Completing the remaining archival comparisons

The original 41 unresolved stretches contain 130 unfinished model fits. Each stretch requires three models under two processing cases. Existing converged fits are preserved and reused; incomplete calculations are the recovery targets.

This metadata-only preparation records the complete workload while the actual-position stability trial runs. It is not ready to launch. The qualification trial must finish and pass all numerical, control and repeated-start checks before the method can be used more widely. Neither partial success nor close track agreement bypasses that gate.

Planned workload after qualification

The qualified subset would supply 15 recovered primaries and three repeated starts. Its three stretches remain in the combined 41-stretch ledger, with their provenance and results identified explicitly. The six completed control fits are run again in the new freeze to check consistency. One repeat per stretch is a numerical robustness check; it does not make a stretch an independent flight or establish global optimality.

The proposed ceiling is 95,718 native sweeps: at most 600 fitting calls and two track predictions per new start. There are no retries, replacement starts or extra attempts for difficult results. Use one worker and two total numerical threads. Matched first-pass timing gives a ceiling estimate of 51–76 hours. This assumes every start spends its entire allowance; actual runtime may be shorter. A new runtime freeze and launcher must enforce the finite limits before execution.

Selection and starts

The stretch inventory preserves filenames, source hashes, UTC intervals, duration, speed, recorded IMU type and the original failure categories. The task ledger records all primary recovery targets, repeated-start checks and controls, including which subset results could be reused.

Original failed-fit status defines the primary tasks. UTC order defines the repeat schedule; the preferred model for each repeat cycles through the three candidate names, using the first originally unfinished fit of that model or a fixed fallback. The subset's three already declared repeats are preserved. Model costs, shape preferences and gyro values never choose the schedule.

New primaries start from the preserved original first-pass endpoints. Repeats use those same endpoints with the fixed 0.01 normalized-coordinate perturbation. Fresh controls use the qualified trial's primary endpoints. Every start and input must be hashed before execution. Existing observations, clock, calibration bounds and receiver-error assumptions remain unchanged.

Reporting

The final combined ledger must distinguish original results, reused qualified subset results and new recovery outcomes. A stretch is newly complete only when all six required fits pass and its predetermined repeated-start check passes. Any incomplete primary or failed repeat leaves it unresolved. Report shape first; report the background-field comparison only after consistent globe preference. Retain every failed or interrupted attempt. No result here is a calibrated detection.

Original published findings remain 12 globe preferences, zero completed flat preferences and 41 unresolved stretches during qualification. This preparation reads saved metadata only, opens no NASA originals and starts no numerical workers.

Reproduce

~/venv/bin/python docs/ilvis0-position-recovery-preparation-20261013/build_plan.py

The receipt hashes the original summary, saved inventory, immutable qualification manifest/proposal and generated plan. It does not consume or hash mutable runtime journals from the running trial.