# Shared document scorer 1.4.0

This study scores task completeness separately from whether a quotation occurs in a PDF. A genuine source quotation cannot supply a fact that the answer omitted. Every requested fact is an explicit obligation in `essential_claims`; multiple obligations can belong to one answer field. All obligations must be satisfied by that field's answer. Alternative phrasings are frozen before either workflow produces an answer.

Each obligation has a unique `id`, `field`, brief `description`, `value`, nonempty `accepted_values`, and grounded `evidence` spans. Values are the shortest complete propositions or values that answer the question in its stated context. A name question accepts the name without requiring an unasked title. A value question accepts the value without repeating the subject. Required units, condition, threshold, deadline and time basis remain obligations. Do not include unstated or unasked requirements. Compound requests have separate obligations, never a single vague broad source span. Evidence spans may be shared where one sentence states several obligations.

`accepted_field_answers` optionally lists full compositions for a field. Every composition must still satisfy every atomic obligation; it never bypasses completeness. The scorer otherwise recognizes complete literal obligation aliases joined by simple punctuation/words or field labels. Exact source prose is acceptable wording only after every obligation matches. Unregistered semantic paraphrases remain unresolved, not silently passed or labelled wrong.

Native answers are parsed by explicit requested field labels. Newlines and semicolons can separate labels. Citation markers attach only to their parsed field. Duplicates, missing fields and unassigned prose cannot disappear. Two fields on one line are a format difference, not a factual failure. Raw outputs remain unchanged.

Avoid ambiguous yes/no questions and conflicting field labels. Freeze affirmative, descriptive labels such as `waiver_status`, and explicit aliases such as `no waiver` or `requirements unchanged`. A bare polarity answer without an unambiguous proposition remains unresolved. Question context can identify the subject but never supplies a missing value, unit, condition or second conjunct.

A correct supported answer requires all requested obligations, no unreviewed factual additions, and source evidence covering every obligation. Citation precision counts emitted citations that uniquely locate a true PDF span supporting at least one registered obligation in their associated field. Complete support coverage is additionally required for answer success. Native automatically shown cards do not count as emitted citations. Page-mapped quotations and explicit page citations remain separate capabilities. Cross-page quotes require consecutive original pages and preserve ordinary word/number boundaries.

For wholly unanswerable requests, every requested field must decline. A correct value in one field plus UNKNOWN in another is not a clean decline. Speculation or any 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 least90% complete supported answers,95% citation precision,90% clean abstentions; every planned attempt is retained. No post-output aliases, answer retries, best-of selection or rescoring of historical studies. Unresolved judgments cannot earn a pass. 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 are40 questions, not120 independent trials. This is an AI-assisted local comparison, not human adjudication or independent hardware reproduction.