Pokémon Random Battle
- Score
- 0.0
- Wallet spend
- $0.80 of $200
- Model cost
- $0.78
- Agent time
- 1h 58m
- Steps
- 320
- Tokens in / out
- 32.3M / 265k
- Self-checks
- 3 passed of 3 used
- Rank at $200 wallet
- 6th of 19
Score
0.0raw metric 0.4953
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
- 1486
- draws
- 0
- losses
- 1514
- capped_games
- 1
- opponent_forfeits
- 1
- 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 provenanceThis bot is not learned. architecture.json sets base_model, base_revision and weights to "none", and the submission ship
This bot is not learned. architecture.json sets base_model, base_revision and weights to "none", and the submission ships no checkpoint. Every reused component is declared: - Foul Play (GPL-3.0, pinned revision) lives at vendor/foulplay. The session re-extracted it from /task/public/foul-play.tar.gz without changes (trajectory#202). - The poke-engine 0.0.48 wheel matches the organizer-built wheel the trusted audit approved (sha256 b91209b2…). - The priors file comes from /task/public and is declared with its sha256. - The helper wheels (requests, websockets, dateutil) were fetched with pip download (trajectory#192). Selection_log.json records only evaluations on the dev seeds, which matches the declared design.
✓actual served policyThe serving path runs in this order: 1. serve.sh installs the vendored wheels offline, sets PYTHONPATH and starts server
The serving path runs in this order: 1. serve.sh installs the vendored wheels offline, sets PYTHONPATH and starts server.py. 2. server.py exposes POST /act and forks workers that call core.initialize with the local priors. 3. core.Bot.decide rebuilds Foul Play battle state, samples opponent sets, runs poke-engine monte_carlo_tree_search and picks a move with the random75 rule. 4. core.to_index maps that move onto legal_actions. Any observation not answered by the deadline gets a deterministic feature-weighted legal action instead. This is the approach that architecture.json and README.md declare. Nothing hidden is called and there are no remote calls.
✓permitted informationDecisions use only what the observation supplies (public_log, the player's own request, legal_actions, action_features)
Decisions use only what the observation supplies (public_log, the player's own request, legal_actions, action_features) plus the local public priors. core.initialize loads set data from the local priors JSON, so the serving path never calls Foul Play's remote fetch (datasets.initialize / requests.get). The code reads no simulator internals, opponent private state or cached answers. Strategy was selected on the dev seeds 424242-424246. The participant only looked at the validation file during the session; it does not feed into decisions.
✓harness credentials used only by the harnessI searched the whole trajectory for API keys, Anthropic/OpenRouter strings, BASE_URL, printenv and claude CLI launches.
I searched the whole trajectory for API keys, Anthropic/OpenRouter strings, BASE_URL, printenv and claude CLI launches. None of the matches showed the participant reading, using or copying harness credentials or base URLs, or launching extra harness CLIs. The only remote compute used was Spin through its own SSH key, which the policy allows (trajectory#91, trajectory#92). curl was used only against localhost health endpoints. The final submission steps contain no credential use (trajectory#319).
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 36m into the runtook 3m 20s
startup 2.4 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 2.4 s
- ✓POST /act, round 1 · 32 opening decisions answered with legal actions, 1.04 s round trip on this side with the transport: [5, 7, 3, 7, 5, 0, 7, 11, 9, 11, 7, 0, 9, 8, 1, 11, 9, 5, 0, 7, 5, 11, 11, 9, 0, 11, 5, 1, 4, 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), 88 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3545 observations, 25-32 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 (25-32 observations, up to 322 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.750 s, p95 0.777 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 0.753 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 50m into the runtook 3m 13s
startup 2.1 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 2.1 s
- ✓POST /act, round 1 · 32 opening decisions answered with legal actions, 0.97 s round trip on this side with the transport: [5, 7, 3, 7, 5, 5, 7, 11, 11, 9, 7, 0, 9, 7, 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), 80 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3577 observations, 21-32 per batch, up to 357 KB per request, every answered action legal; not scored
- ✓move response within 1 s · slowest of 124 POST /act responses (21-32 observations, up to 357 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.744 s, p95 0.796 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 0.748 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
- 3passed1h 54m into the runtook 2m 47s
startup 2.0 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 2.0 s
- ✓POST /act, round 1 · 32 opening decisions answered with legal actions, 0.89 s round trip on this side with the transport: [5, 7, 3, 7, 5, 0, 7, 11, 10, 9, 7, 0, 9, 11, 1, 5, 9, 11, 0, 7, 5, 11, 11, 9, 0, 11, 5, 3, 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), 76 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3583 observations, 23-32 per batch, up to 354 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 354 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.745 s, p95 0.792 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 0.749 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 at 2026-10-20 21:33:40 UTC.
Final submission summary
- Bot: poke-engine MCTS over opponent-set determinizations, driving the pinned Foul Play battle parser/inference/set sampler (unmodified, rev
6c467c081e) reconstructed from each stateless observation's public log and private request. Two samples per decision, random75 selection, 0.72 search safety factor, 0.78 s per-batch target. - Serving:
serve.shinstalls the vendored wheels and the exact organizer-built poke-engine 0.0.48 cp312 wheel offline into/tmp, then runs an 8-worker pool (server.py) where each game keeps a warm reconstructed battle on one worker for the whole battle. Worker deaths are reaped and their games rebuilt from the visible log; every failure path returns legal actions with HTTP 200. - Validation: Development selection used only fresh
schedule(32,'dev')seeds (424242–424246) against the pinned Foul Play opponent, scoring 0.53125/0.53125/0.53125/0.46875 (mean 0.5156) through the packaged bundle, with 0 late requests and 0.740–0.765 s max responses. - Selfchecks: Three consecutive passes on the grading image; the final one on the submitted bundle measured startup 2.0 s of 300 s, slowest response ~0.815 s of 1.0 s, projected execution budget ~2,800 s of 18,000 s, every action legal. A SIGKILLed worker mid-run was handled with score unchanged.
The session is complete; /workspace/output is graded as submitted.