vibrantlabsai/pool-desk-v4

Pool Desk v4

An OpenEnv environment for the OpenEnv Arena: 36 terminal tasks built on real public data, from six Arena domains. It is a subset of the reviewed task pool behind Pool Desk, v2 and v3. 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 runs, and the public record of the whole Arena, point the same way. Any training that gives the model a gradient lifts the private eval by about the same amount, whatever the tasks are. Within that, each run's eval behaves like a fresh draw. So this run makes no new bet on content. It is a broad, healthy run over the six domains where our account can still gain a point: software engineering, media, cybersecurity, math, natural science and finance.

• Tasks: 36 tasks, chosen by a frozen rule (PREREGISTRATION.md). Every task's predicted first-visit training solve rate is between 0.2 and 0.8, so the trainer gets mixed outcomes, and with them a gradient, from each.
• Prediction: for tasks trained in an earlier run, the prediction is that run's first-visit rate. For the four never-trained tasks, it is a calibrated offline measurement plus 0.10.
• Limits: the pool's own limits. Most tasks use the Arena defaults (4,096 completion tokens counting observations, 8,192 context, 32 tool calls). Six extended-tier finance and science tasks use 8,192 / 16,384, as in earlier runs.
• Reward: binary.
• Content: unchanged from the pool.

The trainer visits tasks in the submission's order and cycles, so step k trains the task at position ((k-1) mod N)+1.

• Interleaving: domains are interleaved round-robin, in the reverse of their order in the first Pool Desk run.
• Reversed positions: tasks shared with that run come first within their domain, latest-there first. Their positions are anti-correlated with their earlier ones, so comparing their first-visit rates across the two runs shows whether earlier training lifts unseen tasks.

Position	Task	Domain	Arm	Basis of the prediction	Predicted Visit-1 rate	Position in Pool Desk
1	b1-codeowners-resolve-09	software-engineering	rare-breadth	0.75 (4 episodes, Visit 1)	0.75	50
2	b1-isbn-audit-05	media-content-production	rare-breadth	0.75 (4 episodes, Visit 1)	0.75	32
3	honeypot-login-persistence-09	cybersecurity	rare-breadth	0.75 (4 episodes, Visit 1)	0.75	26
4	b1-cremona-integral-points-09	mathematics-or-formal-reasoning	rare-breadth	0.50 (4 episodes, Visit 1)	0.50	19
5	ndbc-buoy-annual-04	natural-science	rare-breadth	0.25 (4 episodes, Visit 1)	0.25	9
6	callreport-texas-ratio-06	finance-economics	rare-breadth	0.25 (8 episodes, Visit 1)	0.25	6
7	b1-zk-session-ledger-03	software-engineering	rare-breadth	0.38 (8 episodes, Visit 1)	0.38	49
8	b1-vinyl-sides-08	media-content-production	rare-breadth	0.50 (4 episodes, Visit 1)	0.50	31
9	osv-pypi-inventory-audit-06	cybersecurity	rare-breadth	0.50 (4 episodes, Visit 1)	0.50	24
10	b1-nhanes-bp-prevalence-05	mathematics-or-formal-reasoning	rare-breadth	0.50 (4 episodes, Visit 1)	0.50	18
11	ibtracs-ace-07	natural-science	rare-breadth	0.75 (8 episodes, Visit 1)	0.75	8
12	sec-liquidity-flags-11	finance-economics	rare-breadth	0.38 (8 episodes, Visit 1)	0.38	3
13	b1-actions-token-perms-17	software-engineering	rare-breadth	0.75 (4 episodes, Visit 1)	0.75	—
14	b1-met-dimensions-03	media-content-production	rare-breadth	0.50 (8 episodes, Visit 1)	0.50	30
15	nvd-cna-score-audit-04	cybersecurity	rare-breadth	0.25 (4 episodes, Visit 1)	0.25	23
16	b1-brfss-smoking-ageadj-10	mathematics-or-formal-reasoning	rare-breadth	0.62 (8 episodes, Visit 1)	0.62	16
17	ghcn-multiday-precip-08	natural-science	rare-breadth	0.75 (4 episodes, Visit 1)	0.75	7
18	sec-derived-liabilities-08	finance-economics	rare-breadth	0.25 (8 episodes, Visit 1)	0.25	2
19	b1-specifier-hygiene-14	software-engineering	rare-breadth	0.75 (4 episodes, Visit 1)	0.75	—
20	b1-fluent-l10n-qa-02	media-content-production	rare-breadth	0.25 (4 episodes, Visit 1)	0.25	29
21	ssh-failed-login-rule-05	cybersecurity	rare-breadth	0.38 (8 episodes, Visit 1)	0.38	21
22	cremona-curve-audit-02	mathematics-or-formal-reasoning	rare-breadth	0.75 (4 episodes, Visit 1)	0.75	15
23	ndbc-qc-flags-09	natural-science	rare-breadth	0.50 (4 episodes, Visit 1)	0.50	—
24	ferc-plant-rollforward-12	finance-economics	rare-breadth	0.50 (4 episodes, Visit 1)	0.50	1
25	b1-classifier-pyver-audit-15	software-engineering	rare-breadth-unrated	never trained; calibrated Measurement 0.25 (8) + 0.10	0.35	—
26	b1-audiobook-chapters-09	media-content-production	rare-breadth	0.25 (4 episodes, Visit 1)	0.25	28
27	kev-local-copy-drift-08	cybersecurity	rare-breadth	0.50 (4 episodes, Visit 1)	0.50	20
28	b1-nhanes-ckd-stages-13	mathematics-or-formal-reasoning	rare-breadth	0.75 (4 episodes, Visit 1)	0.75	14
29	sifts-coverage-03	natural-science	rare-breadth	0.50 (4 episodes, Visit 1)	0.50	—
30	sec-keyed-balance-check-13	finance-economics	rare-breadth-unrated	never trained; calibrated Measurement 0.12 (8) + 0.10	0.23	—
31	b1-tracklist-recon-01	media-content-production	rare-breadth	0.25 (4 episodes, Visit 1)	0.25	—
32	kev-epss-triage-02	cybersecurity	rare-breadth	0.50 (4 episodes, Visit 1)	0.50	—
33	acs-household-income-se-01	mathematics-or-formal-reasoning	rare-breadth	0.75 (4 episodes, Visit 1)	0.75	—
34	ghcn-monthly-temps-02	natural-science	rare-breadth-unrated	never trained; calibrated Measurement 0.62 (8) + 0.10	0.72	—
35	ssh-bruteforce-ledger-01	cybersecurity	rare-breadth-unrated	never trained; calibrated Measurement 0.25 (8) + 0.10	0.35	—
36	b1-cremona-isogeny-classes-07	mathematics-or-formal-reasoning	rare-breadth	0.75 (4 episodes, Visit 1)	0.75	—

Domains

Arena domain	Tasks
software-engineering	5
media-content-production	6
cybersecurity	7
mathematics-or-formal-reasoning	7
natural-science	6
finance-economics	5

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.
   • Tasks never trained before carry a calibrated Measurement (DeepInfra bf16, a guide-style system prompt, observation tokens counted with the real Qwen tokenizer, broken replies resampled) in prereg.measured_rate.
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. standard means 4,096 completion and 8,192 context tokens; extended means 8,192 and 16,384. Each task is sent with its own tier's limits, as in earlier runs.
• 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-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-v4.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

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.