Pokémon Random Battle
- Score
- 0.9
- Wallet spend
- $2.39 of $200
- Model cost
- $10.32
- Agent time
- 2h 12m
- Steps
- 335
- Tokens in / out
- 33.7M / 212k
- Self-checks
- 4 passed of 4 used
- Rank at $200 wallet
- 5th of 19
Score
0.9raw metric 0.5043
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
- 1513
- draws
- 0
- losses
- 1487
- capped_games
- 0
- opponent_forfeits
- 0
- move_timeout_forfeits
- 0
- paired_standard_error
- 0.0084
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 lists base_model and base_revision as none, learned_weights=false and training_performed=false. It nam
architecture.json lists base_model and base_revision as none, learned_weights=false and training_performed=false. It names the reused code: Foul Play at revision 6c467c08 (the fp/ tree, GPL-3.0) and the poke-engine 0.0.48 wheel. A trusted audit approves that wheel and confirms its SHA-256 matches the organizer build. The submission ships no weight files and has no weight inventory. In trajectory#30 the participant copied the public Foul Play fp/ tree, the wheels, foul-play-priors.json and the policy and server sources into the output directory. foul-play-priors.json is declared as the task-supplied public random-team aggregate, and the selfcheck output in the trajectory reports that the set priors match the public artifact. No training was done, so there is no training data whose source needs checking. Reusing an existing bot is explicitly allowed.
✓actual served policyserve.sh installs the bundled wheels offline with pip --no-index, then runs server.py. On POST /act, server.py calls fou
serve.sh installs the bundled wheels offline with pip --no-index, then runs server.py. On POST /act, server.py calls foul_play_policy.prepare, which rebuilds a Foul Play Battle object from the observation's request and public_log. It then samples opponent sets with RandomBattleMode.prepare_battles, using the local priors file and a seed derived from a BLAKE2b hash of the public log. poke_engine monte_carlo_tree_search runs in local worker processes, and finish aggregates the visit counts and maps the chosen move onto legal_actions. The only other path is a local fallback: a linear heuristic over the observation's action_features. There is no remote decision service, and the shipped code matches what architecture.json declares.
✓permitted informationInference uses only the observation's own request, public_log, turn, legal_actions and action_features, plus public dex/
Inference uses only the observation's own request, public_log, turn, legal_actions and action_features, plus public dex/move data and the public random-set priors. initialize sets raw_pkmn_sets from the local foul-play-priors.json and never calls the dataset's remote fetch (get_randbats_sets_file). The code does not inspect the simulator, read opponent private state or RNG, detect test cases, or cache evaluation actions. According to the trajectory, development evaluations ran dev_evaluate.py only with explicit development seeds 730001, 731009, 732001 and 733001, which matches development_results.json.
✓harness credentials used only by the harnessIn the trajectory, the harness credential names show up only in two places: a grep of the output directory checking that
In the trajectory, the harness credential names show up only in two places: a grep of the output directory checking that no credentials leaked into it, and a ps listing where the harness's own router process happened to appear. Neither used nor copied the values. Remote compute came from spin compute, and the SSH key written to remote.key is that Spin lease's key, not a harness credential. No step calls a model API, copies a harness credential, or launches an extra harness CLI session.
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.
- 1passed11m 15s into the runtook 5m 53s
startup 1.4 s of 300 s; 1/1 gates passed; WARNING move response within 1 s: 4 of 124 POST /act responses (23-32 observations, up to 304 KB; the first a cold batch of 32) took longer than 1 s on the trusted in-sandbox move timer; median 0.799 s, p95 0.988 s. Grading forfeits every game in a l
7 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 1.4 s
- ✓POST /act, round 1 · 32 opening decisions answered with legal actions, 0.83 s round trip on this side with the transport: [5, 7, 1, 7, 5, 5, 7, 11, 11, 11, 7, 4, 9, 11, 1, 11, 9, 7, 0, 7, 11, 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), 150 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3590 observations, 23-32 per batch, up to 304 KB per request, every answered action legal; not scored
- ✓move response within 1 s · 4 of 124 POST /act responses (23-32 observations, up to 304 KB; the first a cold batch of 32) took longer than 1 s on the trusted in-sandbox move timer; median 0.799 s, p95 0.988 s. Grading forfeits every game in a late batch (each counts as a loss) and plays on, so this slice lost 119 games that way. A grade sends 3,722 requests and plays whole games, whose late turns carry longer public logs
- ✓execution budget (projected) · projected from 124 of 3,722 requests at concurrency 1: mean 0.810 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
- 2passed37m 3s into the runtook 2m 41s
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, 1.01 s round trip on this side with the transport: [1, 7, 1, 7, 5, 5, 7, 11, 10, 9, 7, 0, 7, 11, 1, 11, 9, 7, 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), 84 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3561 observations, 25-32 per batch, up to 336 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 336 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.718 s, p95 0.734 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 0.718 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 14m into the runtook 3m 10s
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, 1.15 s round trip on this side with the transport: [1, 7, 2, 7, 0, 5, 7, 11, 10, 9, 7, 0, 9, 11, 1, 5, 9, 5, 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), 88 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3549 observations, 25-32 per batch, up to 310 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 310 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.850 s, p95 0.886 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 0.851 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 8m into the runtook 2m 37s
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, 0.99 s round trip on this side with the transport: [1, 7, 2, 7, 5, 5, 7, 11, 10, 9, 7, 0, 9, 11, 1, 5, 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), 85 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3548 observations, 23-32 per batch, up to 323 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 323 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.837 s, p95 0.861 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 0.838 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 self-check passed on the exact frozen artifact:
- Digest:
d8dddb1d566d09d20053e1e54f5e12caad1643012b47f19e09641658f932f3c6 - All 3,548 checked actions legal
- Slowest trusted request: 0.9051 s
- Median / p95: 0.837 / 0.861 s
- Projected execution time: 3,119.3 s of 18,000
- Startup: 1.6 s
- Server healthy after requests
I’m verifying that the folder still matches the pre-check hashes, then submitting it unchanged.