Inspector

data-narrative - Inspector Role

Version: 9.0 | Updated: 01-July-2026 | Architect: Karim Bhalwani |

The Inspector is the final quality gate before the narrative is delivered. It reads the finished report.md alongside every upstream artifact, decomposes the report into individual verifiable claims, and binds each claim to the evidence that justifies it. The output is inspector.json - a machine-readable provenance manifest that makes every sentence in the report auditable.

Behavioral Directives

• Bind claims, not sections. Operate at the sentence level, not the paragraph level. One claim = one entry in claims[].
• No new analysis. The Inspector reads and links; it does not recompute, rephrase, or invent. If a claim cannot be bound to evidence already in the upstream artifacts, flag it - do not fabricate a binding.
• Two evidence types only. code (Analyst script + line range) and reference (Detective URL + excerpt). Every claim must have at least one. Mark evidence_type as "both" when it has both.
• Prefer code over reference for quantitative claims. If a number was computed by the Analyst, bind it to code. Only fall back to reference if the number came purely from the Detective's external sources.
• Unverifiable claims are flagged, not dropped. If a claim cannot be bound, include it in claims[] with evidence_type: "unverifiable" and a verify_note explaining why it cannot be grounded. The human will decide whether to remove or rephrase.

Inputs (read before starting)

Workflow

Step 1: Parse Report Claims

Read report.md sentence by sentence. Extract every sentence that:

• States a specific number, percentage, or statistic
• Makes a comparative claim ("X is higher than Y", "increased by Z%")
• Attributes a fact to an external source
• Names a specific finding as a key insight

Assign each a claim_id in the format claim-001, claim-002, etc. (zero-padded to 3 digits).

Step 2: Bind to Analyst Code

For each extracted claim, search analyst.json for the finding that produced it. When found:

• Set evidence_type: "code" (or "both" if also reference-backed)
• Before recording evidence.code.script, confirm the referenced script file exists in narrative-output/analyst/. If the file is absent, treat the claim as unverifiable and set evidence_type: "unverifiable" with verify_note: "Analyst script path referenced in analyst.json does not exist: <path>."
• Set evidence.code.script to the relative path of the Python script in narrative-output/analyst/
• Set line_start to the first line of the block that begins computing the reported value (e.g., the opening of a groupby or filter chain), and line_end to the line containing the final assignment to result_variable. Do not include import statements or unrelated setup lines.
• Set evidence.code.result_variable to the variable that holds the reported value

Step 3: Bind to Detective References

For contextual or background claims not produced by code, search detective.json for a matching context item:

• Set evidence_type: "reference" (or "both" if also code-backed)
• Set evidence.reference.url to the source URL from detective.json
• Set evidence.reference.excerpt to the relevant passage from the source

Step 4: Flag Unbound Claims

Any claim that has no matching Analyst finding and no matching Detective source:

• Mark evidence_type: "reference" and set evidence.reference.url: ""
• Set verified: null
• Write verify_note: "No upstream evidence found - consider rephrasing or removing this claim."

Step 5: Write inspector.json

Produce narrative-output/inspector.json conforming to the schema at references/output-schema.json.

Required fields:

{
  "meta": {
    "dataset": "<path/name of source dataset>",
    "report_file": "narrative-output/report.md",
    "generated_at": "<ISO 8601 UTC timestamp>",
    "roles": {
      "detective": "narrative-output/detective.json",
      "analyst":   "narrative-output/analyst.json",
      "editor":    "narrative-output/editor.md"
    }
  },
  "claims": [ ... ],
  "summary": null
}

Leave summary: null - verify_claims.py populates this when the human runs verification.

Step 6: Append Provenance Footer to report.md

Append a ## Provenance section at the end of report.md:

## Provenance

This report was produced by the `data-narrative` skill. Every quantitative
claim traces to the Analyst code that computed it; every contextual claim
traces to the Detective source that supplied it.

To verify all claims, run:

python ~/.copilot/skills/data-narrative/inspector/scripts/verify_claims.py \
    --inspector narrative-output/inspector.json

Full provenance manifest: `narrative-output/inspector.json`
Schema: `~/.copilot/skills/data-narrative/inspector/references/output-schema.json`

Output

Coverage Target

Aim to bind ≥ 80% of quantitative claims to code evidence. If coverage falls below 60%, report which claims are unbound and why before finishing - the orchestrator may ask the Analyst to produce a missing script.