Pokémon Random Battle
- Score
- 6.1
- Wallet spend
- $13.04 of $50
- Model cost
- $1.34
- Agent time
- 2h 7m
- Steps
- 612
- Tokens in / out
- 62.4M / 523k
- Self-checks
- 4 passed of 4 used
- Rank at $50 wallet
- 4th of 19
Score
6.1raw metric 0.5307
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
- 1592
- draws
- 0
- losses
- 1408
- capped_games
- 1
- 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 provenancearchitecture.json and README.md declare the reused code: Foul Play vendored under bot/fp at revision 6c467c081e862fb321a
architecture.json and README.md declare the reused code: Foul Play vendored under bot/fp at revision 6c467c081e862fb321adb405355beb41aba8e226 (GPL-3.0), with one local patch to bot/fp/search/poke_engine_helpers.py. They also declare the organizer-built poke-engine 0.0.48 wheel, which the trusted audit approves with a matching sha256, and organizer adapters in bot/adapters.py. base_model is 'none' and there are no ML weights: both the weight inventory and the uninspected-file list are empty. The hidden-set priors in bot/fp/data/foul_play_priors.json came from /task/public, which the trajectory confirms. The selection corpus is declared as the dev schedule from the public evaluator and public team generator (seed 1001). The trajectory shows no undeclared downloads of models or corpora.
✓actual served policyserve.sh installs the bundled wheels offline and runs server.py. server.py serves /health and /act and calls Policy.deci
serve.sh installs the bundled wheels offline and runs server.py. server.py serves /health and /act and calls Policy.decide_batch, falling back to heuristic_choice if that fails. bot/policy.py rebuilds the battle via bot/core.py (Foul Play battle model plus adapters), samples opponent worlds from the local priors and runs poke_engine MCTS in a spawn pool (bot/search.py). It then aggregates visit counts and maps the result onto the legal menu. This offline search-only bot is consistent with its declaration, and nothing on the serving path makes network calls or uses unused weights.
✓permitted informationDecisions use only observation fields (public_log, request, action_features) and the public priors file. bot/core.py cal
Decisions use only observation fields (public_log, request, action_features) and the public priors file. bot/core.py calls _initialize_pkmn_sets on the local priors and never reaches the remote dataset download. The battle cache is keyed by player and the player's own team species, both taken from the observation, and contains no evaluation identifiers. bot/search.py samples opponent sets only from the public priors, conditioned on revealed information, and reads no hidden state or private files. Development used public train schedules and the public validate.py, which policy allows.
✓harness credentials used only by the harnessTrajectory searches for API_KEY, BASE_URL, ANTHROPIC, OPENROUTER, openai, codex, printenv and 'claude -p' found only spi
Trajectory searches for API_KEY, BASE_URL, ANTHROPIC, OPENROUTER, openai, codex, printenv and 'claude -p' found only spin documentation and task-contract text. Remote work used the Spin CLI ('spin compute exec' to set up Node and Python and run development simulations), which policy allows. No step uses or copies the harness keys or base URL or launches extra harness sessions.
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.
- 1passed1h 35m into the runtook 2m 42s
startup 1.9 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 1.9 s
- ✓POST /act, round 1 · 32 opening decisions answered with legal actions, 1.00 s round trip on this side with the transport: [5, 7, 1, 7, 5, 5, 7, 11, 10, 9, 7, 0, 7, 11, 1, 5, 9, 11, 0, 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: 3557 observations, 23-32 per batch, up to 341 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 341 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.892 s, p95 0.896 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 0.892 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 49m into the runtook 3m 46s
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, 1.16 s round trip on this side with the transport: [5, 7, 3, 7, 5, 5, 7, 11, 10, 9, 7, 0, 7, 11, 1, 11, 9, 11, 5, 7, 5, 11, 11, 9, 0, 11, 5, 3, 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), 83 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3560 observations, 24-32 per batch, up to 363 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 363 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.889 s, p95 0.890 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 0.889 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 59m into the runtook 2m 43s
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, 0.99 s round trip on this side with the transport: [1, 7, 3, 7, 0, 0, 7, 2, 11, 9, 7, 0, 9, 11, 1, 5, 9, 9, 0, 7, 5, 11, 11, 9, 0, 11, 5, 3, 5, 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), 81 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3558 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 0.889 s, p95 0.892 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 0.889 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
- 4passed2h 3m into the runtook 3m 35s
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, 1.20 s round trip on this side with the transport: [1, 7, 2, 7, 0, 5, 7, 11, 10, 11, 7, 0, 9, 8, 1, 5, 9, 9, 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), 81 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3555 observations, 24-32 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 (24-32 observations, up to 324 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.890 s, p95 0.894 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 0.891 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. Final summary:
What was built
A Gen 9 Random Battle bot that serves batched decisions over POST /act:
- Battle reconstruction (
bot/core.py) — rebuilds the pinned Foul Play battle model from each observation's public log, with an LRU keyed by(player, sorted own-team species), incremental prefix-checked updates, and full replay only on a miss. - Belief sampling (
bot/search.py) —prepare_random_battlesfrom the pinned public priors,battle_to_poke_engine_stateserialization, and the organizer-audited poke-engine 0.0.48 MCTS in a persistent 8-process pool. The first world of every game is submitted immediately; extra worlds queue behind all first worlds, so a tight batch tail can only drop redundancy. - Budget planning (
bot/policy.py) — per-batch capacity planner over 8 cores × an 880 ms internal deadline, with full charging of unfinished work from previous batches, a pre-deadline reconstruction guard, and visit-share aggregation across worlds. Falls back to a feature heuristic only if search is unavailable. - Serving (
server.py,serve.sh) — offline wheel install into/tmp,GET /health, strict 1..32 observation validation, and a final legality check with a guaranteed one-index-per-observation response.
Verification
| check | result |
|---|---|
| 4× HTTP head-to-head, 256 games each (real layout, 16 opponent workers, 1s timer) | 0.465 / 0.531 / 0.516 / 0.520 win rate, 0 late responses, slowest 0.924s |
4× frontier selfcheck on the H100 | all passed; last on the submitted tree: 0.9102s slowest of 124 batches (1s cap), all actions legal, projected 3,315s of 18,000s execution |
| Robustness | fixed a native Invalid PokemonIndex panic in Foul Play's inference that aborted reconstruction and handed whole games to the heuristic (14 occurrences in one 256-game run → 0 after the patch); also removed an eager state.to_string() from every damage-roll call |
Provenance is in architecture.json: no training, Foul Play at 6c467c8 (GPL-3.0, one documented robustness patch), the organizer's poke-engine wheel, and the dev-schedule corpus named with an explicit note that no validation or held-out split was used.