Reproduction Scoreboard

This page is the short, shareable evidence view for the GLIL CPU reproduction fork. It is designed for readers deciding whether the fork is worth trying or starring.

Reproduction scorecard

Headline

MegaParticles APE checks: 3/3 PASS for Track B and Track C.
Deterministic hard recipe: outdoor_hard_01a is 5/5 byte-identical with RMSE 0.906313.
Official Ouster sample: completes cleanly with 1122 pose rows, fallback 0, bitset abort 0.
The Ouster smoke result is WARN only because there is no bundled ground-truth APE file for that sample.

Results

dataset	kind	status	RMSE	upstream GLIM	playback mean	Track B	Track C	recommended config
`indoor_easy_01`	APE	PASS	`1.019250`	`3.383012`	`1.000x`	PASS	PASS	`config_fair_glil_true_sample_t128_indoor_d4k_k1_rw_csp15_ct64_lag4`
`outdoor_hard_01a`	APE	PASS	`0.906313`	`4.321651`	`1.000x`	PASS	PASS	`config_fair_glil_true_sample_t128_hard_csp15_ct64_lag4_ffb100_skip16`
`outdoor_kidnap_a`	APE	PASS	`20.349845`	`21.701012`	`1.000x`	PASS	PASS	`config_fair_glil_true_sample_t128_k1`
`os1_128_01_downsampled`	smoke	WARN	NA	NA	`0.201x`	NA	NA	`config_official_os1_128_01_downsampled_acc1`

Track B means the GLIL RMSE is within upstream GLIM RMSE + 20%. Track C means mean playback is at least 0.95x.

What To Try First

Start with outdoor_hard_01a and config_fair_glil_true_sample_t128_hard_csp15_ct64_lag4_ffb100_skip16 if you want the strongest determinism signal.
Use indoor_easy_01 when you want a quick accuracy sanity check with a large margin over the upstream GLIM baseline.
Use os1_128_01_downsampled when you want an official sample smoke check and IMU scaling guardrail.

Shareable Summary

GLIL CPU Reproduction Fork: reproducible LiDAR odometry configs for MegaParticles and Ouster smoke checks. 3/3 Track B+C PASS, outdoor_hard_01a RMSE 0.906313, and the hard recipe is 5/5 byte-identical.

Perception Readiness

For perception-enabled runs, generate a compact readiness table and include it next to RMSE, playback, and trajectory hash evidence:

glil_perception_report_summary \
  --report dataset_a=dataset_a/perception_report.csv \
  --report dataset_b=dataset_b/perception_report.csv \
  --format markdown \
  --output perception_summary.md

This keeps accepted factor counts, rejected-observation reasons, and timestamp match rates visible before a run is promoted into the scoreboard.

Reproduction Bundle

Use glil_reproduction_bundle to package the command, config pointer, perception readiness, trajectory metrics, benchmark CSV, and artifact list for a run before opening a reproduction issue:

glil_reproduction_bundle \
  --run-dir dataset_a \
  --command-log dataset_a/command.txt \
  --ape dataset_a/evo_ape.txt \
  --perception-report dataset_a/perception_report.csv \
  --benchmark-csv dataset_a/coreset_benchmark.csv \
  --format markdown \
  --output dataset_a/reproduction_bundle.md

The bundle keeps paths display-safe by preferring run-directory or repository relative paths, and it marks missing supplied artifacts as warnings instead of silently ignoring them. Use --format json when a script needs to ingest the same evidence.

Example: outdoor_hard_01a reproduction

A bundle generated from the outdoor_hard_01a reproduction run (results/latest_glim/outdoor_hard_01a_reproduce_20260421/) looks like this in Markdown:

# GLIL outdoor_hard_01a reproduction (2026-04-21)

| field | value |
|---|---|
| status | PASS |
| git commit | `a97ec89bd24a98bfd1a1fe494a7b21d1a89398b9` |
| run directory | `results/latest_glim/outdoor_hard_01a_reproduce_20260421` |

## Artifacts

| kind | label | path | exists | bytes |
|---|---|---|---:|---:|
| command | reproduce_glil | `scripts/reproduce_glil.sh` | yes | 6168 |
| metric | evo_ape | `evo_ape.txt` | yes | 498 |
| artifact | git_ref | `git_ref.txt` | yes | 239 |
| artifact | contract_playback | `contract_playback.txt` | yes | 28 |
| artifact | summary | `reproduce_glil_summary_20260421.md` | yes | 925 |
| artifact | reproduce_manifest | `reproduce_manifest.json` | yes | 2933 |

## Trajectory Metrics

| label | extracted metrics |
|---|---|
| evo_ape | max=2.692416, mean=0.796697, median=0.751121, min=0.107598, rmse=0.906313, sse=2347.571431, std=0.432062 |

## Notes

- Track B + Track C PASS (RMSE 0.906313 m vs upstream GLIM 4.321651 m; playback 1.0000x).
- Trajectory byte-identical across 5/5 repeats (BZ5 cached immutable).
- Dataset: MegaParticles outdoor_hard_01a; config `config_fair_glil_true_sample_t128_hard_csp15_ct64_lag4_ffb100_skip16`.

The command that produced it:

glil_reproduction_bundle \
  --repo-root . \
  --run-dir results/latest_glim/outdoor_hard_01a_reproduce_20260421 \
  --ape results/latest_glim/outdoor_hard_01a_reproduce_20260421/evo_ape.txt \
  --command-log scripts/reproduce_glil.sh \
  --git-commit a97ec89bd24a98bfd1a1fe494a7b21d1a89398b9 \
  --artifact git_ref=results/latest_glim/outdoor_hard_01a_reproduce_20260421/git_ref.txt \
  --artifact contract_playback=results/latest_glim/outdoor_hard_01a_reproduce_20260421/contract_playback.txt \
  --artifact summary=results/latest_glim/reproduce_glil_summary_20260421.md \
  --artifact reproduce_manifest=reproduce_manifest.json \
  --title "GLIL outdoor_hard_01a reproduction (2026-04-21)" \
  --note "Track B + Track C PASS (RMSE 0.906313 m vs upstream GLIM 4.321651 m; playback 1.0000x)" \
  --note "Trajectory byte-identical across 5/5 repeats (BZ5 cached immutable)" \
  --format markdown \
  --output results/latest_glim/outdoor_hard_01a_reproduce_20260421/reproduction_bundle.md

This is the bundle shape expected when attaching evidence to a reproduction issue: RMSE and playback evidence is front-loaded, artifact presence is explicit, and every displayed path is repository- or run-directory relative.

The same tool was used to produce bundles for all three MegaParticles reproduction runs:

dataset	run directory	RMSE	Track B	Track C
`indoor_easy_01`	`results/latest_glim/indoor_easy_01_quiet_trackc_retry4_20260421`	`1.019250`	PASS	PASS
`outdoor_hard_01a`	`results/latest_glim/outdoor_hard_01a_reproduce_20260421`	`0.906313`	PASS	PASS
`outdoor_kidnap_a`	`results/latest_glim/outdoor_kidnap_a_reproduce_20260421`	`20.349845`	PASS	PASS

Each run directory contains a reproduction_bundle.md generated from the command above with the per-dataset --run-dir, --title, and --note values adjusted. Attach that bundle directly when filing a reproduction report for that dataset.

Reproduction Reports

If you try another dataset, open a reproduction report with the dataset name, config, RMSE or smoke markers, playback speed, and trajectory hash. Reports that include enough metadata can be promoted into this scoreboard.