[Fix] Helper-model decisions answer yes/no questions with their own confidence instead of a probability

[mrubens]

Problem

When no hosted judgment model is configured, typed decisions fall back to the helper model. The fallback prompt asks for { type: 'noul', noul: number } per yes/no question but never says what noul is. Several small models read it as "how sure am I of my answer" and return ~0.99 whether the answer is yes or no. Every caller then sees a confident yes.

This affects every surface that uses the fallback (channel launch criteria, work-kind routing, thread-reply routing, forum tags, and the rest of the decision-model list), not one feature.

Change

The fallback prompt now defines the fields:

• noul is the probability that the answer is yes, a confident no is near 0, and ~0.5 is for questions the state does not settle.
• score is the zero-based index into the ordered criteria.

The same definition is attached to the noul field of the structured-output schema.

Evidence

Live probe through OpenRouter using the fallback's exact system prompt, prompt, and answer schema: 9 synthetic coding turns x 3 yes/no questions x 2 runs per model, scored against expected answers at a 0.85 threshold.

Model	Wrong before	Wrong after
gpt-5.6-luna	18 of 54	2 of 54
glm-5.3-flash	20 of 51	1 of 48
claude-haiku-4.5	3 of 54	2 of 54
gemini-3.8-flash	1 of 51	1 of 54

Before the change the failures were almost all false yeses at 0.95-1.00 on turns where the right answer was no. The remaining misses are borderline cases scored at 0.85-0.90.

Not addressed: helper models still snap to round numbers, so thresholds tuned on the hosted judgment model are not automatically right for the fallback, and helper latency (2-20 s) is unchanged.

Validation

• New test pins the definitions in the fallback prompt.
• typesafe-judgment.test.ts passes (44), pnpm lint:fast, pnpm check-types:fast.

[roomote-community]

No code issues found. See task

_{Reviewed e3d081a}