# Shared document scorer 1.5.0

This study scores a complete answer separately from quotation grounding. A genuine quotation cannot supply a requested fact omitted from the answer. Every requested fact is a separate frozen `essential_claims` obligation; multiple obligations may belong to a field. All obligations and their source support must pass. The same scorer and format-only adapters apply to both workflows.

## Prospective reference contract

Every answerable reference contains `id`, `question`, `required_fields`, `expected_abstention: false`, and `essential_claims`. Each claim contains a unique `id`, `field`, `description`, canonical `value`, nonempty `accepted_values`, exact grounded `evidence`, and `answer_scope`:

- `field`: permitted only when this field has one obligation. Its explicit field/question context may identify the subject. Physical units, values, dates, per-item bases, event conditions and second conjuncts still must appear in the answer.
- `explicit`: required for each obligation of a compound field. `role_phrases` is a nonempty list of complete, literal role or event phrases. Every accepted alias must retain at least one such phrase, binding its value to the correct role. Context cannot erase that binding. For example, use `75-second ordinary scan cycle` and `15-second reference segment`, rather than two unbound duration aliases. The independent author and reviewer must confirm the phrases actually identify the roles; the structural validator cannot establish semantics by itself.

Aliases describe the shortest complete propositions the task requests, without extra unasked requirements. A person-name field accepts the name. A typed single-claim interval field can accept `45 seconds` for `every 45 seconds`, or for `one sample every 45 seconds` when the sample is explicit in the field. The duration unit stays required. A typed bicycle-count/cap field can accept `2` for `2 bicycles` because its count noun is explicit in the field. This does not remove physical units, a per-booking basis, a time basis, an event, a condition, or anything from a compound obligation. A general duration field does not inherit interval semantics. Inclusive minimum/cap context does not license opposite or exact-equality assertions.

`accepted_field_answers[field]` optionally freezes complete compositions. A string must mechanically satisfy every obligation. A complete semantic paraphrase that does not contain the literal aliases may instead be an object:

```json
{
  "text": "A complete prospectively accepted answer.",
  "obligation_ids": ["the-id-of-every-obligation-for-this-field"],
  "completeness_note": "Explain exactly where each requested value, role and condition is conveyed."
}
```

The independent author and reviewer must attest that each object is true, complete and unambiguous. `obligation_ids` must cover the entire field exactly. Such an object matches only its complete normalized token sequence; its phrases cannot be mined as free aliases or applied to a longer answer. It never changes citation requirements. This explicit semantic annotation is evidence to review, not an automatic entailment claim. Freeze all aliases, roles and complete compositions before either workflow produces output. No post-output additions or historical rescoring.

## Decisions and limitations

A matched subset-only composition is a definite incomplete answer. A short substring alone no longer proves that an answer is wrong. Unregistered wording that could express a missing obligation stays `unresolved`; this includes complete wording copied exactly from a genuine alternative source passage. Exact copying establishes location, not automatic completeness or error. Such answers earn no pass and must not be reported as proven semantic errors. Opposite operators/conditions, conflicting literal values, answerable declines, absent documents and fabricated quotations still fail. Finite recognition cannot establish an overall semantic winner without a separate audit of the actual outputs.

A correct supported answer needs all obligations, no unreviewed factual additions, and frozen evidence for every obligation. Source prose is not exempt from completeness. Native answers are parsed only by requested field labels; raw outputs, duplicate labels, missing fields, extra content and all emitted citation markers remain accounted for. Automatically displayed source cards do not count as emitted citations. A one-field answer need not repeat its field label. Two labels on one line are a format difference, not a factual failure.

Citation precision counts emitted citations that uniquely locate an actual PDF span supporting at least one obligation in their associated field. Complete support coverage is additionally required for answer success. Explicit page citations and quote-mapped pages are separate capabilities. Cross-page quotations require consecutive original pages, the correct document, and unchanged ordinary word/number boundaries. Both outer quotation edges must preserve complete source tokens: `13 hours` cannot be cut out of `113 hours`, a sign or decimal cannot be removed, and `anna` cannot be cut out of `Joanna`. Identifier/name joins such as hyphens, slashes and apostrophes remain bound. These checks also cover a continuation over a page seam. Ordinary punctuation and complete internal extraction line/page joins remain accepted. A value found in an obsolete passage or another document cannot substitute for the frozen current source anchor. Grounded but unregistered same-document support stays unresolved unless a literal contradiction is established; a same-document location is never sufficient for a support pass.

Frozen support anchors use the same unique full-token grounding as emitted quotations. Support means the emitted citation contains that exact original source occurrence, not a compact-text substring found elsewhere. A gold `13 hours` occurrence cannot be satisfied by an archived `113 hours` occurrence. Declared raw character offsets (`char_start`/`char_end`, or the historical `start_char`/`end_char` spelling) must agree with the unique original occurrence during reference validation; malformed, conflicting or substring-directed offsets reject the bank before measurement.

For wholly unanswerable requests, every field must decline. A correct value in one field plus UNKNOWN in another is not a clean decline. Speculation or a supplied missing value fails; ambiguous absence explanations remain unresolved. Numeric zero or a no-fee policy is not documentary absence.

Qualification uses conservative lower bounds in each of three complete Kingy repetitions: at least 90% complete supported answers, 95% citation precision, and 90% clean abstentions. Every planned original attempt is retained. Unresolved judgments cannot earn a pass. No answer retries, best-of selection or post-output rubric changes. Package installation, four fact-free self-tests, runtime authentication and local-only network checks also must pass before launch. Forty unique questions repeated three times are forty questions, not 120 independent trials. This is an AI-assisted local comparison, not human adjudication or independent hardware reproduction.