Why the boundary-aware trial remains incomplete
The trial finished 17/24 fits successfully: all six completed controls, nine of fifteen unfinished primary fits and two of three perturbed repeats. It improved primary recovery over the earlier continuation trial (8/15) and rectangular trial (2/15), but none of the three test stretches became fully resolved. It did not qualify for a full recovery run. Published findings remain 12 globe preferences, zero completed flat preferences and 41 unresolved stretches.
This diagnosis checks sealed saved results, tangents and receiver-epoch errors. It performs zero archival fits and zero native predictions. One small affine calculation per saved tangent examines the optimizer's remaining local correction; these are not additional observations or recovered flight fits.
Six early stops, one budget failure
6 failures already pass the gradient check. Their remaining local correction predicts a very small improvement in squared error, but a change in the residual vector greater than the unchanged stability limit. Their actual nonlinear costs did not decrease after up to 24 halvings of the proposed step. The saved gradient and stability checks therefore disagree about whether work is complete; the stricter combined gate correctly keeps these fits unresolved.
| Stretch / processing | Model / start | Calls | Gradient / limit | Residual correction / limit |
|---|---|---|---|---|
| 03454752 / profiled_removal | flat_still / repeat | 79 | 0.00213 | 1.06 |
| 03454752 / profiled_removal | sphere_still / primary | 117 | 0.00321 | 2.22 |
| 03454752 / unsubtracted | sphere_still / primary | 63 | 0.0073 | 4.19 |
| 233183f6 / profiled_removal | sphere_rotating / primary | 114 | 0.000223 | 24.4 |
| 233183f6 / unsubtracted | sphere_rotating / primary | 54 | 0.00124 | 1.2 |
| b365bdf3 / unsubtracted | sphere_still / primary | 99 | 0.0461 | 9.55 |
For these six, predicted fractional cost improvement is below 1e-8, while the relative residual correction exceeds 1e-6. Squared-error improvement and movement of individual predictions are different quantities; a tiny improvement alone is insufficient. 0/6 have positive boundary slack of 1e-12 or less. The earlier one-floating-point-step boundary-start explanation does not account for these saved endpoints.
1 fit exhausts its tangent budget and still fails stationarity. It accepts many small fractions of its full local correction. Its largest remaining correction includes gravity and gyro gains; a full step produces a much worse actual cost despite the local tangent predicting improvement. This exposes a problem with taking that full local step on the nonlinear objective, distinct from the six nearly stationary stops. Increasing the budget might help this case, but has not been tested and does not fix the early stops.
What the repeated starts show
The comparison below uses the saved predictions at identical receiver timestamps. Track difference is the largest axis-specific RMS difference. The original numerical agreement limits are 1e-6 relative cost and 0.01 m on every axis; both fits must also pass their own stopping checks. No repeat can replace an unsuccessful primary.
| Stretch / repeated model | Primary / repeat passed | Relative cost difference | Maximum axis RMS difference (m) | Repeatability qualified |
|---|---|---|---|---|
| 03454752 / flat_still | True / False | 9.66e-10 | 2.34e-05 | False |
| 233183f6 / sphere_rotating | False / True | 2.16e-09 | 1.36e-06 | False |
| b365bdf3 / sphere_still | False / True | 1.59e-09 | 0.000121 | False |
All three pairs satisfy the cost and track agreement limits, despite opposite pass/fail stopping outcomes. Their largest axis-specific RMS difference is 0.121 mm. This isolates sensitivity of the stopping checks at nearly identical reconstructed tracks. It does not demonstrate a global optimum or validate a weaker acceptance rule.
False in the last column means the full qualification rule failed, even when the
costs and tracks agree. An approximately repeated prediction is useful numerical
evidence; it is not permission to relabel an unresolved fit.
Smallest useful next test
Before another recovery trial, isolate prediction/tangent accuracy near the saved endpoints. On independently generated integrated-motion controls and a separately frozen small native diagnostic, compare fresh residual changes with tangent predictions along predetermined feasible directions over several step sizes. Include a weak calibration direction and a well-conditioned direction. Check receiver-boundary timing behavior explicitly. Record objective changes and individual receiver-epoch prediction changes, rather than relaxing stopping rules to raise completion counts.
The saved artifacts cannot establish a numerical noise floor or inaccurate derivatives: they do not contain the full residual vectors for every rejected step. Large nonlinear curvature, cancellation and timing-interval changes remain distinguishable hypotheses, not established causes. A damped local correction is a candidate to test against known answers, not a validated repair. Keep bounds, observations and physical models unchanged; defer all-41 rollout until recovery and repeatability qualify.
Reproduce
env PYTHONPATH=analysis OPENBLAS_NUM_THREADS=2 OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 ~/venv/bin/python docs/ilvis0-boundary-diagnosis-20261013/build_report.py
The CSV includes all 24 starts; the detailed report preserves local corrections and the three repeat comparisons. The receipt hashes input summaries, journals, tangents, epochs, audit code and generated artifacts, and records the numerical environment.