Matched prediction comparison

The launcher is implemented. Software qualification and a separate source/input/ environment freeze make the 12-fit comparison reviewable before launch. The bounded 12-fit observed-data comparison has been approved and launched. Its final results will be published separately after all charged identities finish.

The method and cost are fixed: at most 12 starts and 3,060 native sweeps, no retries, one worker and two numerical threads. Estimated ceiling runtime is 154–172 minutes. This estimate uses saved timings; the new error objective has not been benchmarked.

See the frozen plan and worker qualification. The scope retains the existing October 5 trajectory, raw inertial measurement unit (IMU) increments and receiver observations. It tests two common assumptions about receiver-error correlation using calibration fitted before the prediction period. It is retrospective development evidence. The existing findings stay unchanged.

The worker refuses to run without an authorization artifact matching its manifest and exact allowance. This binds execution to the recorded study scope; software qualification alone does not establish observational validity.

Prepare or verify the freeze without extracting or integrating NASA originals:

env PYTHONPATH=analysis OPENBLAS_NUM_THREADS=2 OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 ~/venv/bin/python analysis/tests/ilvis0_prediction_covariance_worker.py --prepare-only

For a new, budgeted study, record the allowance with the same command and --record-authorization, then launch with --detach instead. The detached launch must run outside a short-lived command sandbox. The worker retains its lock, journals and metadata after a terminal or network connection ends.

After interruption, check the exclusive lock and unchanged freeze before launching again. An already charged identity without a saved result remains unresolved; there is no retry. Completed studies refuse to run again.

Successful completion automatically publishes results. --package-only verifies and republishes completed evidence without numerical work. It produces a public results summary, fit table, receiver-epoch errors and compressed parameter-probe diagnostics. Local rank and sampled sensitivity ranges are not calibrated confidence intervals. Unfinished fits remain unresolved.