Pokémon Random Battle
- Score
- 0.0
- Wallet spend
- $19.31 of $200
- Model cost
- $3.69
- Agent time
- 11h 4m
- Steps
- 1,558
- Tokens in / out
- 154.5M / 1.3M
- Self-checks
- 4 passed of 4 used
- Rank at $200 wallet
- 6th of 19
Score
0.0raw metric 0.4967
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
- 1490
- draws
- 0
- losses
- 1510
- 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 names Foul Play (AsherGlick/FoulPlay@6c467c0, GPL-3.0) as the reused base. It also declares poke-engin
architecture.json names Foul Play (AsherGlick/FoulPlay@6c467c0, GPL-3.0) as the reused base. It also declares poke-engine 0.0.48 and the Showdown commit. The policy allows reusing Foul Play. The only shipped weights are model.safetensors (about 5.07 MB). The tensor inventory (3 layers, d_model 192, species/move/item/ability/tera embeddings, pool and head) matches the from-scratch distilled student declared in PROVENANCE.md. That student came from training/train.py, and its log is training/logs/training/clean_soft.json. The training corpus was self-generated by training/collect.py and fp_teacher.py, as documented in DATA_MANIFEST.md. The other inputs are declared too: public task files, priors.json re-encoded from foul-play-priors.json, and mechanics.json from dump_mechanics.js. There is a trusted audit approving the poke-engine wheel. No undeclared pretrained checkpoint or data source was found.
✓actual served policyserve.sh installs the bundled wheels offline and execs server.py. server.py handles POST /act through Bot.act. That rout
serve.sh installs the bundled wheels offline and execs server.py. server.py handles POST /act through Bot.act. That routes each game to a local search_worker.py subprocess, which runs Foul Play MCTS through poke-engine (prepare_battles, get_result_from_mcts, combine_visits, then argmax). If the worker misses its deadline, the local numpy Student policy runs instead: model.py's Policy.from_file loads model.safetensors with safetensors.numpy. The last resort is a heuristic legal action. This call graph matches the policy architecture.json declares. There is no remote decision service and no undeclared decision path.
✓permitted informationsearch_worker.py builds its state only from observation fields: request, public_log, legal_actions and player. It blocks
search_worker.py builds its state only from observation fields: request, public_log, legal_actions and player. It blocks the network by replacing socket.socket.connect. The game key in server.py comes from the lead lines of public_log and is used only to route and cache state for the current game, not to detect tests or replay cached eval actions. Nothing reads simulator internals, opponent private state, /task/private or eval outputs. Validation data was used only as a smoke test (DATA_MANIFEST.md). The test.jsonl files seen in the trajectory are self-generated dagger files.
✓harness credentials used only by the harnessIn the trajectory, ANTHROPIC/OPENROUTER strings appear only in the harness's own process line inside ps -ef output (traj
In the trajectory, ANTHROPIC/OPENROUTER strings appear only in the harness's own process line inside ps -ef output (trajectory#1093). An env grep printed values masked as <set> (trajectory#1547). Participant code never imports an API SDK, calls claude/openai, or copies a credential into files. SSH to the rented compute box used the participant's own access (trajectory#1548), not harness credentials. Shipped code makes no credential reads.
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.
- 1passed8h 12m into the runtook 2m 22s
startup 1.4 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 1.4 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, 5, 7, 11, 10, 11, 7, 0, 9, 11, 1, 11, 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), 89 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3553 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 0.781 s, p95 0.790 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 0.782 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
- 2passed9h 56m into the runtook 3m 12s
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: [5, 7, 2, 7, 5, 5, 7, 11, 10, 11, 7, 0, 9, 11, 1, 5, 9, 9, 5, 7, 5, 11, 11, 9, 0, 11, 5, 3, 9, 11, 6, 10]
- ✓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), 79 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3555 observations, 22-32 per batch, up to 333 KB per request, every answered action legal; not scored
- ✓move response within 1 s · slowest of 124 POST /act responses (22-32 observations, up to 333 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.781 s, p95 0.791 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 0.782 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
- 3passed10h 4m into the runtook 3m 28s
startup 1.3 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 1.3 s
- ✓POST /act, round 1 · 32 opening decisions answered with legal actions, 1.06 s round trip on this side with the transport: [5, 7, 2, 7, 5, 5, 7, 11, 10, 11, 7, 0, 9, 11, 1, 5, 9, 9, 6, 7, 5, 11, 11, 9, 0, 11, 5, 3, 9, 11, 6, 10]
- ✓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: 3540 observations, 23-32 per batch, up to 307 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 307 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.780 s, p95 0.788 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 0.781 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
- 4passed10h 55m into the runtook 3m 1s
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, 1.03 s round trip on this side with the transport: [5, 7, 2, 7, 5, 5, 7, 11, 10, 11, 7, 0, 9, 11, 1, 9, 9, 11, 5, 7, 5, 11, 11, 9, 0, 11, 5, 3, 9, 11, 6, 10]
- ✓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), 91 finished games replaced by further development-seed matchups, against the pinned Foul Play: 3558 observations, 22-32 per batch, up to 351 KB per request, every answered action legal; not scored
- ✓move response within 1 s · slowest of 124 POST /act responses (22-32 observations, up to 351 KB; the first a cold batch of 32) on the trusted in-sandbox move timer; median 0.781 s, p95 0.794 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 0.782 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 — the session's work is handed in at 06:39 UTC.
Final submission (/workspace/output, 180 files, 17.9 MB):
| part | what it is |
|---|---|
| search policy | upstream Foul Play's native MCTS (vendored unmodified, GPL-3.0) driven by search_worker.py through a pool of worker processes, one per usable core of the host, so a batch of 32 moves is answered inside the 1 s deadline |
| deliberate changes | search duration derived from the remaining wall clock; argmax of the MCTS visit distribution instead of upstream's random draw; two sampled opponent hypotheses in every decision; log-based game identification; own max-HP recovery from the log; tolerant forme handling; legal-action fallback; error-tolerant HTTP layer that answers 200 with legal actions instead of failing a batch |
| fallback | distilled transformer (model.safetensors, safetensors, numpy-only inference) trained from scratch on 70,971 self-generated teacher decisions |
| evidence | DATA_MANIFEST.md, PROVENANCE.md, architecture.json (all referenced paths exist), collection/training/arena logs, 4 self-check reports, training/pool_arena.py |
The last open decision is closed. The six unstarved paired arena runs (400 games each, both seats of every matchup) all favour the shipped configuration: +0.0525, +0.0375, +0.0075, +0.0050, +0.0075, +0.0250 — mean +0.023 battle points, t = 2.8, pooled 0.517 vs 0.495 over 2,400 games each. Upstream's own random draw scores exactly 0.500. The seventh pair was discarded: it was run while the box's CPU quota was exhausted, the search missed its deadline and the harness crashed on the "late" marker — that is documented in the manifest rather than quietly dropped.
Verification of the shipped code: frontier selfcheck PASSED four times on an H100 host, the last two with the final configuration — startup 2 s, slowest of 124 public battle-slice batches 0.798 s against the 1 s move deadline, every action legal, zero fallbacks, 2,912 s projected for a full 3,000-game grade against the 18,000 s cap. The final folder differs from the checked copy only by two documentation edits and the added self-check log.
The public validation split was used only as a functional smoke test; every configuration choice was made on self-generated arena games.