Skip to content

ADR 0010: v0.5 Direction — Reconcile v0.4 and Ship the Operational Trust Platform

Status

Proposed (2026-09-27). Reconciles ADR 0008, which remains Proposed as a historical record and is not rewritten.

Context

ADR 0008 set three v0.4 pillars: full-scale Camera-LiDAR PASS on real KITTI, Radar as an evaluated modality, and a declared capture-time convention. The implementation has since moved past that snapshot:

  • Pillar 1 (KITTI full scale). An opt-in @pytest.mark.kitti falsification benchmark shows the vendor reference PASS and a declared known-bad FAIL without retuned gates. It is not yet published as a maintained benchmark artifact.
  • Pillar 2 (Radar). Radar has a protocol, holdout, signed controls, robust ego velocity, trajectory rotation/yaw, lever-arm/time, joint spatiotemporal solvers, a nuScenes adapter, radar_msgs/RadarScan intake, and a radar-service replacement gate. Public nuScenes excitation stays too weak for a conclusive full-system PASS. The limit is the data, not missing algorithms.
  • Pillar 3 (capture time). capture_time_reference and per-point capture time exist, but TIERS real-data evidence is INCONCLUSIVE. No end-to-end acceptance run has shown message time and capture time converging under the declared convention.
  • Deferred list. Native continuous-time IMU lever-arm, clock-offset, accelerometer-bias/gravity, and intrinsics factors now exist with synthetic holdout and signed controls. PyPI publication remains skipped. Installation uses the verified GitHub Release wheel.

Work between 2026-08-21 and 2026-08-25 also added an operational trust layer that ADR 0008 did not anticipate:

  • capture manifests;
  • digest-bound raw replay;
  • an append-only lifecycle registry;
  • multi-LiDAR, camera--IMU, and radar service-replacement gates;
  • Autoware promotion and smoke gates;
  • a fail-closed Koide real-pilot handoff;
  • a required slac.result.provenance/v0.1 contract.

On 2026-09-27, calibrex external-run evaluate-camera-imu added the same Calibrex-side falsification evidence for Kalibr outputs. The Koide pilot already applies it to its handoff.

Decision

  1. Close Pillars 1 and 2 as protocol work. Their remaining items are evidence publication (a maintained KITTI benchmark artifact) and data acquisition (Radar captures with stronger excitation). They no longer block a release, and new Radar algorithms are not a priority.
  2. Carry Pillar 3 forward unchanged. The acceptance condition stays as ADR 0008 wrote it: the shared-clock anchored baseline must fall within ±10 ms under the declared convention, and the injection differential must stay exact. Gates are not relaxed.
  3. Keep the trust platform release-ready, but do not cut a release yet. The maintainer decided on 2026-09-27 not to release for now. Any future release still follows the adoption-track gate: a clean wheel, no schema drift, a working quickstart, and a green default test suite. It would be distributed as a GitHub Release wheel, not on PyPI.
  4. Treat the external-run contract as proven for Koide and Kalibr. An external tool's own metrics stay non-comparable. Every imported candidate still needs Calibrex-side holdout and signed-control evidence before adoption. Kalibr lever-arm (translation) evidence stays open.
  5. Favour user-facing paths. Priorities are a single LiDAR-IMU calibration command, lifecycle-registry integration with Calibration CI, and the Pillar 3 acceptance demonstration, ahead of new solver families.

Consequences

The development roadmap becomes the dated inventory for v0.5 work, and this ADR records why the v0.4 pillars close. Large public datasets remain outside the repository. Opt-in tests read them from environment-variable paths and skip when the data are absent, so the default suite stays small and offline.

The project rules still apply: a ROS-independent core, no GPL code in src/calibrex, typed public APIs, schema-validatable artifacts, and provenance on every generated result.