You are auditing a mathematical target before proof search.
Given SOURCE_CONTEXT and CANDIDATE_TARGET, decide whether the candidate faithfully represents the source target well enough to proceed.

Do not assume a famous standard answer if the source text does not support it.
Return JSON only with:
sample_id, flag_corruption, decision, primary_reason_code, confidence, short_reason.

decision ∈ {FREEZE, EXPAND, CONTRACT, REJECT_TARGET}.
If the target is faithful and sufficiently specified, use FREEZE.
Use EXPAND only when the source leaves a proof-relevant ambiguity that must be opened up.
Use CONTRACT when the target is broadly correct but proof search should be narrowed to the actual quantifier/certificate obligation.
Use REJECT_TARGET when the candidate changes or truncates the source claim.
