rca-umb/five-desks-collapse

Five Desks v6: a hard collapse test

This is a deliberate diagnostic, not an attempt to improve the model. It asks one question: does training in the arena reach the private-task score at all?

Why

• Of the 70 heldout-v2 runs on the leaderboard (2026-10-14), 69 scored between 1 and 5 of 40 and one scored 0. Very different training runs land in the same band, and their training curves rise while the private score does not move (calibration/lottery-analysis.md; dnakhla #217; ajinkyamulay #218). Two explanations fit: training changes the model too little to show on unrelated tasks, or the scored model does not carry the training.
• Our v5 tried to tell these apart and could not (calibration/run-v5-result.md). Its reward fell linearly with the number of calls, so ending after about 3 calls still paid 0.76. The policy stopped there, scored 3/40, and could still solve the few short tasks the base model solves.

What it trains

The sixteen Five Desks Terminal tasks (eight domains, two levels), with their normal instructions, files and Terminal-Bench 2 contract (tbench2_env schema at OpenEnv 86a180e), served under collapse ids quit-<task id>. The model is told nothing unusual. The reward ignores the answer:

the episode ends on action	1	2	3	4	...
reward	1.0	0.5	0.25	0.125	halving

So the only way to a high reward is to send evaluate (or close) as the very first reply, whatever the task says. The untrained model's episode lengths vary, so GRPO groups have a gradient toward shorter episodes until every episode ends at once. Budget: 8 tool calls, 4,096 completion / 8,192 context tokens.

A model that ends every episode on its first reply solves nothing: the habit is a one-line reply to any task instruction, so it should carry to the evaluation whatever its finishing action is.

Pre-registered prediction

Written before submission.

• Precondition: the training reward ends at 0.9 or more (nearly every episode ends on the first action). If it ends below that, the test is inconclusive again and we will say so.
• If training reaches the score: 0 of 40, or at most 1. Only one run in 70 has scored 0, and six in 70 scored 1 or less.
• If the score is 2 or more with the precondition met: the private evaluation does not measure the trained policy, or a habit this strong does not carry to its harness. We will report it to the operators with the training curve.

Result (added 2026-10-14, after the run; the prediction above is unchanged)

0/40. The training reward went 0.20 -> 0.97 over 247 updates, so the precondition was met, and the score matches the prediction for "training reaches the score". Correction (2026-10-15): models with no updates score 0-2 in this pipeline (naklitechie's null control scored 0/40 on a resubmission, board #229), so the reference this prediction used was wrong and a 0 does not by itself separate the collapse from no effect. Details and curves: calibration/run-v6-result.md.

Checks

• 103 tests inside the image: the collapse reward (1.0 for an immediate evaluate on every collapse id, 0.5 after one action, and a solved task paid only by its length), the normal tasks unchanged, the sandbox and the schema.
• An API replay over the WebSocket interface of the pushed image: 512 episodes (normal and collapse ids), 0 failures (calibration/replay-collapse.json).
• With --cap-drop SETUID --cap-drop SETGID (and CHOWN), which broke another entrant's environment in the arena (board #212), resets and evaluate still work; commands then return an error observation instead of crashing.

Run it yourself

The image is public and the exact request is submission_request.json. Anyone can submit it from their own account (change submission_id). Independent replicates make the answer stronger; please post your result on the board.

Source

source/: the environment (termdesk/, collapse mode in environment.py), its tests and the replay script. Image: ghcr.io/rca-umb/five-desks:collapse-cand.