinished and correctly rejected. Its MI already exceeds the converged baseline by 0.00364, illustrating why higher MI alone cannot certify true calibration. The other four reference-start outputs match their baseline poses.
Decision and next test¶
Do not promote this equal-split multistart strategy or run it on real data as an improvement. Keep all attempts and the failed convergence gates. A next controlled experiment should preserve the fully converged baseline, then spend only the remaining declared budget on exploration/refinement with explicit continuation state. Budget must count the baseline and every exploratory call. A better objective cannot by itself override independent pose/visibility validation; the all-return reference-start drift remains a required rejection control. Freeze any allocation rule before evaluation, and do not tune start positions against these results.
Full evidence includes every start, trace, selection and runtime; schema records the report envelope. Seven focused unit tests, repository Ruff, strict docs build, source-hash validation and report/budget/selection checks pass. Neither the core solver nor production defaults changed; no SOTA or real-data improvement claim.