Local Level Lab

Airborne measurements. Open methods.

Source derivative component qualification

The completed long-history screen removes the time-limit failures but still overstates horizontal error allowances. This separate component bounds source derivatives on the physical calibration domain, so a mean-value calculation can retain their linear information.

It uses interval automatic differentiation with all 27 physical columns, exact literal inputs and outward binary64 arithmetic. Rotations use a finite series with an explicit value and derivative tail. No accuracy bound for library sine or cosine is assumed. The working state is fixed-size. Both midpoint and end sources are retained. A clipped clock fraction includes both one-sided slopes when the domain crosses a packet boundary.

Check five rotation vectors against independent exact-rational value and derivative references, including zero, mixed axes and eight radians. Check a 13-packet zero-motion source, a four-packet nonzero source against rational rotation predictions, a two-packet clock-boundary case, an independent cubic mean-value example with displaced supplied center/rows, and invalid inputs. Keep domain rejection, immutable storage and the finite packet allowance.

Separate finite allowance

Freeze the tests, implementation, compiler, Python/test-package versions and control catalog before launch. All identities and planned charges must pass. These controls qualify the mathematical component on known answers. They do not install a training constraint, establish useful complete prediction widths or qualify native/observed propagation. Source/domain binding, training intersection and full native qualification remain separate gates.

Derivative-bound rationale · Roadmap.