Pokémon Random Battle
- Score
- 1.8
- Wallet spend
- $3.45 of $50
- Model cost
- $23.08
- Agent time
- 10h 22m
- Steps
- 564
- Tokens in / out
- 67.7M / 802k
- Self-checks
- 4 passed of 4 used
- Rank at $50 wallet
- 4th of 14
Score
1.8raw metric 0.509
The raw metric (mean battle points against Foul Play) is rescaled so that 50% of battle points against Foul Play scores 0 and 100% scores 100, clamped to that range.
Verifier details
- metric
- mean_battle_points
- samples
- 3000
- wins
- 1527
- draws
- 0
- losses
- 1473
- capped_games
- 0
- opponent_forfeits
- 0
- move_timeout_forfeits
- 0
- paired_standard_error
- 0.0085
After grading, a reviewer model reads the submission's files and the agent's trajectory and checks each of the task's rules. A failed check makes the score 0. Step numbers link into the trajectory.
✓implementation provenanceThe submission declares every reused component. architecture.json, PROVENANCE.md, source/DATA_MANIFEST.json and source/l
The submission declares every reused component. architecture.json, PROVENANCE.md, source/DATA_MANIFEST.json and source/logs/artifact-acquisition.json list Foul Play at revision 6c467c08 under vendor/fp, the poke-engine 0.0.48 wheel and the task-provided priors.json. The critic is initialized from the public MintOcha critic.onnx (revision 14eccd10, base sha 2ec12c85…) and then fine-tuned on private generated games: seeds 947132 and 398375, 170,668 examples, seed 32719, 4 epochs. The trajectory matches this account. It shows the MintOcha download and its metadata (trajectory#210) and the copy into output (trajectory#220). The fine-tune ran on the remote machine with --base output/critic.onnx, 4 epochs and seed 32719 (trajectory#485). Its log reports 170,668 samples (trajectory#502). The resulting graph and safetensors were copied into output (trajectory#503). A trusted audit approves the poke-engine wheel as organizer-built. critic.safetensors has a transformer-critic tensor layout consistent with the declared fine-tune. Research-only downloads were not shipped. No public validation or test data was used.
✓actual served policyserve.sh installs the shipped wheels offline, sets PYTHONPATH to vendor and runs server.py. server.py loads the critic (
serve.sh installs the shipped wheels offline, sets PYTHONPATH to vendor and runs server.py. server.py loads the critic (critic_model.load_model) and serves /act. It returns policy.search results through a ProcessPool, with a feature-based fallback under the deadline. policy.search runs Foul Play belief-world sampling and poke-engine MCTS. value_model.rank_candidates then reranks the top candidates by expanding afterstates and scoring them with critic_model. critic_model runs the shipped critic.onnx graph with its parameters loaded from the shipped critic.safetensors. The settings in config.json (value, critic and critic_public on; mix 0.75) match this call graph. The served policy is the declared one, and there are no outbound network calls.
✓permitted informationThe policy builds its state only from the observation's public_log and the bot's own request (state_builder.py). It adds
The policy builds its state only from the observation's public_log and the bot's own request (state_builder.py). It adds opponent sets sampled from the task priors.json and a damage posterior. critic_features.encode masks hidden opponent information when public mode is on: unrevealed reserves, stats, exact HP, PP and roles are zeroed, and tera type appears only after terastallization is revealed. No process inspection, simulator peeking or hidden-state access was found. The only file writes are a temp snapshot cache keyed by observation and env-gated traces. The training corpora are private self-generated games declared in source/DATA_MANIFEST.json.
✓harness credentials used only by the harnessI searched the trajectory for API_KEY, BASE_URL, OPENAI, printenv, codex exec and openai imports. No step used the harne
I searched the trajectory for API_KEY, BASE_URL, OPENAI, printenv, codex exec and openai imports. No step used the harness model credentials for the participant's own purposes. The remote compute used the participant's own vast.ai SSH keys, which is permitted. The shipped server makes no outbound API calls.
LLM adversarial review is not proof of arbitrary program behavior. Host storage and task-specific deterministic checks remain required.
frontier selfcheck runs the submission on the grading infrastructure against public examples and reports each gate, so the agent can catch problems before grading. Each run gets four. This agent used 4 of 4.
- 1failed22m 32s into the runtook 11sfree (stopped before any GPU time)
failed review prerequisites: The review needs architecture.json and serve.sh at the top of the submission
1 gates
- ✗review prerequisites · The review needs architecture.json and serve.sh at the top of the submission
- 2passed24m 36s into the runtook 2m 21s
startup 1.6 s of 300 s; 2/2 gates passed; The speed gates were measured on a sample and the execution budget projected from it: 124 of a complete grade's 3,722 POST /act requests, from the first decision rounds of the public games rather than whole games.
7 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 1.6 s
- ✓POST /act, round 1 · 32 opening decisions answered with legal actions, 0.82 s round trip on this side with the transport: [5, 7, 1, 7, 5, 5, 7, 11, 9, 9, 7, 0, 9, 11, 9, 5, 9, 11, 5, 7, 5, 11, 11, 9, 0, 11, 5, 0, 9, 11, 6, 9]
- ✓battle slice · 124 POST /act batches over 124 decision rounds, 32 games at a time: the first 32 public games (teams from development seed 91827461), 90 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3594 observations, 24-32 per batch, up to 322 KB per request, every answered action legal; not scored
- ✓move response within 1 s · slowest of 124 POST /act responses (24-32 observations, up to 322 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.696 s, p95 0.863 s. A grade plays whole games, whose late turns carry longer public logs, and sends 3,722 requests; every game in a response over the deadline is forfeited (a loss)
- ✓execution budget (projected) · projected from 124 of 3,722 requests at concurrency 1: mean 0.719 s per request x 3,722, as many as one complete 3,000-game grade sent; grading sums every request's in-sandbox time
- ✓healthy after requests
- 3passed4h 35m into the runtook 3m 7s
startup 4.9 s of 300 s; 2/2 gates passed; The speed gates were measured on a sample and the execution budget projected from it: 124 of a complete grade's 3,722 POST /act requests, from the first decision rounds of the public games rather than whole games.
7 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 4.9 s
- ✓POST /act, round 1 · 32 opening decisions answered with legal actions, 1.09 s round trip on this side with the transport: [5, 7, 1, 7, 5, 5, 7, 11, 10, 11, 7, 0, 9, 11, 9, 11, 9, 11, 0, 7, 5, 11, 11, 9, 0, 11, 5, 3, 7, 11, 8, 9]
- ✓battle slice · 124 POST /act batches over 124 decision rounds, 32 games at a time: the first 32 public games (teams from development seed 91827461), 82 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3532 observations, 24-32 per batch, up to 329 KB per request, every answered action legal; not scored
- ✓move response within 1 s · slowest of 124 POST /act responses (24-32 observations, up to 329 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.919 s, p95 0.928 s. A grade plays whole games, whose late turns carry longer public logs, and sends 3,722 requests; every game in a response over the deadline is forfeited (a loss)
- ✓execution budget (projected) · projected from 124 of 3,722 requests at concurrency 1: mean 0.919 s per request x 3,722, as many as one complete 3,000-game grade sent; grading sums every request's in-sandbox time
- ✓healthy after requests
- 4passed8h 25m into the runtook 3m 1s
startup 5.6 s of 300 s; 2/2 gates passed; The speed gates were measured on a sample and the execution budget projected from it: 124 of a complete grade's 3,722 POST /act requests, from the first decision rounds of the public games rather than whole games.
7 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 5.6 s
- ✓POST /act, round 1 · 32 opening decisions answered with legal actions, 1.07 s round trip on this side with the transport: [5, 7, 1, 7, 5, 0, 7, 11, 10, 9, 7, 4, 7, 11, 1, 11, 9, 9, 0, 7, 5, 11, 11, 9, 0, 11, 5, 0, 9, 11, 8, 9]
- ✓battle slice · 124 POST /act batches over 124 decision rounds, 32 games at a time: the first 32 public games (teams from development seed 91827461), 91 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3529 observations, 25-32 per batch, up to 333 KB per request, every answered action legal; not scored
- ✓move response within 1 s · slowest of 124 POST /act responses (25-32 observations, up to 333 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.772 s, p95 0.815 s. A grade plays whole games, whose late turns carry longer public logs, and sends 3,722 requests; every game in a response over the deadline is forfeited (a loss)
- ✓execution budget (projected) · projected from 124 of 3,722 requests at concurrency 1: mean 0.775 s per request x 3,722, as many as one complete 3,000-game grade sent; grading sums every request's in-sandbox time
- ✓healthy after requests
- 5passed10h 17m into the runtook 2m 55s
startup 5.0 s of 300 s; 2/2 gates passed; The speed gates were measured on a sample and the execution budget projected from it: 124 of a complete grade's 3,722 POST /act requests, from the first decision rounds of the public games rather than whole games.
7 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 5.0 s
- ✓POST /act, round 1 · 32 opening decisions answered with legal actions, 0.93 s round trip on this side with the transport: [5, 7, 2, 7, 5, 5, 7, 11, 10, 9, 7, 0, 9, 11, 2, 5, 9, 5, 5, 7, 5, 11, 11, 9, 0, 11, 5, 0, 9, 11, 8, 10]
- ✓battle slice · 124 POST /act batches over 124 decision rounds, 32 games at a time: the first 32 public games (teams from development seed 91827461), 89 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3563 observations, 24-32 per batch, up to 330 KB per request, every answered action legal; not scored
- ✓move response within 1 s · slowest of 124 POST /act responses (24-32 observations, up to 330 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.773 s, p95 0.826 s. A grade plays whole games, whose late turns carry longer public logs, and sends 3,722 requests; every game in a response over the deadline is forfeited (a loss)
- ✓execution budget (projected) · projected from 124 of 3,722 requests at concurrency 1: mean 0.776 s per request x 3,722, as many as one complete 3,000-game grade sent; grading sums every request's in-sandbox time
- ✓healthy after requests
The agent's last message before its session ended.
Submitted /workspace/output successfully.
- Belief-sampled native search with a fine-tuned public-view critic.
- Readable source, provenance, training logs, and safetensors weights included.
- Final H100 selfcheck passed; all actions legal, slowest batch 0.849 seconds.
- Public validation completed using a relaxed local diagnostic deadline.