Qualification away from the equator

Goal: test whether the completed 200 Hz correction improves calibration recovery away from the equator, and demonstrate why increment conventions cannot be mixed. This implements the next-step proposal. No NASA input, observed optimizer starts, production promotion or git operations.

Frozen matrix and budget

Use 12-second, 220 m/s constant-latitude tracks at 1,000 m, at +45° and −45°. Generate all three physical truths independently from analytic motion and eight-point Gaussian quadrature. Ordinary body-rate integrals are supplied to the unchanged combined-correction candidate. Keep the six original calibration limits, 24 off-grid receiver observations, original solver and 200-sweep allowance including two final prediction sweeps. No attempts are repeated.

Two constructions: nominal 200 Hz with interior calibration and pitch/yaw; variable spacing (the original ±50 ppm control) with near-boundary calibration and added ±0.35 radian roll. Both constructions run from zero and the original opposed start. The crossed identity matrix is 3 truths × 2 latitudes × 2 constructions × 2 starts: 24 starts maximum, at most 4,800 sweeps for fits. These are software controls, not flight simulations or independent scientific trials. Expected runtime is a few minutes; the sweep allowance, rather than elapsed time, is binding.

Before fitting, make 24 known-calibration predictions: original and combined integration for each of the 12 constructions. Six additional already-compensated controls use endpoint receiver observations, one per truth/latitude. For each, make two predictions: the explicitly supported original treatment and a deliberately mislabelled-input diagnostic demonstrating double correction. The convention guard must reject enabled compensation when input is correctly labelled already compensated. These 12 predictions add no optimizer starts. Total ceiling: 4,836 native sweeps.

Pass criteria for each fit: independently generated track prediction below 2 mm, known-calibration prediction below 2 mm, numerical convergence, and objective no higher than the known-calibration objective plus 0.01. Report calibration errors, active limits and conditioning separately; these pass criteria do not imply unique calibration. Preserve failures and charged interrupted starts without retries.

Implementation and verification

• Add analysis/lll/ilvis0_latitude_controls.py: independent latitude-aware Euler motion, local Earth/transport vectors, receiver metrics and quadrature. Construct a precisely defined finite-increment convention from analytic attitude endpoints and independently integrated force, rather than calling a production predictor.
• Add analysis/tests/test_ilvis0_latitude_controls.py: both hemispheres, all models, axes/signs via finite differences, force/frame identity, known-calibration native predictions, independent ODE validation of finite increments, explicit convention guard, unequal spacing and invalid inputs. Watch new tests fail before implementation.
• Add build_report.py here: verify prior frozen sources, seal source/environment/ build hashes before numerical starts, exclusive lock, fsynced append-only starts, atomic results and completion, no retries, saved-only publication mode.
• Publish CSV comparisons, findings, exact budget and a separate hash receipt. Packaging must preserve numerical files and journals. Update current roadmap links.

The already-compensated convention here is an explicitly constructed mathematical control, not an assertion about Applanix firmware. Endpoint observations isolate full-packet treatment; arbitrary partial finite increments require a separately defined convention. Do not silently reinterpret those inputs or invent archival samples. Full nuisance freedom, long records, noise and independence remain later qualifications.