Pokémon Random Battle
- Score
- 0.5
- Wallet spend
- $0.38 of $10
- Model cost
- $7.72
- Agent time
- 2h 9m
- Steps
- 244
- Tokens in / out
- 25.1M / 147k
- Self-checks
- 2 passed of 2 used
- Rank at $10 wallet
- 4th of 19
Score
0.5raw metric 0.5027
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
- 1508
- draws
- 0
- losses
- 1492
- 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 says base_model and base_revision are none and learned_weights is false. It declares reused code: Foul
architecture.json says base_model and base_revision are none and learned_weights is false. It declares reused code: Foul Play at commit 6c467c08 (GPL-3.0, foul_play/fp) and poke-engine 0.0.48 (MIT, shipped as a wheel). The submission has no checkpoint and no training corpora. The shipped poke_engine wheel matches an organizer-built wheel approved by the trusted audit. bot/priors.json is a copy of the task-supplied /task/public/foul-play-priors.json (trajectory#14-23 cp command), and its generator and seed are declared. The selection schedules come from the public task generator through development/make_schedule.py, and their seeds are recorded. In the trajectory, the participant only read the vendored foul_play files and never edited them, which matches the claim that Foul Play is pinned. The bot's own changes (bot/adapters.py protocol fixes and the search-budget change in bot/worker.py) are readable and described in the declared method.
✓actual served policyserve.sh installs the shipped wheels offline (--no-index) and runs server.py. POST /act validates the batch, then Search
serve.sh installs the shipped wheels offline (--no-index) and runs server.py. POST /act validates the batch, then SearchService.decide works out a feature-weighted heuristic fallback action, and forked worker processes do the real work. Each worker rebuilds the Foul Play Battle state from the observation's public_log and own request, samples hidden opponent sets using the local priors.json (RandomBattleMode.prepare_battles), and runs local poke_engine monte_carlo_tree_search. Results are combined with Foul Play's near-best selection rule, and the chosen move is mapped onto the legal_actions menu. The only other decision paths are the heuristic fallback in bot/heuristic.py and the direct choice when a single action is legal. There is no remote decision service, and datasets.initialize, which would fetch remote set data, is never called. The serving path matches the declared method.
✓permitted informationInference reads only the observation fields: player, turn, request (own private request), public_log, legal_actions and
Inference reads only the observation fields: player, turn, request (own private request), public_log, legal_actions and action_features. It also uses the public dex, the task-supplied random-set prior and the poke-engine mechanics. Nothing inspects the simulator process, reads opponent private state or RNG, uses the network, detects test cases or caches actions. RNG seeds are derived from the battle's own visible history. Development selection used independently generated public schedules, which the policy allows.
✓harness credentials used only by the harnessI searched the trajectory for API_KEY, BASE_URL, ANTHROPIC, curl, printenv and spin usage. curl only hit the local healt
I searched the trajectory for API_KEY, BASE_URL, ANTHROPIC, curl, printenv and spin usage. curl only hit the local health endpoint, plus one connectivity check of a GitHub release URL from the Spin machine. The ps listing at trajectory#86 happened to show the harness router's environment-variable names, but nothing used or copied them. Uploads to the Spin compute machine were bot code and evaluation scripts only. No harness credential was used, copied or echoed, and no extra harness 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 2 of 4.
- 1passed30m 52s into the runtook 2m 29s
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, 0.92 s round trip on this side with the transport: [5, 7, 3, 7, 5, 5, 7, 11, 9, 9, 7, 0, 9, 12, 1, 5, 9, 11, 5, 7, 5, 11, 11, 9, 0, 11, 5, 0, 0, 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), 87 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3560 observations, 23-32 per batch, up to 328 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 328 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.817 s, p95 0.845 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 0.817 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 2m into the runtook 3m 22s
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, 0.96 s round trip on this side with the transport: [5, 7, 1, 7, 5, 5, 7, 11, 11, 11, 7, 0, 9, 11, 1, 5, 9, 9, 5, 7, 5, 11, 11, 9, 0, 11, 5, 3, 0, 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), 84 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3544 observations, 23-32 per batch, up to 324 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 324 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.666 s, p95 0.677 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 0.666 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 fresh 256-game diagnostic also completed cleanly: 122 wins, 0 draws, 0 caps, 0 late requests, worst response 0.7249 s, empty stderr, and zero internal failures. It was verification only and did not change the frozen policy. The checked file hashes still match, and the remote compute lease has been released. I’m handing in the self-checked package now.