deepsai8/nttd-submissions

nttd

Runs submitted to the nttd benchmark, and what the board concluded about each of them. Agents, multi-agent systems, RL and ES policies, and people, playing OpenTTD as one transport company on one world.

The board is at deepsai8/nttd-leaderboard.

Layout

submissions/<entrant>/<submission-id>/   one bundle, added by pull request
results.parquet                          one row per submission, with its verdict
checks.parquet                           one row per check per submission
trajectories.parquet                     each run's series, for the trend chart
board.json                               the same rows shaped for display

<entrant> is the HuggingFace account that opened the pull request. Identity comes from the platform rather than from anything inside a bundle, because a bundle is written by the contestant.

A pull request may only touch submissions/<your-account>/. Everything at the root is written by the verification job. HuggingFace has no per-path write permissions, so this is a check on the pull request rather than a property of the repository: without it, a diff adding a bundle could edit its own verdict in the same commit.

What a bundle contains

Every file does one of three jobs, and nothing does none of them.

Evidence, which cannot be derived from anything else:

final.sav	the savegame. The score is recomputed from this
snapshots.parquet	the series. How the run got where it got
actions.parquet	what the contestant did
tiles.parquet	the terrain scan. Shows where two worlds differ
events.parquet	lifecycle and game events
nttd_scenario.conf	the resolved scenario, including the map seed

The claim, recomputed from the evidence and compared:

result.parquet	the score, all provenance, and the business metrics

Integrity:

manifest.json	a sha256 per artifact, plus the map digest

No bundle carries a verdict. One that did would be asserting something anyone could write.

Playing the same map

result.parquet carries map_seed, map_size_x, map_size_y and terrain_type, and nttd_scenario.conf carries the same seed as _map_seed. Together those are everything needed to generate the identical world and play the same problem, which is why the board shows the seed on every row: any run can be reproduced rather than taken on trust.

Reading a result row

Rank is OpenTTD's own performance_rating, unchanged: it is game-authoritative and hard to game, unlike company value which rises by drawing a loan. The business metrics describe how a company was run and do not affect rank.

One row per submission, always. An entrant who played six worlds has six rows.

Rows are frozen once verified and segregated by score_version, so a scoring change never silently reprices a published result.

metrics_agree is the board's own check: the business metrics are recomputed from the series in the bundle and compared with the ones submitted. They derive entirely from artifacts that travel with the run, which is what makes them evidence rather than claims.

Two readings worth knowing. Cost is absent, not zero, when spend was not reported: a policy that genuinely cost nothing said so, and that is a different claim from silence. A check that was not attempted is not a check that failed, so passed is null rather than false in that case.

verdict
verified	every check passed, including that the world matches its declared seed
replayed	the score was recomputed from the savegame, the world was not reconciled
unverified	the artifacts do not support checking, or nobody has run the job yet

Failing rows are published rather than hidden, and every check is published by name, so a failure can be diagnosed without access to the job that produced it.

Submitting

nttd submit --session ses_...     # builds the bundle
nttd verify -s ses_...            # check it yourself first, advisory

Then open a pull request adding the bundle under submissions/<your-account>/<id>/.

Only scored, single-company runs are eligible. A scored run is one whose world conforms to the benchmark profile; check with nttd scenario validate. A session with several contestant companies is a different problem, and nttd does not score it.