Temporal Hallucination Detection Prompt

This prompt is a guardrail for output validation pipelines. Its job is to inspect a generated text and its cited sources to detect fabricated dates, unsupported claims of recency, and misrepresentations of when a source was published or what time period it covers. The ideal user is an AI safety engineer or validation pipeline builder who needs structured hallucination flags, not free-text critiques. You should use this prompt after answer generation and before the response reaches the user. It assumes you already have a model output and the evidence passages that were used to generate it. It does not retrieve evidence or rewrite answers. It only detects temporal grounding failures.

To use this prompt effectively, you must provide the full model-generated answer and the exact evidence passages that were cited or used during generation. The prompt will return a structured JSON object containing a list of detected temporal hallucinations, each with a specific claim identifier, the type of temporal failure (e.g., fabricated date, unsupported recency claim, temporal scope misrepresentation), the offending text span, and a severity score. Wire this into your validation pipeline as a post-generation check. If any hallucination flag has a severity score above your threshold, block the response from reaching the user and route it for human review or regeneration with corrected temporal constraints. For high-stakes domains like financial reporting or medical summarization, always require human review when temporal hallucinations are detected, even at low severity.

Do not use this prompt as a substitute for proper retrieval date filtering or evidence freshness scoring earlier in your pipeline. It is a detection tool, not a prevention tool. If your retrieval system is serving outdated documents, this prompt will flag the resulting hallucinations, but it will not fix the root cause. Use it alongside temporal relevance scoring and stale evidence detection prompts to build a complete temporal grounding system. The next section provides the copy-ready prompt template you can adapt for your own validation harness.

This template is designed to be dropped into a hallucination detection pipeline immediately after answer generation. It forces the model to act as a strict temporal auditor, comparing every date, recency claim, and time-bound assertion in a candidate output against the provided source evidence. The prompt is structured to produce machine-readable flags rather than prose, making it suitable for automated guardrails, eval harnesses, and human-review queues. Adapt the placeholders to match your application's evidence schema, temporal sensitivity requirements, and output format.

textCopy
You are a temporal hallucination auditor. Your job is to examine a candidate model output and flag every statement that fabricates a date, claims recency without evidence support, or misrepresents the temporal scope of cited sources.

## INPUT
Candidate Output:
[OUTPUT_TEXT]

Source Evidence (with publication dates and content):
[EVIDENCE_LIST]

Query Temporal Sensitivity:
[TEMPORAL_SENSITIVITY]

## INSTRUCTIONS
1. Extract every date, time range, recency claim (e.g., "recently," "last week," "current"), and temporal comparison (e.g., "before," "after," "since") from the candidate output.
2. For each temporal claim, check whether it is supported by the source evidence. A claim is supported only if:
   - The exact date or range appears in a source published at or before that date.
   - A recency claim is backed by a source whose publication date falls within the recency window implied by the claim.
   - A temporal comparison is consistent with the dates in the cited sources.
3. Flag any temporal claim that lacks evidence support. For each flag, provide:
   - The exact text of the claim.
   - The type of hallucination: FABRICATED_DATE, UNSUPPORTED_RECENCY, TEMPORAL_MISALIGNMENT, or STALE_EVIDENCE.
   - A brief explanation of why the claim fails.
   - The closest evidence passage, if any exists.
4. If the candidate output uses a source to support a claim about a time period that the source does not cover, flag it as TEMPORAL_MISALIGNMENT.
5. If the candidate output cites a source whose publication date makes it too old for the claim, flag it as STALE_EVIDENCE.
6. If no temporal claims are found, return an empty flags list.
7. If no source evidence is provided, flag ALL temporal claims as UNSUPPORTED_RECENCY.

## OUTPUT SCHEMA
Return a JSON object with this exact structure:
{
  "temporal_claims_found": <integer>,
  "flags": [
    {
      "claim_text": "<exact text from output>",
      "flag_type": "FABRICATED_DATE | UNSUPPORTED_RECENCY | TEMPORAL_MISALIGNMENT | STALE_EVIDENCE",
      "explanation": "<why this claim is unsupported>",
      "closest_evidence": "<relevant passage or null>"
    }
  ],
  "overall_temporal_faithfulness": "FAITHFUL | PARTIALLY_FAITHFUL | UNFAITHFUL | INSUFFICIENT_EVIDENCE"
}

## CONSTRAINTS
- Do not flag claims that are explicitly hedged with uncertainty language (e.g., "may have occurred," "reportedly") unless the hedge itself is fabricated.
- Do not flag relative dates that are mathematically derivable from provided evidence dates.
- If the query temporal sensitivity is "historical" or "time-agnostic," apply looser recency standards.
- If the query temporal sensitivity is "real-time" or "recent," apply strict recency standards.
- Return ONLY the JSON object. No markdown fences, no commentary.

Adapting the template: Replace [OUTPUT_TEXT] with the candidate model response you are auditing. [EVIDENCE_LIST] should contain each retrieved passage with its publication date and full text, formatted consistently so the auditor can match claims to sources. [TEMPORAL_SENSITIVITY] should be one of real-time, recent, historical, or time-agnostic, ideally set by an upstream classification prompt. If your application uses structured evidence objects, adjust the evidence format instructions inside the prompt to match your schema. For high-risk domains such as financial reporting or clinical documentation, add a [RISK_LEVEL] placeholder and tighten the flagging thresholds for high risk.

What to do next: Wire this prompt into a post-generation validation step. If overall_temporal_faithfulness is UNFAITHFUL or PARTIALLY_FAITHFUL, route the output to a human review queue or trigger a regeneration with stricter temporal constraints. Log every flag for eval analysis—patterns in FABRICATED_DATE flags often indicate a need to improve retrieval freshness or adjust the answer-generation prompt. Avoid using this prompt as the sole defense in regulated workflows; pair it with citation verification and source date extraction prompts for a complete temporal grounding pipeline.

This prompt is designed as a post-generation guardrail, not a real-time conversational filter. It should be wired into your output validation pipeline so that every model response is checked for fabricated dates, unsupported recency claims, and temporal misrepresentations before it reaches the user. The prompt expects a model-generated text and the evidence context that was used to produce it. If your system uses RAG, pass the retrieved passages as [EVIDENCE_CONTEXT]. If your system generates answers without retrieval, you must still provide the source material the model was conditioned on, or explicitly mark the context as empty and rely on the model's internal knowledge boundary detection.

Integration pattern: Place this prompt as a validation step after answer generation and before user delivery. In a typical pipeline: (1) user query → (2) retrieval → (3) answer generation → (4) temporal hallucination detection → (5) if flags are raised, either regenerate with corrected temporal constraints, append a temporal uncertainty disclaimer, or escalate for human review depending on your [RISK_LEVEL] configuration. For high-stakes domains like financial reporting or medical summarization, always require human review when temporal hallucination flags fire. Log every detection result with the original output, evidence context, flag type, and specific temporal claims identified for audit trails and prompt improvement cycles.

Validation and retries: Parse the JSON output from this prompt and check the hallucination_flags array. If flags.length > 0, decide your remediation path based on flag severity. For severity: critical (fabricated dates, claims of recency with no evidence), block the response and regenerate with explicit temporal grounding instructions. For severity: warning (vague temporal language, minor date mismatches), you may append a disclaimer or request user clarification. Implement a maximum of 2 retries before escalating to human review to avoid infinite regeneration loops. Model choice: This prompt benefits from models with strong instruction-following and structured output capabilities. GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro are suitable. Avoid smaller models (<7B parameters) for this task unless fine-tuned on temporal verification data, as they struggle with the nuanced distinction between implied and explicit temporal claims.

Testing before deployment: Before wiring this into production, run the provided test harness against known temporal hallucination patterns: fabricated dates ('as of June 2025' when evidence is from 2023), unsupported recency claims ('recently announced' with no source date), temporal scope misrepresentation (citing a Q1 report for Q4 figures), and missing temporal qualifiers where the evidence is old but the output implies currency. Measure precision (did we flag real hallucinations?) and recall (did we miss any?). A healthy system should achieve >90% recall on critical fabrications, even if precision is lower initially. Tune the prompt's [CONSTRAINTS] section to adjust sensitivity based on your domain's tolerance for false positives versus missed hallucinations.