data-provenance-audit

Use this skill to ensure every material number in a stock analysis is sourced, derived from sourced inputs, user-provided, or explicitly labeled as an assumption. Trigger for valuations, source checks

SKILL.md

GitHub

Data Provenance Audit Skill

Purpose

Ensure every number used in investment analysis is traceable to a human-verifiable source.

Trigger Examples

  • "Audit my AAPL valuation note for orphan numbers."
  • "Check whether every number in this stock analysis has a source."
  • "Verify data provenance before using these valuation inputs."
  • "Find unsourced revenue, FCF, debt, share count, or price assumptions."

Trigger Conditions

Activate whenever:

  • A valuation model is run
  • Fundamental metrics are calculated
  • A report includes current price, revenue, earnings, cash flow, debt, shares, or margins
  • A scorecard uses numerical inputs
  • User asks to verify sources or prevent AI-fabricated data

Required Inputs

  • Text, table, report, model input packet, or list of numbers to audit
  • Source metadata where available: document, URL/path, filing date, period, provider, and confidence
  • Formulas for derived metrics, or enough input IDs to reconstruct them
  • User-provided assumptions explicitly marked as assumptions

If the user asks to audit a note but does not provide the note, ask for it or return a blocked audit with the missing input.

Hard Rule

No orphan numbers.

Every number must be one of:

  • Sourced raw data
  • Derived data with formula and sourced inputs
  • Explicit model assumption
  • Explicit user input

Unsourced numbers must not drive valuation.

Workflow

  1. Collect all numerical inputs.
  2. Assign data IDs.
  3. Attach source metadata.
  4. Classify raw / derived / assumption / user input.
  5. Check freshness.
  6. Check confidence.
  7. Reconcile conflicts across sources.
  8. Produce Source Annotation Table.
  9. Block valuation if critical source data is missing.

Script Mapping

  • Provenance implementation: scripts/audit/data_provenance_audit.py
  • Input packet schema: schemas/valuation_input_packet.schema.json
  • Data provenance policy: references/data_source_policy/data_provenance_policy.md

Minimal Output Contract

Return sourced raw data, derived metrics, assumptions/user inputs, and blockers. If a critical number is unsourced, say it cannot drive valuation or action guidance.

Output

## Data Provenance Audit

### Source Annotation Table

| Data ID | Metric | Value | Period | Source | URL / Path | Confidence |
|---|---:|---:|---|---|---|---:|

### Derived Metric Table

| Data ID | Metric | Formula | Input IDs | Result |
|---|---|---|---|---:|

### Unverified / Assumption Table

| Item | Value | Reason | Impact |
|---|---:|---|---|

### Blockers

- Missing source:
- Conflicting source:
- Stale source: