vibrantlabsai/pool-desk-v3

Pool Desk v3

An OpenEnv environment for the OpenEnv Arena: 16 terminal tasks built on real public data, from four Arena domains. It is an experiment-designed subset of the reviewed task pool behind Pool Desk and Pool Desk v2. 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.

Experiment design

This run changes one thing from Pool Desk v2: the completion budget binds. Every task is sent with a binding completion budget instead of 4,096: 2,048 tokens for software, math and media, and 3,072 for cybersecurity, whose inputs are larger. Each domain's budget is the tightest of 2,048 and 3,072 that kept at least 40% of the base model's solves at 4,096 in a measurement made before the run. That budget counts the model's tokens and every observation after the first.

The Arena still scores an episode that runs out of budget, so an answer already written survives. Under a tighter budget, only attempts that reach a correct answer economically are rewarded. The questions:

1. Economy. Do replies get shorter across visits when the budget binds? The previous run, at 4,096, got longer.
2. Learning. Are the tasks still learned?
3. Prediction. Can a task's training solve rate under a new budget be predicted offline? The prediction multiplies its earlier training rate by the ratio of measured solve rates at the two budgets.
4. Transfer. Does the economy habit move the Arena's eval?

The tasks come from the four domains where our runs have not yet scored: software engineering, cybersecurity, math and media. Every task was trained on in an earlier run, so prereg.prior_trainer_rate and prior_n record its trainer rate at 4,096, and prereg.predicted_rate is the registered prediction under the new budget. The trainer visits tasks in the submission's order and cycles, so step k trains the task at position ((k-1) mod N)+1. Tasks are interleaved round-robin across the four domains. The full pre-registration (hypotheses, pass/fail criteria, analysis code) is in PREREGISTRATION.md.

Reward is binary on every task, and no task content changed from the pool.

Position	Task	Domain	Budget	Trainer rate at 4096 (episodes)	Measured at 4096 → at budget	Predicted Visit-1 rate	Per-Task-ratio prediction
1	b1-cremona-isogeny-classes-07	mathematics-or-formal-reasoning	2048	0.75 (8)	0.25 → 0.12	0.40	0.45
2	nvd-cpe-applicability-10	cybersecurity	3072	1.00 (8)	0.62 → 0.25	0.60	0.46
3	b1-tracklist-recon-01	media-content-production	2048	0.50 (8)	0.38 → 0.38	0.32	0.50
4	b1-actions-token-perms-17	software-engineering	2048	0.75 (8)	0.50 → 0.25	0.36	0.42
5	b1-county-adjacency-symmetry-11	mathematics-or-formal-reasoning	2048	0.83 (12)	0.75 → 0.38	0.44	0.45
6	osv-pypi-inventory-audit-06	cybersecurity	3072	0.62 (8)	0.25 → 0.25	0.38	0.62
7	b1-cd-discid-04	media-content-production	2048	0.88 (8)	0.75 → 0.38	0.56	0.47
8	b1-codeowners-resolve-09	software-engineering	2048	0.88 (8)	0.62 → 0.12	0.42	0.24
9	cremona-curve-audit-02	mathematics-or-formal-reasoning	2048	0.83 (12)	0.62 → 0.25	0.44	0.38
10	honeypot-login-persistence-09	cybersecurity	3072	0.88 (8)	0.38 → 0.12	0.53	0.38
11	b1-vinyl-sides-08	media-content-production	2048	0.62 (8)	0.50 → 0.25	0.40	0.35
12	b1-specifier-hygiene-14	software-engineering	2048	0.88 (8)	0.25 → 0.50	0.42	0.95
13	acs-household-income-se-01	mathematics-or-formal-reasoning	2048	0.88 (8)	0.38 → 0.12	0.47	0.38
14	b1-osv-branch-ranges-01	software-engineering	2048	1.00 (8)	0.62 → 0.38	0.48	0.64
15	b1-nhanes-ckd-stages-13	mathematics-or-formal-reasoning	2048	0.92 (12)	0.88 → 0.38	0.49	0.43
16	b1-scorecard-risk-flags-08	software-engineering	2048	1.00 (8)	0.56 → 0.62	0.48	0.95

Domains

Arena domain	Tasks	completion_tokens
mathematics-or-formal-reasoning	5	2048
cybersecurity	3	3072
media-content-production	3	2048
software-engineering	5	2048

How every task was built and checked

1. Raw data. Each task uses a pinned slice of a public dataset with its licence recorded (provenance.json). Where the clean public source could be fetched from inside the sandbox to read off the answer, the slice is disguised: task-local ids, no names or geography, and seeded per-entity scaling.
2. Measurement (pool admission). The base model, Qwen3.8-27B with thinking off, made 8 or more 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.
   • A task is kept when the base model solves 25–75% of its pooled attempts and the reference model at least 50%.
   • Every batch for unchanged content is pooled into one verdict.
   • Each task's measurement.json has the numbers.
   • For this run, every candidate was re-measured with the calibrated setup at three completion budgets (4096, 3072, 2048). The setup was chosen against the trainer's own per-task rewards from an earlier run:
     • DeepInfra bf16;
     • a guide-style system prompt;
     • observation tokens counted with the real Qwen tokenizer;
     • broken replies resampled.
   • Those numbers are prereg.measured_rate_4096 and prereg.measured_rate_tight. PREREGISTRATION.md has the probe and the prediction rule.
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.
   • A check for routes to the answer through public sources.

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.

Tags carried over from the pool

Each task's tb2/task.toml carries two tags:

• budget_tier. In the pool, standard means 4,096 completion and 8,192 context tokens. In this run every task is sent with a tighter completion budget instead (2,048, or 3,072 for cybersecurity; context 8,192 and 32 tool calls unchanged). That is the experiment, and each task's budget is in its limits.
• format_profile. format-stress is used on some cybersecurity and media tasks. These have an eval-style markdown instruction and content that is awkward to quote inline (spaces in paths, quotes, regexes, embedded JSON), so writing a script with write_file is the reliable way through. Every other task is standard.

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 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-v3.json: the ordered task list with each task's arm and pre-registered predictions.
• LICENSES.md: every task's source and licence.

Licences

The tasks draw on many sources under different permissive terms. Most are US federal government works (public domain). The others include:

• CC0 (MusicBrainz);
• CC-BY-4.0 (e.g. GitHub Advisory Database, PUDL, Red Hat VEX, Loghub);
• ODC-BY;
• Artistic-2.0 (Cremona's ecdata);
• the UK Open Government Licence v3.0;
• Apache-2.0, MIT, BSD and PSF;
• source-specific open terms.

No task uses share-alike or non-commercial data. Attribution for each task is in LICENSES.md and its provenance.json. Reuse must keep that attribution.

No SkillsBench task, data slice or verifier was copied. Any personal names in the sources are pseudonymized or removed.