Tax Authority Evidence Ranking Prompt Template

This prompt is designed for tax and accounting software teams building AI features that must reason over a corpus of tax authority documents. It ranks retrieved statutes, IRS rulings, treasury regulations, private letter rulings, and other guidance by jurisdiction, authority level, effective date, and taxpayer-specific applicability. The output is a hierarchical evidence brief that downstream answer-generation prompts can consume. Use this prompt when generic relevance ranking is insufficient and you need the model to apply domain-specific authority rules before synthesizing an answer.

The ideal user is an engineering lead or AI product manager integrating tax research capabilities into a professional tax preparation, audit support, or corporate tax workflow product. Required context includes a pre-retrieved set of tax documents, the taxpayer's jurisdiction and entity type, the tax year or period in question, and the specific tax issue being researched. The prompt expects the model to weigh primary authority (Internal Revenue Code, Treasury Regulations, judicial decisions) above secondary authority (IRS publications, private letter rulings, commentary), and to flag superseded or obsolete guidance. It is not a substitute for a tax professional's judgment; always route final determinations to a qualified human reviewer.

Do not use this prompt when the retrieved document set is empty, when the tax question is purely computational, or when the user expects a final tax position rather than an evidence brief. This prompt is a ranking and structuring step in a multi-stage pipeline, not an end-to-end answer generator. If your application requires a direct answer to a taxpayer question, chain this prompt's output into a downstream synthesis prompt that includes explicit grounding and citation instructions. For high-risk scenarios involving penalty exposure, criminal tax matters, or positions contrary to published guidance, add a mandatory human-review gate after the evidence brief is generated and before any answer reaches the user.

This template is the core instruction set for ranking tax authority evidence. It is designed to be copied directly into your model request, with placeholders replaced by your application's runtime data. The prompt enforces a strict hierarchy: jurisdiction first, then authority level, effective date, and taxpayer-specific applicability. It also requires the model to flag superseded guidance and circular cross-references, which are common failure modes in tax research.

textCopy
You are a tax research evidence ranking assistant. Your task is to analyze the provided tax documents and produce a structured JSON brief that ranks evidence by authority and applicability.

## INPUT
[QUERY]
[RETRIEVED_DOCUMENTS]

## CONSTRAINTS
- Rank evidence using this strict hierarchy:
  1. Jurisdiction match (federal > state > local; must match taxpayer jurisdiction in [TAXPAYER_CONTEXT])
  2. Authority level (Statute > Regulation > Revenue Ruling > Private Letter Ruling > IRS Notice > Publication > Commentary)
  3. Effective date (most recent applicable period first; flag any document not yet effective or expired)
  4. Taxpayer-specific applicability (directly addresses facts in [TAXPAYER_CONTEXT] > analogous situation > general guidance)
- For each piece of evidence, assign a confidence score from 0.0 to 1.0 reflecting how directly it applies.
- Flag any evidence that has been superseded, revoked, or modified by later authority. Cite the superseding document if present in the retrieved set.
- Detect and flag any circular cross-references where Document A cites Document B and Document B cites Document A as primary authority.
- If two documents conflict, surface the conflict explicitly and explain which authority prevails under the hierarchy.
- Do not invent citations, code sections, or rulings not present in the input.

## OUTPUT_SCHEMA
Return a single JSON object with this structure:
{
  "query_summary": "string summarizing the tax question",
  "jurisdiction_analysis": {
    "taxpayer_jurisdiction": "string",
    "matching_documents": ["list of document IDs matching jurisdiction"],
    "non_matching_documents": ["list of document IDs from other jurisdictions"]
  },
  "ranked_evidence": [
    {
      "rank": 1,
      "document_id": "string",
      "authority_type": "Statute | Regulation | Revenue Ruling | Private Letter Ruling | IRS Notice | Publication | Commentary",
      "title": "string",
      "effective_date": "YYYY-MM-DD",
      "applicability_score": 0.0-1.0,
      "key_excerpt": "relevant quote from document",
      "reasoning": "why this evidence is ranked here",
      "flags": ["SUPERSEDED_BY_DOC_X" | "EXPIRED" | "NOT_YET_EFFECTIVE" | "CIRCULAR_REFERENCE" | "CONFLICT_WITH_DOC_Y"]
    }
  ],
  "conflicts": [
    {
      "document_a_id": "string",
      "document_b_id": "string",
      "conflict_description": "string",
      "prevailing_authority": "document_id of prevailing authority",
      "resolution_rationale": "string"
    }
  ],
  "superseded_guidance": [
    {
      "superseded_document_id": "string",
      "superseding_document_id": "string or null if not in retrieved set",
      "supersession_evidence": "quote or explanation"
    }
  ],
  "circular_references": [
    {
      "document_a_id": "string",
      "document_b_id": "string",
      "description": "string"
    }
  ],
  "gaps": ["list of missing authority types or unanswered sub-questions"],
  "overall_confidence": 0.0-1.0
}

## TAXPAYER_CONTEXT
[TAXPAYER_CONTEXT]

## EXAMPLES
[FEW_SHOT_EXAMPLES]

To adapt this template, replace the square-bracket placeholders with your runtime data. [QUERY] should contain the user's tax question. [RETRIEVED_DOCUMENTS] should contain the full text and metadata of each retrieved document, including document IDs, titles, dates, and authority types. [TAXPAYER_CONTEXT] must include jurisdiction, entity type, tax year, and relevant facts. [FEW_SHOT_EXAMPLES] should contain 2-3 worked examples showing correct ranking, conflict resolution, and flagging behavior. If your retrieval system does not provide document IDs, generate them from a hash of the document content before calling the prompt.

Before deploying, validate the output against the schema using a JSON schema validator. Run eval suites that include known superseded documents, circular references, and jurisdiction mismatches. If the model fails to flag a known superseded document or misses a circular reference, add that case to your few-shot examples. For high-stakes tax workflows, route outputs with overall_confidence below 0.7 or with unresolved conflicts to a human reviewer before surfacing them to users. Do not rely on this prompt alone for final tax positions—always require human review for interpretive guidance and PLR applicability determinations.

This prompt is designed to operate as a stateless ranking microservice within a larger RAG pipeline. It expects pre-retrieved passages with complete metadata as input and does not perform document retrieval itself. The calling application is responsible for assembling the [RANKED_EVIDENCE] array, ensuring each passage object contains the required fields (passage_id, text, jurisdiction, authority_type, effective_date, cross_references) before the model call. Any missing or malformed fields must trigger a pre-call validation error that halts the workflow and returns a structured error to the upstream retriever, not a degraded prompt to the model.

After the model returns a ranked evidence brief, run a post-processing validator that checks for superseded guidance. This validator must compare effective_date values across passages and flag any source that has been explicitly overridden by a later authority of equal or higher level within the same jurisdiction. It must also detect circular cross-reference chains (e.g., Ruling A cites Ruling B, which cites Ruling A) and mark those passages for human review. The validator should output a validity_flags array alongside the ranked brief, with each flag linked to a specific passage_id and a machine-readable reason code (SUPERSEDED, CIRCULAR_REF, JURISDICTION_MISMATCH).

Every ranking decision must be logged for auditability. Capture the model version, a SHA-256 hash of the prompt template, the full list of input passage IDs, the ranked output order, and any validity flags in a structured log entry. Store these logs in an append-only datastore with a retention period matching your organization's audit requirements. For high-stakes tax positions—defined as any ranking involving penalty exposure, statute of limitations analysis, or positions contrary to published IRS guidance—route the ranked evidence brief to a human tax professional via a review queue before it reaches the end user. The review interface should display the ranked passages, validity flags, and a diff view showing which passages the model deprioritized and why.

Implement a retry loop with exponential backoff for malformed JSON outputs. If the model response fails JSON schema validation, retry up to three times with backoff intervals of 1s, 4s, and 16s. On each retry, append a brief repair instruction to the prompt context indicating the specific schema violation. After three failed retries, escalate to a human queue with the full prompt payload, the failed outputs, and the validation error log. Do not silently fall back to an unranked passage list or a degraded response. For model selection, prefer models with strong JSON mode and structured output support (e.g., GPT-4o with response_format, Claude 3.5 Sonnet with tool-use mode). Avoid models that require fragile regex post-processing for JSON extraction.