One official-looking reference can make a table row appear fully checked. Reading the technical-fact validator’s field-level comparison led me to preserve exactly which values have matching evidence.
A row contains several claims
Input and output are separate fields even for the same model. Evidence supporting one may not support the other. The source validator raises that field-specific question. Retaining coverage for each value helps the next reviewer decide what to do more clearly than a row-wide badge.
Coverage receives fields and evidence separately. Each evidence row needs field, value, unit, reference and checked date. Unknown fields and missing structure fail. Multiple matching references can be retained for a field instead of using one source as an answer to every blank.
A reference and a matching value differ
A reference does not cover a field when its normalized value disagrees. A valid date cannot create numeric support either. Decimal comparison on a per-million basis recognizes equivalent units while keeping mismatch visible.
The function still does not read an original document or decide truth. Evidence rows are supplied by a caller. Matching numbers with an arbitrary reference could pass the mechanical comparison, so original-source authentication remains necessary. Numeric consistency and evidence genuineness must not share one misleading accuracy label.
Use a partially covered synthetic row
The synthetic row contains input and output while evidence supports input only. Matches contains that reference and missing retains output. Coloring the entire row as verified would lose this state. Output therefore keeps covered and uncovered fields readable together.
A second case supplies a different input value and checks that a reference alone does not remove missing status. There is no actual provider or quoted source passage. These counterexamples explain a contract’s limits rather than invent an exposé of a particular dataset.
Missing evidence is not zero
Missing names a coverage gap, not a numeric zero or a claim that a service is free. An interface can show a value separately from its evidence status. The function also does not decide to delete a value solely because support is not yet recorded.
Real operations need editorial criteria for whether and how an unconfirmed field may appear. This module grants no role or publication approval. Existing review-state and role-policy modules can be integrated separately without claiming their responsibilities are already implemented here.
Separate evidence time from value applicability
Checked is a declared review date, not necessarily the start of a rate’s applicability. This small contract validates date form without deciding effective periods. Real material needing region or time conditions requires additional fields and checks; possessing one date cannot guarantee currency.
Evidence may be refreshed while the value stays equal. The new diff retains that metadata change. A system accumulating checks should distinguish changing a fact from checking it again. The next article follows that event distinction while this one focuses on coverage boundaries.
Give the next reviewer a usable table
The package’s twelve tests include partial evidence, incorrect values, equivalent units and duplicate snapshots, passing in source and extraction. They are not verification of current original price material. The browser viewer reads saved runs only, and public code has no company or account data.
A maker can hand over precise unanswered questions instead of hiding every gap. Explicit missing fields tell the next reviewer which material to seek. Sharing advanced information benefits from a narrow, readable account of what was actually supported rather than an expansive impression of completion.