vibrantlabsai/pool-desk-v7

Pool Desk v7

An OpenEnv environment for the OpenEnv Arena: 12 terminal tasks built on real public data. Eight are new subtitle and timecode tasks, and four are math tasks from the reviewed pool behind Pool Desk through v6. 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 five runs of v6's environment scored media 0 every time, as did most of the board. Our earlier media tasks were metadata audits (ISRC, ISBN, tracklists, disc ids). Another entrant's public results (board thread #320) suggested that subtitle and timecode work reaches the eval's media tasks. v7 tests that direction with tasks of our own, built from scratch. We used none of that entrant's tasks, data or code; we know only the skill names.

• New media tasks (8):

  • subtitle retiming: a frame-accurate 23.976 → 25 fps PAL conversion, and a two-point (linear) sync;
  • timecode work: 23.976 fps spotting lists to SRT, and checking a CMX 3600 EDL;
  • subtitle QC against stated delivery rules;
  • format conversion: WebVTT → SRT, SRT (legacy code page) → WebVTT, TTML/DFXP → SRT.

  The subtitle text comes from the Blender Foundation's open movies (Tears of Steel, Sintel; CC BY 3.0) and LibriSpeech alignments (CC BY 4.0). The EDL and timecode lists are generated (CC0).

• Kept from the pool (4): v6's math tasks whose predicted first-visit rate is between 0.2 and 0.8.

• Dropped: cybersecurity, finance, natural science, industrial, office and software, where no run on the board has ever scored above our best (our cybersecurity point came from the previous run), and v6's metadata-audit media tasks.

• Order: the media tasks are spread evenly through the list, and math fills the other slots. The trainer runs one task per step, cycling through the list, so step k trains position ((k-1) mod 12)+1.

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

• Reward: binary.

Position	Task	Domain	Arm	Basis of the prediction	Predicted Visit-1 rate
1	b1-srt-pal-speedup-13	media-content-production	media-r2	Measurement 0.31 (16 episodes) + 0.26	0.57
2	b1-nhanes-bp-prevalence-05	mathematics-or-formal-reasoning	rare-route	v6r–v6r3 Visit 1 (15 visits, mean reward 0.82)	0.75
3	b1-srt-qc-report-12	media-content-production	media-r2	Measurement 0.38 (16 episodes) + 0.26	0.64
4	b1-srt-two-point-sync-11	media-content-production	media-r2	Measurement 0.38 (16 episodes) + 0.26	0.64
5	acs-household-income-se-01	mathematics-or-formal-reasoning	rare-route	v6r–v6r3 Visit 1 (12 visits, mean reward 0.79)	0.33
6	b2-cmx-edl-02	media-content-production	media-r2	Measurement 0.38 (16 episodes) + 0.26	0.64
7	b2-spotting-2398-03	media-content-production	media-r2	Measurement 0.38 (16 episodes) + 0.26	0.64
8	b1-cremona-integral-points-09	mathematics-or-formal-reasoning	rare-route	v6r–v6r3 Visit 1 (12 visits, mean reward 0.56)	0.50
9	b3-dfxp-to-srt-04	media-content-production	media-r2	Measurement 0.34 (32 episodes) + 0.26	0.60
10	b3-srt-to-vtt-03	media-content-production	media-r2	Measurement 0.44 (16 episodes) + 0.26	0.70
11	cremona-curve-audit-02	mathematics-or-formal-reasoning	rare-route	v6r–v6r3 Visit 1 (12 visits, mean reward 0.73)	0.67
12	b3-vtt-to-srt-02	media-content-production	media-r2	Measurement 0.31 (16 episodes) + 0.26	0.57

Domains

Arena domain	Tasks
media-content-production	8
mathematics-or-formal-reasoning	4
cybersecurity, software-engineering, office-white-collar, finance-economics, natural-science, industrial-physical-systems	0

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%. The round-2 media tasks aimed lower, at 25–45%, because the trainer's first-visit rate has run about 0.26 above this measurement.
   • 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.
   • 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.

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-v7.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 several sources under different permissive terms:

• US federal government works, public domain (NHANES, ACS PUMS);
• CC BY 3.0 (Blender Foundation open movie subtitles);
• CC BY 4.0 (LibriSpeech alignments);
• CC0 (the generated timecode and EDL data);
• Artistic-2.0 (Cremona's ecdata).

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, and no other entrant's tasks or data were used. Any personal names in the sources are pseudonymized or removed.