Pokémon Random Battle
- Score
- 0.0
- Wallet spend
- $1.84 of $50
- Model cost
- $7.99
- Agent time
- 1h 52m
- Steps
- 280
- Tokens in / out
- 26.5M / 154k
- Self-checks
- 2 passed of 2 used
- Rank at $50 wallet
- 7th of 19
Score
0.0raw metric 0.4863
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
- 1459
- draws
- 0
- losses
- 1541
- capped_games
- 0
- opponent_forfeits
- 0
- move_timeout_forfeits
- 0
- paired_standard_error
- 0.0083
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 provenancearchitecture.json declares no base model and no learned weights. It names the reused code: Foul Play at revision 6c467c0
architecture.json declares no base model and no learned weights. It names the reused code: Foul Play at revision 6c467c08 (GPL-3.0, from /task/public/foul-play.tar.gz) and poke-engine 0.0.48, shipped as the task-provided wheel. The trusted audit approves that wheel and its SHA-256 matches the declared one. No weight inventory exists, and nothing in the evidence shows a checkpoint. In the session, the participant copied fp/ from the extracted upstream source, the priors from /task/public, and the engine wheel from /task/public/engine (trajectory#7). No edits to the shipped fp/ files appear; the session only reads them. No training was done. The development schedules used to choose settings came from the public generator, with seeds declared in architecture.json and provenance/.
✓actual served policyserve.sh installs the bundled wheels offline and starts server.py. POST /act checks the request and calls policy.choose_
serve.sh installs the bundled wheels offline and starts server.py. POST /act checks the request and calls policy.choose_batch, which hands observations to persistent worker processes started locally with fork. Each worker rebuilds Foul Play battle state from public_log and the request, samples possible hidden opponent teams from the local foul-play-priors.json, runs poke_engine.monte_carlo_tree_search locally, and picks the move with the best visit-weighted value. Other paths are a forced single-action shortcut, a revival search that also uses MCTS, and a tactical_action fallback that scores the supplied action_features. All of these are declared in architecture.json and run offline. No remote decision service is called.
✓permitted informationInference reads only the observation fields player, public_log, request, turn, legal_actions and action_features. Hidden
Inference reads only the observation fields player, public_log, request, turn, legal_actions and action_features. Hidden opponent sets are sampled from the public generator-frequency priors. policy.py sets raw_pkmn_sets directly from the local JSON, so the networked datasets.initialize/get_sets_file path is never called while serving. I found no simulator inspection, RNG access, test detection or cached evaluation actions. Selection used dev schedules made independently by the public generator (trajectory#70-94 area). The official validator was run only as a final smoke check (trajectory#258-270).
✓harness credentials used only by the harnessSearching the trajectory found no access to the harness's API keys or base URL, no direct model API calls, no extra clau
Searching the trajectory found no access to the harness's API keys or base URL, no direct model API calls, no extra claude, codex or other agent CLI sessions, and no credentials copied into files or machines. The only remote use was Spin compute for development battles (e.g., trajectory#24, trajectory#38), which the policy allows.
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 2 of 4.
- 1passed1h 14m into the runtook 2m 36s
startup 1.4 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 1.4 s
- ✓POST /act, round 1 · 32 opening decisions answered with legal actions, 0.85 s round trip on this side with the transport: [1, 7, 3, 7, 5, 5, 7, 11, 10, 9, 7, 0, 7, 11, 1, 11, 9, 9, 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), 81 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3544 observations, 24-32 per batch, up to 314 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 314 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.667 s, p95 0.713 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 0.672 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
- 2passed1h 48m into the runtook 2m 40s
startup 1.7 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 1.7 s
- ✓POST /act, round 1 · 32 opening decisions answered with legal actions, 0.88 s round trip on this side with the transport: [5, 7, 3, 7, 5, 5, 7, 11, 10, 9, 7, 0, 9, 11, 9, 11, 9, 11, 5, 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), 79 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3549 observations, 24-32 per batch, up to 334 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 334 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.681 s, p95 0.751 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 0.688 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.
The final H100 selfcheck passed on digest c077e5d6a90b76b746d594d7c17e242740b4bf0197edef539d35620bad2481a4: 1.7 s startup, every sampled action legal, 0.8153 s slowest response, and a projected 2,562.5 s response budget. The server remained healthy. I’m releasing the development host and submitting that exact package.