Scoring — Combiner

mercury.match.evaluate(link: BankLink, company: MercuryCompany) -> LinkResult

Weighted sum

$\text{total}=s_{\text{name}}\cdot w_{\text{name}}+s_{\text{email}}\cdot w_{\text{email}}+s_{\text{phone}}\cdot w_{\text{phone}}$

Weights:

Field	Weight	Constant
Name	2.5	NAME_WEIGHT
Email	1.5	EMAIL_WEIGHT
Phone	1.5	PHONE_WEIGHT

Threshold:

Constant	Value	Comparison
MATCH_THRESHOLD	2.5	total >= threshold → Match

Invariants

Signal shape	Total	Verdict	Example
Three strong agreements	5.0	Match	Link 1
Full name + phone	4.0	Match	Links 3, 7
Full name + email	4.0	Match	Link 5
Full name alone	2.5	Match	Link 6 (nickname-resolved)
Partial name + phone	2.75	Match	Link 8
Phone alone	1.5	Mismatch	Link 2
Email alone	1.5	Mismatch	—
Partial name alone	1.25	Mismatch	Link 9
Nothing agrees	0.0	Mismatch	Links 4, 9

Rationale for the weights

The rebalanced weights reflect a specific judgment: a full first+last name match is two tokens of bio evidence on a highly specific signal, so it clears the threshold by itself (exactly at 2.5). Every narrower signal — phone alone, email alone, first-name-only, surname-only — needs a second source of agreement.

This is a simplified expression of the Fellegi–Sunter intuition: signals with high discriminating power (log m/u ratio) should dominate the decision statistic. See matching-approach for the F&S comparison.

LinkResult

@dataclass(frozen=True)
class LinkResult:
    linkId: int
    verdict: Verdict       # Match | Mismatch
    name_score: float      # [0, 1]
    email_score: float     # [0, 1]
    phone_score: float     # [0, 1]
    total: float           # weighted sum, [0, 5.5]

The CLI prints only linkId and verdict. The component scores are kept on the result so a future caller can emit graded output or audit per-field contributions without rescoring.