Local Level Lab

Airborne measurements. Open methods.

Complete joint-adapter known-answer qualification

The immutable subdomain transfer passes its rejection tests, and every one of the 48 planned source domains passes the transfer and cap preflight. The joint adapter also passes local tests. This separate freeze checks its complete path with durably charged work before the large prediction comparison.

Use six fixed 13-packet controls: all three candidates under both processing cases, 5 ms packets, training at 15 and 30 ms and prediction at 50 ms. Keep the original known-answer fixture recipe. Perturb the predeclared calibration in mount, bias and gain coordinates, and narrow one velocity coordinate while retaining all parent constraints. These are software controls, not flights.

The fixture generator may use its known calibration to generate control observations. The target adapter must build its training constraints and future range before the final truth audit. This is unblinded software qualification, not independent confirmation. Record the native fixture outputs, all old/new bounds, parent support verification, subdomain proof and exact row/endpoint audits. Preserve all 51 expression columns and all 78 enclosing coordinates.

Separate finite allowance

Freeze implementation, test recipe, catalog, compiler, environment and test package version. Charge native and bound operations before each call and reserve auxiliary work before execution. Never disable the real worker's cooperative continuation checks. All six identities and their truth, column, row, endpoint and ordering checks must pass. No aircraft originals, protected ESA data, observed fits or calibrated confidence claim follow from this test.

Source/domain preflight · Roadmap.