Source Provenance Chain Reconstruction Prompt

This prompt is for investigators and verification engineers who need to reconstruct the full provenance chain of a claim, tracing it backward through republishing layers to identify the original evidence source. Use it when a claim appears in a news article, social media post, or report that cites another source, which in turn cites another, and you need to map every link in that chain. The prompt identifies where attribution breaks, where sources are cited without access, and where circular references create false corroboration. It belongs in verification pipelines after claim extraction and before evidence matching, giving downstream systems a clear map of what was actually sourced versus what was merely repeated.

The ideal user has a specific claim and a set of source documents that reference each other. You must provide the claim text and the content of each document in the chain, with enough metadata (publication dates, URLs, author bylines) to let the model distinguish between independent reporting and syndicated repetition. The prompt works best when you can supply at least two layers of citation—for example, a news article that cites a press release, which in turn references a study. Without at least one explicit citation link between documents, the model cannot build a chain and will return an empty or single-node result. Do not use this prompt for general source credibility scoring, for verifying the truth of the claim itself, or for single-document analysis where no republishing chain exists.

Before running this prompt, ensure your pipeline has already extracted discrete claims and gathered the documents that cite or are cited by each other. The prompt expects structured input with clear document boundaries. After receiving the output, validate that every link in the chain includes a specific citation reference (a quote, a URL, or a named attribution). If the model produces a chain link without a citation, treat that link as unverified. For high-stakes verification, route any chain with missing links, circular references, or anonymous-source endpoints to human review before using the provenance map in downstream evidence matching. The next section provides the copy-ready template you can adapt for your own verification harness.

This prompt template is designed to be pasted directly into your system instructions or as a user message in an AI harness. It instructs the model to trace a claim back to its original source by analyzing a provided set of intermediary documents. The core task is to build a linear provenance chain, identify where attribution breaks, and detect circular citations. This is a high-stakes verification workflow; the output must be auditable, and the model should abstain rather than invent missing links.

markdownCopy
You are a source provenance investigator. Your task is to reconstruct the full provenance chain for a specific claim by analyzing a set of provided source documents that may represent different stages of republishing.

**Inputs:**
- [CLAIM]: The exact claim text to trace.
- [SOURCE_DOCUMENTS]: A list of source documents, each with a unique ID, full text, and available metadata (publication date, author, publisher, URL).

**Instructions:**
1.  **Identify the Earliest Instance:** Scan all [SOURCE_DOCUMENTS] to find the document that contains the [CLAIM] and has the earliest verifiable publication date. This is the starting point for the chain.
2.  **Build the Chain Forward:** From the earliest instance, trace the claim forward in time. For each subsequent document that contains the claim, determine if it explicitly cites a prior source in the chain. If it does, link them. If it does not, mark it as a **Provenance Gap**.
3.  **Detect Circular Citations:** Check if any document in the chain cites another document that appears later in the chain, creating a loop. Flag any circular citations clearly.
4.  **Flag Attribution Gaps:** For any link where a document does not cite a prior source, note the missing attribution and state that the provenance is broken at that point.
5.  **Abstain if Uncertain:** If you cannot determine the relationship between documents or the claim's origin is ambiguous, state that the chain is incomplete. Do not fabricate links.

**Output Format:**
Produce a JSON object with the following structure:
{
  "claim": "[CLAIM]",
  "provenance_chain": [
    {
      "source_id": "string",
      "publication_date": "ISO 8601 string or null",
      "relationship_to_prior": "origin" | "cites" | "uncited_republication",
      "cited_source_id": "string | null",
      "notes": "string explaining the evidence for the link or the nature of the gap"
    }
  ],
  "circular_citations_detected": [
    {
      "source_id": "string",
      "cited_source_id": "string",
      "explanation": "string"
    }
  ],
  "overall_assessment": "string summarizing chain completeness and key risks"
}

**Constraints:**
- Only use the provided [SOURCE_DOCUMENTS]. Do not introduce external knowledge.
- If publication dates are missing, note this ambiguity and explain how it affects the chain.
- Prioritize precision over recall. A short, verified chain is better than a long, speculative one.

To adapt this template, replace the [CLAIM] and [SOURCE_DOCUMENTS] placeholders with your specific inputs. The [SOURCE_DOCUMENTS] should be a structured list, likely injected by your application's retrieval or data-loading step. For high-risk domains like legal or financial verification, you must add a human review step after this prompt runs. The JSON output is designed for programmatic consumption, so wire it to a validation function that checks for the required schema and flags any uncited_republication entries for immediate analyst attention. Do not treat this prompt's output as a final verdict; it is a structured input for a downstream review or audit logging system.

This prompt is designed to be a single step in a larger verification pipeline, not a standalone chat interaction. The primary integration point is after claim extraction and evidence retrieval, where you have a target claim and a set of candidate sources that may have been republished, syndicated, or cited through multiple layers. The harness should call this prompt once per claim-source pair that requires provenance tracing, feeding in the claim text, the source content, and any available metadata such as publication dates, author bylines, and outlet names. Because provenance reconstruction is inherently recursive—a source cites another source which cites another—you should plan for the harness to make multiple calls, using each identified upstream source as the input for the next iteration until an original source is reached or the chain breaks.

The implementation should wrap the prompt in a function that enforces a strict JSON output contract. Define a schema that includes: provenance_chain (an ordered array of nodes, each with source_identifier, publication_date, attribution_type such as 'original', 'citation', 'syndication', or 'unknown', and content_snippet), chain_completeness (an enum of 'complete', 'broken', or 'circular'), broken_link_index (the array index where the chain breaks, or null), and circular_reference_detected (boolean). After each model response, validate the JSON against this schema. If validation fails, retry with the error message appended to the prompt as a correction instruction, up to a maximum of two retries. For high-stakes verification workflows, route any response where chain_completeness is not 'complete' to a human review queue with the partial chain and the original source content attached.

Model choice matters here. This prompt requires strong instruction-following and structured output discipline. Prefer models with native JSON mode or function-calling capabilities. If using a model without guaranteed JSON output, implement a repair step that attempts to extract and fix malformed JSON before falling back to retry. Log every call—including the input claim, source metadata, raw model response, parsed output, validation result, and retry count—to your observability platform. This audit trail is critical for debugging chain breaks and for demonstrating verification rigor to downstream consumers. Avoid using this prompt on sources that are already classified as primary or original by a prior source-type classification step; reserve it for cases where republishing is suspected or the source lineage is explicitly unclear.