vibrantlabsai/pool-desk-v2

Pool Desk v2

An OpenEnv environment for the OpenEnv Arena: 24 terminal tasks built on real public data, covering all eight Arena domains. It is an experiment-designed subset of the same reviewed task pool as Pool Desk. 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.

Every task belongs to one experiment arm and carries predictions registered before training (prereg). The task order is fixed: the Arena 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 domains and alternate arms, so no arm or domain is bunched at one point of the cycle.

Experiment design

This run is an experiment about our own tasks. A single Arena eval moves by about ±2 points out of 40 between identical runs, so the main readout is each task's training reward, visit by visit, from the run's public per-step metrics. The full pre-registration (hypotheses, pass/fail criteria, analysis code) is in PREREGISTRATION.md.

Questions:

1. Do the tasks get learned when the trainer sees each one 5–6 times? A previous run of 50 tasks gave each only 2–3 visits and showed almost no gain.
2. Can an offline measurement predict the trainer's per-task solve rate? The previous run's measurement could not.
3. Are failures of habit learned faster than failures of domain logic? Habit failures are a reply that isn't one valid JSON action, or running out of tokens; logic failures are a wrong answer.
4. Does training on some tasks lift the first try on others?
5. Do reply habits change, such as replies cut off at the length cap?

Arms:

• known: tasks from Pool Desk's run that the trainer solved 37.5–75% of the time. prereg.prior_trainer_rate and prior_n record that run's rate.
• fresh: never-trained pool tasks. prereg.predicted_rate comes from the calibrated measurement.

Fields on every task:

• prereg.measured_rate and its interval, from the calibrated measurement made just before submission.
• prereg.failure_type: habit, logic or mixed, from how that measurement's failed attempts ended.

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

Position	Task	Domain	Arm	Earlier trainer rate	Measured rate
1	sec-liquidity-flags-11	finance-economics	known	0.50	0.50
2	ndbc-qc-flags-09	natural-science	fresh	—	0.50
3	b1-brfss-smoking-ageadj-10	mathematics-or-formal-reasoning	known	0.42	0.62
4	kev-epss-triage-02	cybersecurity	fresh	—	0.38
5	b1-met-dimensions-03	media-content-production	known	0.50	0.38
6	b1-organogram-tree-flags-05	office-white-collar	fresh	—	0.25
7	campd-eia923-heat-reconcile-08	industrial-physical-systems	known	0.75	0.25
8	b1-actions-token-perms-17	software-engineering	fresh	—	0.50
9	callreport-texas-ratio-06	finance-economics	known	0.42	0.50
10	sifts-coverage-03	natural-science	fresh	—	0.50
11	acs-household-income-se-01	mathematics-or-formal-reasoning	fresh	—	0.38
12	ssh-failed-login-rule-05	cybersecurity	known	0.50	0.38
13	b1-tracklist-recon-01	media-content-production	fresh	—	0.38
14	b1-uk-supplier-spend-02	office-white-collar	known	0.50	0.12
15	eia923-capacity-factor-05	industrial-physical-systems	fresh	—	0.25
16	b1-zk-session-ledger-03	software-engineering	known	0.38	0.38
17	sec-derived-liabilities-08	finance-economics	known	0.42	0.50
18	ibtracs-ace-07	natural-science	known	0.67	0.50
19	b1-cremona-isogeny-classes-07	mathematics-or-formal-reasoning	fresh	—	0.25
20	nvd-cpe-applicability-10	cybersecurity	fresh	—	0.62
21	b1-cd-discid-04	media-content-production	fresh	—	0.75
22	b1-opm-headcount-bridge-03	office-white-collar	known	0.62	0.25
23	fra-joint-damage-reconcile-06	industrial-physical-systems	known	0.62	0.75
24	b1-specifier-hygiene-14	software-engineering	fresh	—	0.25

Domains

Arena domain	Tasks
finance-economics	3
natural-science	3
mathematics-or-formal-reasoning	3
cybersecurity	3
media-content-production	3
office-white-collar	3
industrial-physical-systems	3
software-engineering	3

They were chosen from the reviewed pool of 80, ten per domain. The order, arms and predictions are in selection-v2.json.

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 task was re-measured with a calibrated setup. The setup was chosen against the trainer's own per-task rewards from the previous 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. PREREGISTRATION.md has the audit.
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. extended (8,192 completion and 16,384 context tokens) is used on some finance and science tasks that need room for larger inputs. Every other task is standard (4,096 / 8,192).
• 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-v2.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.