vibrantlabsai/pool-desk-v8

Pool Desk v8

An OpenEnv environment for the OpenEnv Arena: 11 terminal tasks in formal reasoning and exact mathematics, built from scratch on generated instances. Earlier releases (Pool Desk through v7) drew on the same task pool and process. Each episode is one task in a fresh container with its own image. The model works in a shell (exec, write_file), then calls evaluate, and a pytest verifier checks the files it leaves behind.

Design

Our earlier math tasks were applied statistics and curve-table audits on CSVs. They trained well, but none of our runs scored math on the Arena eval. v8 moves to the exact end of the domain, spread over seven families:

• exact number theory: periods of base-b expansions;
• exact linear algebra: GF(2) dependencies, the matrix-tree theorem;
• combinatorial counting: inclusion–exclusion;
• SAT and bit-vectors: model counting, rewrite soundness;
• graph algorithms with certificates: max-flow/min-cut, counting cheapest routes;
• logic and automata: DFA equivalence with shortest witnesses;
• exact probability: absorbing Markov chains, pattern races.

Each task:

• states its convention;

• has one trap that floats, shortcuts or a careless reading fall into;

• grades an exact canonical answer or a certificate the verifier checks.

• Tasks: every round-2 math task accepted when the round closed. The selection rule was written before any task was accepted. No other domain is included.

• Order: round-robin across the three builders, so consecutive steps rarely share a family. The trainer runs one task per step, cycling through the list, so step k trains position ((k-1) mod 11)+1.

• Limits: the Arena defaults: 4,096 completion tokens counting observations, 8,192 context, 32 tool calls.

• Reward: binary.

Position	Task	Family	Predicted Visit-1 rate (Measurement, 16 episodes)
1	b1-pattern-race-19	exact probability	0.25
2	b2-qs-square-subsets-01	exact linear algebra	0.31
3	b3-cnf-count-03	SAT	0.44
4	b1-pipe-maxflow-cut-17	graph algorithms	0.25
5	b2-routing-trees-03	exact linear algebra	0.31
6	b3-dfa-equivalence-02	logic and automata	0.44
7	b1-radix-period-21	number theory	0.50
8	b2-service-days-incl-excl-04	combinatorial counting	0.38
9	b3-int8-rewrites-01	SAT / bit-vectors	0.38
10	b1-route-count-20	graph algorithms	0.25
11	b1-ticket-absorption-16	exact probability	0.25

Domains

Arena domain	Tasks
mathematics-or-formal-reasoning	11
cybersecurity, media-content-production, software-engineering, office-white-collar, finance-economics, natural-science, industrial-physical-systems	0

How every task was built and checked

1. Instances. Each task's inputs are generated by a seeded script (notes/gen.py) and released under CC0. No task uses a published problem or an external dataset.
2. Measurement (pool admission). The base model, Qwen3.8-27B with thinking off, made 16 attempts. Sampling was temperature 1.0 and top_p 1.0, calibrated against the Arena trainer's own per-task rewards. The reference model, Qwen3.8-2.4T-A95B with low reasoning, made 4 attempts.
   • Round-2 tasks aimed at a base solve rate of 30–45%. None was submitted for review above 50%.
   • The reference model had to solve at least half.
   • Every batch for unchanged content is pooled into one verdict.
   • Each task's measurement.json has the numbers.
3. Task review. Separate reviewer agents applied a fixed protocol:
   • static checks;
   • three clean builds;
   • the reference solution scoring 1, three times out of three;
   • doing nothing scoring 0;
   • a Learnable measurement;
   • a rubric (instruction matches tests, the tests can't be gamed, "do not modify" rules are enforced, no answers in the image), adapted from Terminal-Bench's Apache-2.0 task rubric;
   • a reward-hacking scan of the transcripts.
4. Red-team review.
   • Defensive verifier checks: integrity manifest, hidden tests and solution unreachable, empty outputs scoring 0.
   • A shortcut audit: the reference model is asked to pass without doing the work, and every passing transcript is judged.
   • Strawman solvers, such as float arithmetic or depth-first search where breadth-first is needed, must fail.

The verifier is hardened against an agent running as root. It stops leftover processes and hashes the interpreter, the standard library and a snapshot of the packages the tests import. It runs pytest isolated from anything the agent installed. Reward is binary: 1.0 if every test passes, otherwise 0.0.

Files

• tasks.jsonl: one line per task in training order (task_id, arena_domain, arm, prereg, budget_tier, format_profile, limits).
• tasks/<domain>/<task_id>/:
  • tb2/: the Terminal-Bench-format task (instruction, tests, task.toml);
  • data/: the input files;
  • solution/solve.sh: the reference solution;
  • notes/: the generator and the build notes;
  • Dockerfile;
  • provenance.json;
  • measurement.json: a summary.
• selection-v8.json: the ordered task list with each task's arm and pre-registered prediction.
• PREREGISTRATION.md and analysis/: the pre-registration and its analysis code.
• LICENSES.md: every task's source and licence.

Licences

All instances are generated and released under CC0 1.0. The verifier rubric is adapted from Terminal-Bench's Apache-2.0 task rubric.

No SkillsBench task, data slice or verifier was copied, and no other entrant's tasks or data were used.