Prediction precision at the shared-recorder endpoints

All six endpoints completed, using 60 native integration evaluations and zero optimizer starts. The published finding remains 30 globe preferences, zero flat and 23 unresolved. These checks do not promote an old fit or change its acceptance rules.

The finite plan tests the three failed model/processing endpoints and their three successful-recorder counterparts. Each retains its original complete stretch, observations, clock, covariance and calibration bounds. The saved recorder-pair audit compares measurements on common UTC seconds and explains the exact stopping failures.

Central prediction checks

Both integration implementations are evaluated at identical saved parameters. The value implementation predicts the track; the tangent implementation predicts the track and its derivatives. Neither uses fused Group-1 navigation as an observation. Two identical evaluations of each implementation test deterministic repeatability.

Stretch ID prefix	Original endpoint passed	Reproduced tangent cost	Central value/tangent relative cost difference	Largest coordinate difference (µm)
3df34ea3	False	27.218	2.11e-14	5.457e-06
233183f6	True	13.7973	0	0
e6629f14	False	22663.4	0	0
abe881b5	True	18951.2	0	0
a953c8dd	False	2.37385	1.12e-15	3.638e-06
31d7e501	True	0.978621	1.9e-12	0.0001473

The largest repeat difference is 0 µm. The largest difference between value and tangent coordinates, including all noncentral probes, is 0.0002474 µm. These comparisons test the two implementations at the specified points; they do not independently verify every derivative or establish a precision floor for nearby parameters.

Full, half and quarter local corrections

The correction is the same bounded affine least-squares direction used by the stability test. Each fraction receives a fresh prediction and derivative evaluation. Ratios at most one pass the original cost and gradient limits; position RMS must remain at most 1 mm per axis, and no epoch may move more than 1 cm.

Stretch ID prefix	Correction fraction	Fresh cost change / limit	Fresh gradient / limit	Largest axis RMS change (mm)	These probe checks pass
3df34ea3	1.00	1.057	0.0002476	0.02625	False
3df34ea3	0.50	0.813	0.03559	0.0131	True
3df34ea3	0.25	0.3291	0.05352	0.006494	True
233183f6	1.00	0.397	0.000102	0.00133	True
233183f6	0.50	0.272	0.0006119	0.000664	True
233183f6	0.25	0.3108	0.0009398	0.000327	True
e6629f14	1.00	0.02961	6.578e-06	3.766	False
e6629f14	0.50	0.06987	0.04961	1.883	False
e6629f14	0.25	0.04505	0.07442	0.9415	True
abe881b5	1.00	0.0163	3.437e-06	0.01217	True
abe881b5	0.50	0.008037	0.0002041	0.006082	True
abe881b5	0.25	0.0001339	0.0003062	0.003038	True
a953c8dd	1.00	4.014	0.0004536	0.001773	False
a953c8dd	0.50	1.257	0.00395	0.0008865	False
a953c8dd	0.25	1.686	0.006054	0.0004441	False
31d7e501	1.00	5.474	0.0003909	0.0009594	False
31d7e501	0.50	0.6494	0.0003422	0.0004807	True
31d7e501	0.25	0.3992	0.0002733	0.0002445	True

Half or quarter steps passing here are diagnostic observations, not accepted fits. Shrinking a correction arbitrarily can hide remaining work; a revised stopping procedure needs independent known-answer qualification and a separately frozen observed comparison. The CSV retains affine predicted costs alongside actual costs, rather than assuming a locally linear prediction is exact.

What these checks establish

Every central objective reproduced its saved endpoint within the frozen reproduction limit. Both implementations produced identical repeated coordinates.
At the October 2009 IMU8 failure, the half correction passes all measured probe checks, while the full correction misses the relative-cost limit narrowly.
At the September 2015 DMS disc failure, the quarter correction moves vertical predictions by about 0.94 mm RMS and passes these checks; the full correction moves them by about 3.77 mm RMS.
At the October 2015 DMS background-field failure, full, half and quarter corrections all miss the relative-cost check. The progression is not monotonic.
The October 2015 loaner endpoint was complete under the original solver. It fails this newer full-correction check, although its half and quarter probes pass. Original completion and a fresh check under a later policy are distinct.

The two implementations disagree far too little at these evaluated points to explain the failed cost probes. Exact repeated evaluations also rule out random variation at the same parameters in this diagnostic. They do not rule out deterministic roundoff as parameters change, inaccurate derivatives or nonlinear curvature. The saved local systems are poorly conditioned, with weak directions dominated by gravity correction and correlated calibration parameters.

The useful next test is derivative and accumulation precision in those directions, before qualifying a revised numerical strategy. More attempts or accepting whichever step fraction passes would not establish that the original fit is fully resolved.

The next numerical investigation should distinguish integration roundoff, derivative accuracy and nonlinear curvature in the weak gravity/calibration directions. The September disc endpoint also needs its vertical stability failure addressed. This study supplies no new nuisance freedom or statistical significance.

Reproduce the report

~/venv/bin/python docs/ilvis0-shared-leg-probes-20261014/build_report.py

Individual probes preserve objective values, physical movements and checks. Report rendering performs no integration or optimization. Completed native work is frozen and cannot resume; interrupted identities receive no retries.