How decisions are made

Every planned action is resolved to apply, suggest, or suppress. First match wins:

#	Condition	Outcome
1	Rule disabled, or trigger does not match	skipped silently
2	if expression throws	skipped, logged, run continues
3	Budget spent	budget_exhausted
4	Provider returned nothing	no_answers
5	Depends on a question that was not asked	unknown_question
6	Destructive op not in allowDestructive	destructive_not_opted_in
7	Executor cannot perform the op	unsupported_by_executor
8	Named template does not exist	missing_template
9	confidence < gate.suggest	below_suggest_threshold
10	mode == 'shadow'	shadow_mode
11	mode == 'suggest', or confidence < gate.auto	suggest
12	otherwise	apply

Rule confidence

A rule is as trustworthy as its least certain input, so its gate value is the minimum confidence across the questions it depends on — gateOn when given, otherwise the ids its if expression references. Averaging would let one confident answer carry a coin flip.

A rule with neither is unconditional and has confidence 1.

Answer types

Type	Value	Confidence
noul	probability 0–1	derived as |2p − 1|
choice	the selected option	reported by the model
score	a fractional level	reported by the model

A noul carries no confidence of its own, so Dispatch derives one: 0.5 is maximally uncertain, 0 and 1 maximally certain. That lets one gate expression work across all three types.

Scores stay fractional. A score of 1.6 means the distribution sits between two levels. Rounding it before comparing against >= 2.5 would silently move your threshold.

Suppression is the product

suppressed is not diagnostics. In shadow mode it is the entire output — the list you read to decide whether you trust this. Every entry carries the rule that wanted the action, the reason, and the numbers compared:

  Suppressed
    x label: add needs-repro    [needs-repro] shadow_mode
        Shadow mode decides and logs but writes nothing.

Idempotency

A decision is keyed by sha256(target, item, repo context, question ids, config hash). Two consequences worth knowing:

• Re-delivering the same event serves the stored decision without a provider request.
• Editing your config re-triages, because the config hash is an input. Renaming a rule or moving a threshold invalidates cached decisions, as it should.

The mode is deliberately not part of the key, so promoting shadow → auto does not throw away everything Dispatch has already worked out.