Solution — match (binary)

uv run mercury match <customers.json> <banks.json>

The solution the interview prompt asks for: for each link, emit either Match or Mismatch.

Decision rule

verdict = Match   if total >= MATCH_THRESHOLD
          Mismatch otherwise

MATCH_THRESHOLD = 2.5 (inclusive). See scoring-combiner for how total is constructed.

Output format

Total matches: X
Total mismatches: Y

Link 1: Match
Link 2: Mismatch
...

Ground-truth verdicts on the 9-link fixture

Link	Fraud-team comment (paraphrased)	Verdict
1	Name + company + phone match	Match
2	Phone matches, no other connection; calling customer	Mismatch
3	Name + phone match	Match
4	No connection, possible fraud?	Mismatch
5	Business name + user-as-business + email match	Match
6	Cy ↔ Cyril, last names match (nickname)	Match
7	Direct name match + phone	Match
8	Phone + first-name Cyril	Match
9	Doesn’t seem to match at all	Mismatch

Totals: 6 Match, 3 Mismatch. Verified by tests/test_solution_match.py::test_match_full_output.

Caveat — inferred verdicts

Three of the nine verdicts (links 2, 6, 8) are inferred from hedged fraud-team language (“probably good”, “going to call”). The binary output forces those into Match or Mismatch, but the fraud team was arguably not deciding. solution-triage exposes that ambiguity directly.