ADF dispatch: unattended long runs require a machine-checkable acceptance criterion and a known-answer control run #968

Proposal

A dispatch rule for ADF: a task may run unattended for a long period only if it has (a) a machine-checkable acceptance criterion and (b) a control run on a case with a known answer, executed through the same pipeline before the real run. Tasks without both stay attended or tightly bounded (open-ended refactors, issue triage).

Why

Anthropic's nine-loop computation ran overnight on nothing but "keep going" prompts and produced a correct frontier result. Our runaway-loop failures have the same shape and the opposite outcome. The difference was not the model. The task carried its own oracle:

• an eight-loop control run, checked against the published result;
• exact arithmetic repeated modulo three primes;
• two representations cross-checked on all 107,053 nonzero coefficients;
• an independent group reproducing part of the result.

The loop detector should treat "long and making verifiable progress" differently from "long and repeating itself". A passing control run plus a checkable criterion is what makes that distinction possible.

Tasks

1. Add an acceptance field (the command or check) and a control field (a known-answer case) to the task and agent definitions used for long or unattended dispatch.
2. Spawner: refuse unattended mode when either field is missing, or when the control run fails.
3. Loop detection (#3080): use progress against the acceptance check as a signal that a long run is still productive.

Acceptance

• An unattended dispatch without both fields is rejected, with a clear message.
• A deliberately broken pipeline is caught by its control run before the real task starts.

Source

https://www.anthropic.com/research/yes-claude-can-do-nine-loops ; result page https://smsharma.io/cosmic-nine-loops/

Mirror of Gitea terraphim/terraphim-ai# (source of truth).