KITTI camera-LiDAR v3: cross-start agreement rule (development protocol)

Status: frozen before any replay of this rule on real inputs; screened once, not carried forward. This page fixes one opt-in rule, how it is screened on the 55 spent KITTI raw drives and what the result can and cannot decide. It is development, not a pre-registration or an audit. No evaluation data is read.

Why this rule

On all 55 spent drives, the current rule judges ten turned-run axes more than 0.5 deg off. The 2011_09_28_0002 start trace shows the mechanism behind the largest errors. A single-axis fit from one profile start and a multi-axis fit from another are both evidenced. They agree on pitch (−2.85 and −3.05 deg) but differ on roll by 1.55 deg, and the adopted multi-axis fit reports that roll. No scalar adoption margin separates the two cases. The 3.5 margin also moves an error to 0036.

The rule

EdgeAlignmentOptions(start_agreement_deg=0.5). Every search start whose optimum observes at least one rotation DoF yields an evidenced optimum: the candidate start and each distinct profile start. The adoption rule and the final fit are unchanged. A rotation axis the final fit would report is withheld, recorded as unobservable with a reason, when any other evidenced optimum differs from the final fit on that axis by more than 0.5 deg. All optima are corrections relative to the same candidate. An optimum that observes no rotation DoF carries no evidence and is not compared.

0.5 deg is the check's existing observable rotation std bound (observable_rotation_std_deg). It is not chosen from any result: two supported optima farther apart than the std bound mean the axis is not determined to that bound. The option defaults to None, which gives bit-identical v2 behavior and is omitted from recorded options, so existing captures and caches stay valid. A unit test reproduces the 0002 pitch +3 trace's expected effect. That traced effect motivated this rule, so 0002 is not independent evidence for it.

Screen

1. Cached inputs, 23 drives. Replay the current rule and the new rule on the 23 drives whose captured NPZ inputs remain. The current-rule replay must match the captured evidence (1e-9).
2. Re-capture, 32 drives. Run tools/kitti_camera_lidar_v3_batch.py with --proposed-options '{"start_agreement_deg": 0.5}' on the other 32 drives, using the existing pins and retry rules.
3. Score all 55. tools/kitti_camera_lidar_v3_rule_screen.py takes every spent drive exactly once and applies the criteria below.

The rule is carried forward to drafting a v3 pre-registration only if every check holds on all 55 drives. The values are v2's gate values, fixed here before the replay:

Check	Requirement for the proposed rule
Baseline not flagged	no scored baseline is warn or fail
Baseline judged accuracy	every judged baseline axis within 0.75 deg
Turned judged accuracy	every judged turned-run axis within 1.5 deg
3 deg turns flagged	at most 10 % of counted 3 deg trials missed
1 deg turns flagged	at most 25 % of counted 1 deg trials missed
Judged-axis coverage	at least half of the scored baseline axes judged
Scored-drive coverage	at least two thirds of the 55 drives scored
Fewer large errors	fewer turned-run judged axes beyond 0.5 deg than the current rule (10)
No new large errors	none beyond 0.5 deg that the current rule did not have
No failed drives	no capture, parity or scoring failure

Otherwise it is not carried forward. No tolerance, criterion or rule detail changes after the screen. A different rule needs a new protocol. Failures and every per-drive result are published either way, as before.

Limits

All 55 drives are spent: they were used to motivate this rule and they include v2's evaluation drives. A pass is development evidence only. It neither supports a SOTA claim nor replaces a pre-registered test on unread recordings (KITTI-360 after its exposure audit). Translation is still not judged. Running the rule natively through calibrex check would need its CLI and schema wiring and a real-driver confirmation, both deferred until the rule is carried forward.

PYTHONPATH=src:. .venv/bin/python -c "from pathlib import Path; \
from tools import kitti_camera_lidar_v3_replay as r; \
r.compare_drives(Path('CACHE'), Path('OUT/cached'), drives=[...], \
proposed={'start_agreement_deg': 0.5})"
PYTHONPATH=src:. .venv/bin/python tools/kitti_camera_lidar_v3_batch.py run \
  docs/assets/kitti_camera_lidar_v3/spent_batch_20261031/pins.json WORK \
  --proposed-options '{"start_agreement_deg": 0.5}' --drive ...
PYTHONPATH=src:. .venv/bin/python tools/kitti_camera_lidar_v3_rule_screen.py \
  SCREEN.json OUT/cached/comparison.json LATEST_ATTEMPT_COMPARISONS...