Final-state trajectory gradient reporting¶

The native sparse screw_linear fitter now reports gradient_norm at its returned state. This is a reporting-correctness and development record on already-spent synthetic data, with no timing, independent accuracy or SOTA claim.

What changed¶

The normal equations use scalar square-root information, weight = 1 / sigma. The reported quantity is max(abs(J.T @ (weight**2 * residual))), half the squared-objective gradient in the fitter's local coordinates. Information-form priors are already whitened and carry unit row weights. The previous result could retain this quantity from before the last accepted update.

The fitter tracks whether the last linearization still describes the current state. An accepted update or restoration invalidates it. The final gradient refreshes that model when necessary; otherwise the existing model and gradient are reused. Covariance shares that same final model. A gradient stop, rejection, or failed solve does not itself change the state and needs no refresh.

The maintained checks cover independent four-anchor pose gradients in identity and rotated world frames, weighted targets, accepted updates at the iteration limit, tiny accepted steps, rejected steps, acceptance followed by rejection, the damping limit, and exceptions/nonfinite solves, with and without calibration parameters. Of the 24 checks, 15 fail on the preceding fitter and all 24 pass after correction. This count includes model-reuse checks as well as reporting.

Stopping criteria, iteration limits, accepted updates, the residual model, covariance and serialized schemas are unchanged. A final accepted update in a fit without calibration parameters can require one final linearization that the old, stale report omitted. An already-current model is reused.

Frozen paired evidence¶

The baseline is commit 1e6da103bf1d268d6469cb60dce269099dae53ea. The corrected numerical source is aa4a0c24069020624e4d6bb8c519c2d9d9838fcd. All 401 source modules were hashed with normalized LF text before and after each cohort. Only src/calibrex/core/continuous_time_sparse.py differs.

The receipt, schema, source manifest, original cohort producer and archive producer bind all inputs, outputs and source. The complete archive retains all 171 declared members, including both three-seed cohorts, full input/options and fit records, native artifacts, final normal matrices and gradients, schemas, CLI logs, the preceding source, and exact-test red/green receipts and logs.

The producer invokes calibrex.cli.main.main with the recorded CLI arguments in its process. Pass-through wrappers call the original maintained fitter and assembler, recording their arguments and results while returning their outputs unchanged. Additional dense products record the final normal model, so this instrumented cohort is not a latency measurement. No residual, Jacobian, sampling, stopping, or state override is applied.

Runtime was Python 3.12.10, NumPy 2.3.4, SciPy 1.17.1 and Pydantic 2.12.5. OpenBLAS/MKL/OMP limits were one; concurrent host activity was allowed.

Seed	Previous reported gradient	Corrected final gradient	Iterations	Assembly count, previous/current	Current native artifact
20261010	0.0650938792306	0.00356875260312	21	22 / 22	JSON
20261011	0.0426090979809	0.0430917463485	23	24 / 24	JSON
20261012	0.0719102156829	0.0740917147636	29	30 / 30	JSON

Each pair has exactly equal canonical problem/options JSON. All fitted result fields except gradient_norm match exactly, including knot poses, calibration, covariance, objective, status and iteration count. The final normal matrices and gradients match exactly. Native JSON excluding provenance also matches exactly, including all holdout metrics and signed controls. All three policies pass, with all 32 controls detected in each case. Every corrected report equals the recorded final normal-gradient norm. The SHA-256 inventory binds every archive member and all three current native artifacts.

Limits and previous attempts¶

Every final raw gradient above exceeds the configured 1e-8 gradient threshold. The existing converged label can also result from a small accepted damped step. Correct reporting discloses this limitation; it does not certify stationarity. These gradients mix physical parameter units, so normalized stopping needs a separately declared contract. Clock behavior at boundaries and discarded marginalization gradients/cross-knot information remain separate work in the reliability study.

The archive also retains six named historical diagnostic directories and their index. The first aggregation harness referred to an incorrect SE3 field name; its six raw fits and all six aggregation errors remain intact. The corrected aggregation reuses exactly those same raw fit bytes without rerunning them. The private stopping draft has three expected failing checks and two positive controls. A reporting-only prototype fixes two failures while deliberately leaving premature stopping red. Both later four-case model-reuse prototypes and the three saved-window model diagnostics are retained. These private stopping drafts are separate from the 24 maintained reporting/reuse checks.

Assembly counts do not establish a wall-clock speedup. The controlled prototype shows that an already-current calibration model can avoid a repeated assembly; the native three-seed assembly counts are unchanged. Earlier speed records and SOTA verdicts remain bound to their original source and results.

Replay¶

At the pinned corrected source, the direct CLI recipes are:

calibrex trajectory recover-lidar-imu-joint --trajectory screw_linear --seed 20261010 --output replay-20261010.json --json
calibrex trajectory recover-lidar-imu-joint --trajectory screw_linear --seed 20261011 --output replay-20261011.json --json
calibrex trajectory recover-lidar-imu-joint --trajectory screw_linear --seed 20261012 --output replay-20261012.json --json

For instrumented replay, copy the original cohort producer to its named outputs/final-gradient-development-v1/run.py path in a checkout with fresh output directories. At the baseline commit run it with --stage baseline --source 1e6da103bf1d268d6469cb60dce269099dae53ea; at the numerical commit use --stage current --source aa4a0c24069020624e4d6bb8c519c2d9d9838fcd. Keep the thread limits above and retain every outcome. The producer refuses to overwrite its stage directories. Timestamps, paths and platform floating-point behavior can differ on replay. Historical producer paths and source bindings are retained as originally recorded; none is a current stopping-policy implementation.