Temporal Grounding and Outdated Claim Detection Prompt

This playbook is for RAG developers and quality engineers who need to audit generated outputs for temporal accuracy. The prompt inspects a production trace containing the user query, retrieved documents with timestamps, and the final generated answer. It flags claims that are outdated, anachronistic, or presented without proper temporal qualification. Use this when your knowledge base contains time-sensitive information such as financial reports, legal statutes, medical guidelines, product documentation, or news articles where staleness directly causes user harm.

This prompt belongs in an offline trace review or continuous monitoring pipeline, not in the real-time generation path. It assumes you have already captured a trace with retrieval metadata including document publication or last-updated timestamps. The ideal user is an engineer or analyst who can interpret the trace structure and act on the findings—whether that means updating a knowledge base, adjusting retrieval freshness filters, or tuning the model's instructions to respect temporal context. Do not use this prompt when your knowledge base is static, when timestamps are unavailable in your traces, or when you need a real-time guardrail. For real-time intervention, you would need a separate, latency-sensitive classifier that checks document recency before generation, not a post-hoc trace auditor.

Before running this prompt, ensure your trace format includes a query_timestamp (when the user asked the question), a retrieved_documents array with each document's pub_date or last_updated field, and the generated_answer text. The prompt works best when documents have explicit, machine-readable timestamps rather than free-text date mentions. If your retrieval pipeline already filters by recency, this prompt serves as a verification layer to catch filter misconfigurations, edge cases where stale documents slip through, or model behaviors where the LLM ignores timestamp metadata and treats old information as current. After running the audit, route findings to your knowledge base maintenance queue if the root cause is stale source data, or to your prompt engineering backlog if the model is failing to qualify temporal claims despite having fresh context.

This prompt is designed to be pasted directly into your trace review tool, evaluation harness, or orchestration framework. It instructs the model to act as a temporal accuracy auditor, cross-referencing every factual claim in a generated output against the publication or modification timestamps of the source documents retrieved during that specific trace. The goal is to flag statements that are factually outdated, presented without necessary temporal qualification, or anachronistically applied to a time period for which the source evidence is no longer valid.

textCopy
You are a temporal grounding auditor. Your task is to review a single production trace and identify claims in the generated output that are outdated, anachronistic, or presented without required temporal qualification.

## INPUT
- Generated Output: [GENERATED_OUTPUT]
- Retrieved Documents with Timestamps: [RETRIEVED_DOCUMENTS]
- Current Date: [CURRENT_DATE]

## TASK
For each factual claim in the Generated Output, determine if it is temporally grounded in the Retrieved Documents. A claim is temporally grounded if the source document's timestamp is recent enough to support the claim's implied currency and the claim does not apply old information to a current context without qualification.

## OUTPUT_SCHEMA
Return a single JSON object with the following structure:
{
  "temporal_accuracy_report": {
    "overall_score": "PASS" | "FAIL" | "WARN",
    "claims": [
      {
        "claim_text": "string",
        "source_document_id": "string | null",
        "source_timestamp": "ISO8601 | null",
        "temporal_status": "CURRENT" | "OUTDATED" | "ANACHRONISTIC" | "UNQUALIFIED" | "NO_SOURCE",
        "explanation": "string",
        "recommended_action": "string"
      }
    ],
    "summary": "string"
  }
}

## DEFINITIONS
- CURRENT: The claim is supported by a source whose timestamp is within an acceptable recency window for the claim's domain.
- OUTDATED: The source is too old to support the claim as a current fact. A newer source is likely required.
- ANACHRONISTIC: The claim applies information from one time period to another where it is not valid (e.g., using a 2019 statistic to describe a 2024 event).
- UNQUALIFIED: The claim is presented as a timeless or current fact, but the source is old and the output lacks necessary temporal qualifiers like "As of 2021..." or "At the time of publication...".
- NO_SOURCE: No retrieved document supports this claim, regardless of timestamp.

## CONSTRAINTS
- Only flag a claim if the temporal mismatch would mislead a reasonable user.
- Do not flag claims that are explicitly historical in nature.
- If a document timestamp is missing, treat the claim as UNQUALIFIED and note the missing metadata.
- For [RISK_LEVEL] set to "HIGH", any OUTDATED or ANACHRONISTIC claim must result in an overall_score of "FAIL".

To adapt this template, replace the square-bracket placeholders with values extracted from your production trace. [GENERATED_OUTPUT] should contain the final text the user saw. [RETRIEVED_DOCUMENTS] must be a structured list of documents, each with a unique ID, content snippet, and a timestamp field. If your trace captures document metadata separately, inject it here. [CURRENT_DATE] sets the reference point for recency. The [RISK_LEVEL] variable controls the strictness of the final verdict; set it to "HIGH" for regulated, financial, or medical use cases where presenting outdated information as current is a compliance risk. After running the prompt, always validate that the output conforms to the OUTPUT_SCHEMA before surfacing results in a dashboard or alerting system. For high-risk domains, route any output with an overall_score of "FAIL" or "WARN" to a human review queue before the end user sees it.

This prompt is designed to operate as a post-hoc analysis step within a trace review pipeline, not as a real-time guardrail during generation. It expects a structured trace object containing the generated output, the retrieved context with document timestamps, and the current evaluation date. The implementation harness should extract these fields from your observability store, assemble the prompt, and route the resulting temporal accuracy report to the appropriate review queue or dashboard.

Wire the prompt into a batch processing job that queries your trace database for sessions where temporal accuracy is critical—for example, RAG responses citing financial reports, regulatory documents, or product specifications with known effective dates. The harness must validate that each trace contains retrieved_documents with a publication_date or effective_date field before invoking the prompt. If timestamps are missing, log a schema violation and skip the trace rather than feeding incomplete data to the model. For high-risk domains such as healthcare or compliance, route all OUTDATED and ANACHRONISTIC findings to a human review queue with the trace ID, the flagged claim, and the conflicting document timestamps attached. Use a structured output parser to extract the JSON report from the model response, and implement a retry with a stricter format reminder if parsing fails on the first attempt.

For continuous monitoring, deploy this prompt as a scheduled evaluation job that samples recent production traces and appends temporal accuracy scores to your observability dashboard. Set alert thresholds on the temporal_accuracy_score field—for instance, trigger an on-call notification if the 7-day rolling average drops below 0.85. Avoid using this prompt on traces where the retrieved context lacks timestamp metadata, as the model will be forced to guess temporal relationships and may produce unreliable classifications. Pair this prompt with a retrieval pipeline audit to ensure document timestamps are consistently populated before scaling the monitoring job across your full trace volume.