Hallucinated Date and Number Detection Prompt

This prompt is designed for reporting and analytics pipelines where model-generated dates, timestamps, numeric values, and statistics must be verified against source data before they enter databases, dashboards, or downstream systems. Use it as a post-generation validation step when a model has produced a summary, report, or analysis that includes quantitative claims. The prompt instructs the model to act as an auditor, cross-referencing every date and number in the output against the provided input context, flagging unsupported values with confidence scores, and suggesting corrections when evidence exists. This is not a prompt for generating content. It is a repair and validation prompt that belongs after generation, before ingestion.

The ideal user is a data engineer, ML engineer, or backend developer responsible for an automated reporting pipeline. The required context is a pair of inputs: the original source data or document that grounds the facts, and the model-generated output that needs auditing. Without both, the prompt cannot perform its cross-reference function. The output is a structured audit report—ideally JSON—containing each flagged value, its location in the text, a confidence score for hallucination, the evidence (or lack thereof) from the source, and a suggested correction when possible. This structured output is designed to be consumed by a downstream validation harness, not read in isolation.

Do not use this prompt when the source data is unavailable, when the output contains only qualitative claims, or when the model's generation step already has perfect grounding guarantees. It is also inappropriate for real-time chat applications where latency constraints prevent a secondary validation pass. In high-risk domains such as financial reporting or clinical documentation, this prompt is a necessary but not sufficient control—always pair it with human review for flagged items above a defined severity threshold. For maximum reliability, implement this as a gating step: if the audit report returns any high-confidence hallucinations, block the output from downstream ingestion and route it for manual correction.

This prompt template is designed to be dropped directly into your prompt library or orchestration code. It accepts a block of model-generated text and the original source input, then returns a structured audit of every date, timestamp, numeric value, and statistic that cannot be traced back to the source. Use it as a post-generation validation step in reporting pipelines, analytics systems, or any workflow where fabricated temporal or quantitative data would corrupt downstream decisions.

textCopy
You are an output auditor specialized in detecting hallucinated dates, timestamps, numeric values, and statistics in AI-generated text.

Your task is to compare the [OUTPUT_TEXT] against the [SOURCE_INPUT] and identify every date, timestamp, numeric value, or quantitative claim in the output that cannot be traced to evidence in the source.

## Input

**Source Input (ground truth):**
[SOURCE_INPUT]

**Model Output (to audit):**
[OUTPUT_TEXT]

## Instructions

1. Extract every date, timestamp, numeric value, percentage, monetary amount, statistic, or quantitative claim from the model output.
2. For each extracted value, search the source input for supporting evidence.
3. Classify each value as one of:
   - **GROUNDED**: The exact value or a mathematically derivable equivalent appears in the source.
   - **PARTIALLY_GROUNDED**: The value is consistent with source data but not explicitly stated (e.g., a sum of stated figures, an inferred date range).
   - **HALLUCINATED**: The value has no traceable basis in the source input.
   - **CONTRADICTED**: The value directly conflicts with source data.
4. For each HALLUCINATED or CONTRADICTED value, provide a corrected value if the source contains evidence for one. If no correction is possible, set the corrected_value to null.
5. Assign a confidence score from 0.0 to 1.0 for each classification.

## Output Schema

Return a JSON object with this exact structure:

{
  "audit_summary": {
    "total_values_extracted": <integer>,
    "grounded_count": <integer>,
    "partially_grounded_count": <integer>,
    "hallucinated_count": <integer>,
    "contradicted_count": <integer>,
    "overall_hallucination_rate": <float 0.0-1.0>
  },
  "findings": [
    {
      "value": "<the extracted value as it appears in output>",
      "value_type": "<date | timestamp | percentage | currency | integer | float | statistic | duration | other>",
      "location_in_output": "<surrounding context or sentence fragment>",
      "classification": "<GROUNDED | PARTIALLY_GROUNDED | HALLUCINATED | CONTRADICTED>",
      "confidence": <float 0.0-1.0>,
      "source_evidence": "<quoted evidence from source input, or null if none found>",
      "corrected_value": "<correct value from source, or null>",
      "correction_rationale": "<explanation of correction, or null>"
    }
  ],
  "requires_human_review": <boolean>,
  "review_triggers": ["<list of reasons human review is recommended, empty if none>"]
}

## Constraints

- Do not flag values that are clearly marked as hypothetical, projected, or estimated in the output text if the source contains the basis for those projections.
- Treat relative date expressions ("last quarter," "next month") as values requiring source grounding against a reference date.
- If the source input is empty or contains no quantitative data, classify all extracted values as HALLUCINATED.
- If [RISK_LEVEL] is "high," set requires_human_review to true for any finding with confidence below 0.9.
- Preserve the exact formatting of extracted values in the "value" field.

## Risk Level
[RISK_LEVEL]

Adaptation notes: Replace [SOURCE_INPUT] with the original document, retrieved context, or database record that the model was supposed to work from. Replace [OUTPUT_TEXT] with the model-generated text you need to audit. Set [RISK_LEVEL] to "low", "medium", or "high" to control the human-review threshold. For high-stakes pipelines like financial reporting or clinical documentation, keep [RISK_LEVEL] at "high" and route flagged outputs to a review queue before they reach consumers. If your source input is structured data rather than free text, pre-serialize it into a text representation that preserves all quantitative fields and their relationships.

This prompt is designed to sit in a post-generation validation layer within your analytics or reporting pipeline. After a model produces output containing dates, timestamps, numeric values, or statistics, route that output—along with the original input data—through this detection prompt before the values enter any downstream database, dashboard, or user-facing report. The prompt expects two inputs: the model-generated text and the source input data that should have grounded those values. Without both inputs, the detection logic cannot function, and the prompt will correctly flag most values as unverifiable.

Integration pattern: Implement this as an asynchronous validation step. After your primary generation call completes, extract all date and numeric fields from the output, package them alongside the source context, and send both to this prompt. Parse the JSON response into a structured validation record. For each flagged value, log the confidence score, the reason code, and any corrected value the prompt suggests. If the confidence score for a hallucination flag exceeds your threshold (start with 0.7), either auto-correct the value using the prompt's suggested correction or quarantine the record for human review. Never silently drop flagged values—always preserve the original output, the flag, and the correction in an audit log so you can measure false positive rates over time.

Model choice and retries: Use a model with strong instruction-following and JSON output capabilities (GPT-4o, Claude 3.5 Sonnet, or equivalent). Set temperature=0 to minimize variance in detection results. If the prompt returns malformed JSON or missing required fields, implement a single retry with the error message appended to the prompt context. After two failures, escalate the record to a human review queue rather than looping indefinitely. For high-throughput pipelines, batch multiple output segments into a single prompt call, but keep each batch under 8,000 tokens of combined input to avoid attention dilution on individual values.

Validation and monitoring: Before deploying, run this prompt against a golden dataset of 50–100 examples containing known hallucinated dates and numbers alongside correctly grounded values. Measure precision (what fraction of flagged values were truly hallucinated) and recall (what fraction of actual hallucinations were caught). Track these metrics in production by sampling flagged outputs weekly and having a subject-matter expert review them. Common failure modes include false positives on paraphrased values (the model flags a date that is semantically equivalent but formatted differently) and missed hallucinations in nested JSON structures where the value is buried inside a larger string field. If your pipeline handles regulated data, require human approval on any auto-correction before it reaches external consumers.