Pokémon Random Battle
- Score
- 3.3
- Wallet spend
- $0.42 of $10
- Model cost
- $21.39
- Agent time
- 2h 58m
- Steps
- 305
- Tokens in / out
- — / —
- Self-checks
- 3 passed of 3 used
- Rank at $10 wallet
- 2nd of 19
Score
3.3raw metric 0.5163
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
- 1549
- draws
- 0
- losses
- 1451
- capped_games
- 0
- opponent_forfeits
- 0
- move_timeout_forfeits
- 0
- paired_standard_error
- 0.0082
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 and NOTICE.md declare the reused components with their revisions and licenses: Foul Play (pmariglia, r
architecture.json and NOTICE.md declare the reused components with their revisions and licenses: Foul Play (pmariglia, rev 6c467c08…), poke-engine 0.0.48 and pokemon-showdown. Reusing existing bots is allowed. Trusted audits approved both poke-engine artifacts (the sdist and the cp312 wheel), which also covers the uninspected engine tarball. The only learned component is a small NumPy MLP. Its inventory in model.safetensors (feat_mean/std[36], w_lin[36], w1[36,48], b1[48], w2[48,1], b2[1]) matches features.py and train.py. It was trained from a random init by MCTS-teacher distillation on self-generated games from schedule(240,'train') with seed 7777777. trajectory#188 launched that training and trajectory#193 shows its log, which matches train.log and training.json (8534 states, 80 epochs, top1 0.393→0.487). trajectory#195–196 downloaded the 8180-byte weights. One reproducibility gap: train.py imports prepare_stratified_battles, which disappeared from bot_engine.py when it was rewritten after training (trajectory#198). This is not a provenance violation, because the run and the weights are otherwise consistent.
✓actual served policyserve.sh installs local wheels with --no-index and runs server.py. server.py's /act sends each observation to bot_engine
serve.sh installs local wheels with --no-index and runs server.py. server.py's /act sends each observation to bot_engine.decide_observation, which rebuilds the battle from public_log and the request using Foul Play's protocol parser, samples opponent sets with prepare_random_battles, and runs poke-engine monte_carlo_tree_search. It picks the legal command with the most weighted visits; POLICY_NET (features.PolicyNetwork loading model.safetensors) adds a 1e-5 tiebreak. The MLP argmax is the fallback on timeouts or exceptions. Set priors come from the local foul-play-priors.json through EnhancedRandomBattleMode, so Foul Play's remote randbats fetch is bypassed. No remote decision service or undeclared path exists in the serving call graph.
✓permitted informationDecisions use only the observation's public_log, the player's own request (including the Zoroark preprocessing of the pl
Decisions use only the observation's public_log, the player's own request (including the Zoroark preprocessing of the player's own team), legal_actions and action_features, plus the static public set priors. I found no simulator inspection, opponent private state, test-set detection or cached evaluation actions in server.py, bot_engine.py or features.py. Training labels came from the 'train' schedule split. The participant viewed validation data and ran /task/validate.py against the local server, which is ordinary development validation. Nothing from validation is cached into the serving path.
✓harness credentials used only by the harnessThe trajectory (antigravity-cli harness) contains no participant reads, copies or uses of ANTHROPIC/OPENAI/OPENROUTER/GE
The trajectory (antigravity-cli harness) contains no participant reads, copies or uses of ANTHROPIC/OPENAI/OPENROUTER/GEMINI keys or *_BASE_URL variables. Searches found only the Spin documentation's SPIN_API_KEY instructions and Foul Play's PKMN_RANDBATS_REMOTE_BASE_URL constant. Remote compute used the permitted Spin CLI: spin compute upload/run/download at trajectory#183, #188 and #195. curl was used only for local 127.0.0.1 health checks. No extra harness CLI sessions were launched.
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 3 of 4.
- 1passed1h 56m into the runtook 2m 33s
startup 1.8 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 1.8 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, 8, 1, 5, 9, 11, 5, 7, 5, 11, 11, 9, 0, 11, 5, 0, 5, 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), 83 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3534 observations, 23-32 per batch, up to 338 KB per request, every answered action legal; not scored
- ✓move response within 1 s · slowest of 124 POST /act responses (23-32 observations, up to 338 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.756 s, p95 0.761 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 0.756 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
- 2passed2h 39m into the runtook 3m 3s
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, 1.02 s round trip on this side with the transport: [5, 7, 3, 7, 5, 5, 7, 11, 10, 11, 7, 0, 9, 11, 1, 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), 81 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3533 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 0.774 s, p95 0.778 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
- 3passed2h 54m into the runtook 2m 33s
startup 1.5 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 1.5 s
- ✓POST /act, round 1 · 32 opening decisions answered with legal actions, 0.87 s round trip on this side with the transport: [1, 7, 3, 7, 5, 5, 7, 11, 10, 11, 7, 0, 9, 11, 1, 5, 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), 80 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3568 observations, 24-32 per batch, up to 349 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 349 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.775 s, p95 0.784 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