modules/deep-research-validator/references/workflow-reference.md

# Validate Answer Workflow Reference

This reference expands the core workflow in `SKILL.md`. Load it when a run needs detailed review categories, source-support wording, or packaging interpretation.

## Claim Selection

Review material claims rather than every sentence. Prioritize:

- conclusions and recommendations;
- legal, tax, compliance, or eligibility positions;
- numeric claims, percentages, dates, thresholds, and deadlines;
- causal claims and risk statements;
- claims used as premises for later conclusions.

Skip purely introductory, stylistic, or duplicated statements unless they carry a material conclusion.

Claim selection is itself a semantic review step. Read the full answer and
record reviewed sections, omitted sections, limitations, and whether the scope
is `all_material_claims`, `selected_material_claims`, or `limited`. Mechanical
sentence extraction is only navigation help.

## Answer-Contract Conformance

Before claim-by-claim review, assess separately whether the answer addresses the
question, uses the contracted document type, suits the intended audience, and
follows the evidence-display requirement. A source-supported answer can still
fail because it answers the wrong question or produces the wrong artifact.

## Semantic Support Statuses

Use these support statuses in `claims_review.json`:

- `supported`: the cited source directly supports the same fact or conclusion.
- `partially_supported`: the source supports part of the claim or supports a narrower conclusion.
- `not_supported`: the source does not support the claim.
- `contradicted`: the source conflicts with the claim.
- `uncertain`: support cannot be determined because sources are unavailable, gated, too short, or ambiguous.

## Review Dimensions

For each reviewed claim, separate:

- source access: URL/file reachable and captured;
- source identity: model-led assessment that the captured item is the authority
  actually cited, including version, jurisdiction, and period;
- exact passage presence: mechanical observation made only within the
  specifically cited source snapshot;
- semantic support: model-led assessment that the source actually entails,
  narrows, qualifies, or contradicts the claim;
- reasoning: model-led assessment that the conclusion follows from supported
  premises and identification of missing premises;
- professional judgment: explicit factors and alternative interpretations;
- issue treatment and final claim disposition.

## Official Archive Coverage And Reproductions

Treat these as separate questions:

1. Does the decision or authority exist, and is its identifier corroborated?
2. Was the official full text obtained?
3. Is the public official portal exhaustive, rolling, partial, or access-gated?
4. Is the reviewed item an official text, official summary, non-institutional
   reproduction, or commentary?
5. Does the available text substantively support the claim?
6. Was the reproduction's text verified against the official full text?

Do not infer that a decision is absent or nonexistent merely because an
official public search returns no result. First establish the portal's coverage
from official documentation or other reliable evidence. A visible year facet
describes the current result set; by itself it does not prove why older records
are absent or whether a separate historical archive exists.

When the official historical archive is outside a documented rolling public
window or is authentication-gated, record that access condition without
treating it as a defect in the decision. A non-institutional reproduction can
still support a claim when the authority's identity and substance are
corroborated. Unless the official text was obtained or an exact comparison was
possible, record `text_fidelity: corroborated_not_text_verified`, explain the
corroboration, and disclose a residual text-fidelity limitation. This source
limit does not automatically make otherwise supported reasoning unsound.

Use `not_found_in_complete_official_archive` only when reliable evidence shows
that the searched official archive is complete for the relevant authority,
document type, jurisdiction, and period. A search failure, a rolling public
window, or a gated archive is insufficient.

Exact or fuzzy text matching is not semantic support. For example, "this is a
terrier" can support "this is a dog" without literal overlap, while "this is
not a dog", "this will be a dog", and "this was a dog" have different
negation or temporal meaning despite strong lexical overlap.

## Validation Boundary

Keep four dimensions explicit:

- mechanically observed: document or source access, identifiers, exact quote
  presence in the available text, and JSON/schema shape;
- model-led source identity and semantic support: authority identity, claim
  relation, archive-coverage significance, text fidelity, claim meaning,
  entailment, contradiction, scope, qualification, time, and modality;
- model-led reasoning: whether the conclusion follows from supported premises
  and whether intermediate premises are missing;
- judgment-dependent: legal applicability, materiality, competing
  interpretations, strategy, reasonableness, and uncertain professional
  outcomes.

Deterministic observations may collect evidence but must not decide semantic
support. A passing deterministic audit means the review record is complete
enough to inspect; it does not certify legal correctness.

## Issue Types And Treatments

Use one or more issue types for each claim:

- `source_unavailable` → obtain the source or mark evidence blocked;
- `source_not_identified` → identify the cited authority;
- `wrong_source`, `wrong_source_version`, or
  `wrong_jurisdiction_or_period` → replace the source and reassess support;
- `missing_source_support` → add support, remove the claim, or state
  uncertainty;
- `partial_or_overbroad_support` → narrow the claim;
- `source_contradiction` → correct or remove the claim;
- `qualification_or_scope_distortion` → restore the source's qualification;
- `temporal_or_modality_distortion` → correct tense, timing, certainty, or
  modality;
- `reasoning_gap` → supply the missing premise when supported, otherwise
  narrow, caveat, or remove the conclusion;
- `judgment_dependent` → state uncertainty and require professional review;
- `answer_contract_failure` → revise the answer to the contracted question,
  document type, audience, or evidence display.

Use `none` alone when no issue was identified. Record one of these treatment
actions for every issue: `obtain_source`, `identify_source`, `replace_source`,
`add_support`, `narrow_claim`, `correct_claim`, `restore_qualification`,
`correct_time_or_modality`, `add_reasoning`, `state_uncertainty`,
`remove_claim`, `professional_review`, or `revise_answer_contract`. Treatment
status is `proposed`, `applied`, `blocked`, or
`professional_review_required`; `none` uses `none` / `not_needed`.

## Review-State Consistency

The deterministic layer may reject contradictions among explicit review codes
because that is record consistency, not semantic judgment:

- any support, reasoning, source-identity, or professional-judgment attention
  requires a non-`none` issue and treatment;
- `supported`, `partially_supported`, and `contradicted` assessments require a
  source check;
- `not_supported`, `contradicted`, and `unsound` records cannot retain the
  claim unchanged;
- answer-contract attention and `answer_contract_failure` treatment must match
  in both directions;
- reviewer rejection, proposed or blocked treatment, and unresolved contract
  attention cannot produce `reviewed_answer_ready`;
- `document_revision.status: completed` requires
  `overall_assessment.outcome: corrected`, while unresolved attention cannot
  coexist with `no_material_defect_identified`.

These rules do not decide whether a claim is supported, reasoning is sound, or
professional judgment is required. They ensure only that the semantic
reviewer's own recorded decisions do not contradict each other.

## Output Guidance

The validation package is not a new substantive answer. It is a review record.
State assumptions and source limits clearly. Preserve the user's document
structure when writing a corrected Markdown document. A proposed fix in the
review record is not a corrected answer: the answer must be regenerated
semantically and packaged again.

## Deterministic Audit Interpretation

`validation_audit.json` is a guardrail, not a legal conclusion. Its
`record_integrity_status` says whether required review fields and treatments
were recorded. Its `delivery_readiness` aggregates those explicit statuses; it
does not infer legal correctness. Passage observations are bound to the selected
verified capture and its extraction scope. Literal checks preserve case for
verbatim passages and translated originals while normalizing whitespace. A
missing paraphrase does not invalidate semantic support. A missing quotation
requires a documented original inspection or a disclosed evidence limit; an
unexplained discrepancy or changed capture prevents a ready review.