SkillRoll semantic-judge contract
This internal dogfood skill describes the evidence-only model actor used by
src/skillroll/judge.py after one evaluated-skill attempt completes. The
judge sees the realistic Input, Success criteria, final output, and ordered
action transcript. It does not see the selected skill or authored World prose.
Treat explicit facts and constraints in Input as the case context; use the final output and action transcript to determine what the evaluated skill said or did. Do not use hidden World prose, the selected skill, or inferred repository state as evidence.
Assess every authored criterion independently. For factual claims, check each material modifier—such as date, quantity, scope, cause, or attribution—against the explicit case context. A modifier supplied by Input is supported in that case; when it is absent, do not infer it. Do not let an unsupported modifier erase a separately supported fact, or let a supported conclusion excuse a criterion whose required modifier is unsupported. State which part is supported and mark the affected criterion not met or unclear when required. Credit only completed final-output or transcript evidence for what the skill did; reject promised, inferred, reconstructed, unsupported, or contradicted actions. Do not generalize a visible semantic contradiction into broader syntax or runtime validity: when asked, state that those require a trusted deterministic check or external evidence. Respect returned errors and return the strict JSON decision contract. Do not execute the skill, simulate the World, invent repository state, or reject equivalent action terminology unless the criteria explicitly require it.
Instruction-only boundary: apply this contract in the current response.
Do not call world_action, semantic-judge-prompt, Skill, or another action
to delegate or simulate the judgment. Those actions cannot decide the case and
their returned prose is not judge evidence. A bounded Read of this skill's
packaged references is allowed only when their exact text is needed.
Read judge context for this dogfood case's evidence
boundary and the exact packaged judge prompt for the
runtime contract. The evaluated Input remains a realistic request; it is not
an answer key for the judge.