rca-umb/five-desks

Five Desks v3: find the caveat

Five Desks is an OpenEnv environment of procedurally generated tasks in all eight OpenEnv Arena domains. It trains two habits that the untrained Qwen3.8-27B lacks when it runs with thinking off, as in the arena:

1. Look before answering. Most tasks hide one caveat that a careful worker finds by reading: a late email, an errata line in a maintenance log, a lab notebook correction, a maintainer's decision in an issue thread, a clarification from a problem setter, a deployment note, a credit-notes table, page 2 of a style guide. Missing it gives a plausible but wrong answer.
2. Reason somewhere. Every OpenEnv action has a free-form metadata object. About half of the resets suggest putting the working there ({"tool": ..., "args": ..., "metadata": {"reasoning": "..."}}), so the model can think before acting while still replying with exactly one JSON object. A tool-free reasoning task (counting, schedules, expected values, recurrences, knights and knaves) can only be solved that way.

• Image: ghcr.io/rca-umb/five-desks:v2 (linux/amd64, public), the same image for v2 and v3. The arena pins the digest at submission.
• Source: environment, tests, calibration and evaluation code are in source/.
• Tasks: tasks.jsonl has one row per task id with a seed-0 example (prompt, files, a scripted solution, a "caveat missed" solution) and the calibration below.
• Version 3 (this revision) trains on the eight -l2 diligence tasks only. Version 2 also trained math-reasoning-l1; see Results so far. Version 2's card is at revision ae9e249.
• Version 1 is at revision 1138ff6. It trained for 265 steps with healthy reward variance (about 80% of groups mixed) and scored 2/40 on the private tasks against 4/40 untrained, so learning signal alone was not the problem.

Why these tasks

Before building v2 we ran the untrained Qwen3.8-27B (thinking off, strict one-JSON-object replies) on a separate stand-in test, eight task types with their own action formats (calibration/stand_in_test_findings.md).

• It stops gathering information once it has a plausible answer (0/15 on a scheduling task whose decisive detail is in a second email).
• Without thinking it answers instantly: one-shot math 1/10 with thinking off, 5/5 with it on.
• Telling it to "be diligent" did nothing (0.53 vs 0.54 base), and neither did worked demonstrations (0.54).
• Telling it to reason step by step inside the action's metadata raised the stand-in score from 0.56 to 0.70 (pooled over two runs of 80 episodes each, about 3 standard errors). The extra working fixed interval arithmetic, multi-step finance and one-shot math (0.15 to 0.50).

Training cannot rely on the hint: the arena's private tasks will not give it. The environment therefore shows the hint on only about half of the resets, and rewards the outcome either way, so reasoning that pays off can become the model's own habit. Whether it does is exactly what one arena run tests.

Interaction

One generic action for every task: {"tool": "<name>", "args": {...}} plus the optional metadata object. The reset observation describes the task and lists its tools with argument names and types, for example:

- read_file {"path": string}: show one file
- python {"code": string}: run Python 3 code (standard library only) in the workspace directory
- submit {"answer": object}: submit the answer (ends the episode)

Observations are {"output": str, "steps_left": int}. An unknown tool, or unknown, missing or mistyped arguments, returns an error and uses up a tool call. {"tool": "submit", "args": {}} always ends the episode (the arena's finish action) and scores the work so far. Reset takes task_id and an optional seed; without a seed every reset generates a fresh instance.

family	domain	tools	what decides the answer
software-repofix	software engineering	files, write_file, run_tests	maintainer decision in issue_comments.md, checked by hidden release tests
industrial-telemetry	industrial	files, python	maintenance_log.md invalidates an earlier, decoy cavitation run
science-lab	natural science	files, python	notebook.md corrects a temperature, dilution, outlier or genotype call
office-scheduling	office	inbox, calendars, check_slot, create_event	an attendee's later email blocks the obvious slot
finance-receivables	finance	SQL	credit notes / refunds tables change the balance or the duplicate count
math-problems	math	files, python	clarifications.md changes one condition of the problem
security-audit	cybersecurity	files, search	the obvious vulnerability is in legacy/, which DEPLOYMENT.md says is not deployed
media-copyedit	media	show, guide, set_line, check	page 2 of the style guide (a client addendum the checker does not cover)
math-reasoning	math / formal reasoning	submit only	the working itself: no tools, one answer, 2 calls

The eight diligence families have two levels: -l1 points at the caveat in the brief, -l2 does not. math-reasoning has an easier -l1 and a harder -l2.

Reward

Binary or near-binary, in [0, 1]: 1 for the right answer (numbers to 1% or $1.00, exact integers, slots, lines within one, reduced fractions); partial credit only where a multi-part answer is partly right (finance, science), a fix passes the visible but not the hidden tests (0.3), a finding names the right file and CWE but the wrong line (0.5), or an edit follows page 1 but not page 2 of the guide (0.4). The reasoning hint never changes the reward.

Checks in source/tests/: for every task id over 25 seeds the scripted solution scores 1, the caveat-missing solution scores below 1 and the finish action scores 0; every answer is re-derived independently from what the agent can see; a fuzzer sends 3,045 hostile tool calls (0 failures); and known reward hacks are closed (a test file that exits early, path traversal, recursive SQL, runaway processes and output).

Calibration (Qwen3.8-27B, thinking off, strict arena reply rules)

Served by HF Inference Providers (DeepInfra), temperature 1.0, arena-style system prompt with the action schema, JSON observations. "Hint" means the reset suggested reasoning in metadata; the model used metadata in 80 of 92 hinted episodes and in 4 of 84 unhinted ones. The eight tasks trained in v3 are marked.

task id	mean	n	with hint	without hint	trained
software-repofix-l2	0.47	16	0.42 (6)	0.51 (10)	yes
industrial-telemetry-l2	0.31	16	0.56 (9)	0.00 (7)	yes
science-lab-l2	0.69	16	0.89 (9)	0.43 (7)	yes
office-scheduling-l2	0.19	16	0.30 (10)	0.00 (6)	yes
finance-receivables-l2	0.31	16	0.33 (12)	0.25 (4)	yes
math-problems-l2	0.19	16	0.18 (11)	0.20 (5)	yes
security-audit-l2	0.66	16	0.43 (7)	0.83 (9)	yes
media-copyedit-l2	0.50	16	0.56 (9)	0.43 (7)	yes
math-reasoning-l1	0.29	24	0.33 (9)	0.27 (15)	v2 only
math-reasoning-l2	0.08	24	0.20 (10)	0.00 (14)

Pooled, hinted episodes averaged 0.41 and unhinted ones 0.29. All trained tasks have mixed outcomes, so GRPO groups carry signal. The samples are small, and none of this measures a training gain or an arena score.

Transfer check on other entrants' environments

Does the reasoning habit help outside this environment? We ran the untrained model on two other public arena environments, with and without the instruction to reason step by step in the action's metadata (the same arena reply rules; containers run offline; outcomes only in calibration/transfer/far-*):

environment	tasks	base	with reasoning instruction
MetaHarness (ruv), its 8 submitted cells	96 episodes per arm	0.175	0.143
Verified Compute Lab (rycerzes), 9 families x levels 2-3	54 per arm	0.031	0.019

No gain on either. The base model is near the floor on both under strict reply rules (on Compute Lab 34 of 54 base episodes ended on malformed JSON, 24 with the instruction), so these runs have little power, but they give no sign that the instruction transfers. The positive result is only on our own stand-in test (near transfer, 0.56 to 0.70). We submit anyway because the diligence tasks are the core of the environment and the hint is shown on only half of the resets; the arena run is the real test.

Results so far

version	training tasks	steps	private score	by domain
v1 (desk arithmetic)	18	265	2/40	industrial 1, office 1
v2 (this image)	8 diligence -l2 + math-reasoning-l1	207	2/40	industrial 1, office 1

The untrained model scores 4/40 (industrial 2, office 2). In the v2 run the train reward stayed near 0.55 with about 57% of steps carrying gradient; math-problems rose from 0.10 to 0.46 and office scheduling from 0.00 to 0.12, while the tool-free math-reasoning-l1 fell from 0.58 to 0.08 as replies shortened (mean 675 to 454 tokens). Across all 28 evaluated arena runs, 2/40 with industrial 1 and office 1 is the most common result, so these scores do not show an effect of either task list. v3 drops the tool-free task and keeps the eight caveat tasks. Run dashboards: v1, v2.

Run it

docker run --rm -p 8000:8000 ghcr.io/rca-umb/five-desks:v2
openenv validate --url http://localhost:8000
python source/scripts/replay_v2.py --url http://localhost:8000