verification.json Fields

Complete reference for every field in verification.json — the deterministic verification report produced by the extract and upgrade commands. No LLM is involved in generating this file; it is pure string matching and arithmetic against the source bill text.

Top-Level Structure

{
  "amount_checks": [ ... ],
  "raw_text_checks": [ ... ],
  "arithmetic_checks": [ ... ],
  "completeness": { ... },
  "summary": { ... }
}

Amount Checks (amount_checks)

One entry for each provision that has a text_as_written dollar string. Checks whether that exact string exists in the source bill text.

Status Values

Example

{
  "provision_index": 0,
  "text_as_written": "$2,285,513,000",
  "found_in_source": true,
  "source_positions": [431],
  "status": "verified"
}

Counts in Example Data

Raw Text Checks (raw_text_checks)

One entry per provision. Checks whether the provision’s raw_text excerpt is a substring of the source bill text, using tiered matching.

Match Tiers

Example

{
  "provision_index": 0,
  "raw_text_preview": "For an additional amount for ''Compensation and Pensions'', $2,285,513,000, to r",
  "is_verbatim_substring": true,
  "match_tier": "exact",
  "found_at_position": 371
}

Arithmetic Checks (arithmetic_checks)

Group-level sum verification — checks whether line items within a section or title sum to a stated total.

Note: This field is deprecated in newer extraction files. It may be absent or empty. When present, it uses this structure:

Old files that include this field still load correctly. New extractions and upgrades omit it.

Completeness (completeness)

Checks whether every dollar-sign pattern in the source bill text is accounted for by at least one extracted provision.

UnaccountedAmount

Each entry represents a dollar string found in the source text that wasn’t matched to any extracted provision:

Example

{
  "total_dollar_amounts_in_text": 2,
  "accounted_for": 2,
  "unaccounted": []
}

For a bill with unaccounted amounts:

{
  "total_dollar_amounts_in_text": 1734,
  "accounted_for": 1634,
  "unaccounted": [
    {
      "text": "$500,000",
      "value": 500000,
      "position": 45023,
      "context": "pursuant to section 502(b) of the Agricultural Credit Act, $500,000 for each State"
    }
  ]
}

The unaccounted amounts are typically statutory cross-references, loan guarantee ceilings, struck amounts in amendments, or prior-year references in CRs. See What Coverage Means (and Doesn’t) for detailed interpretation.

Coverage Calculation

Coverage = (accounted_for / total_dollar_amounts_in_text) × 100%

Verification Summary (summary)

Roll-up metrics for the entire bill — these are the numbers displayed by the audit command.

Example (H.R. 9468)

{
  "total_provisions": 7,
  "amounts_verified": 2,
  "amounts_not_found": 0,
  "amounts_ambiguous": 0,
  "raw_text_exact": 5,
  "raw_text_normalized": 0,
  "raw_text_spaceless": 0,
  "raw_text_no_match": 2,
  "completeness_pct": 100.0,
  "provisions_by_detail_level": {
    "top_level": 2
  }
}

Mapping to Audit Table Columns

How verification.json Is Used

By the audit command

The audit command reads verification.json for each bill and renders the summary metrics as the audit table.

By the search command

Search uses verification data to populate these output fields:

By the summary command

The summary footer (“0 dollar amounts unverified across all bills”) counts the total amounts_not_found across all loaded bills.

When verification.json Is Generated

• By extract: Automatically after LLM extraction completes. Verification runs against the source XML with no LLM involvement.
• By upgrade: Re-generated when upgrading extraction data to a new schema version. The source XML must be present in the bill directory for verification to run.

If the source XML (BILLS-*.xml) is not present, verification is skipped and verification.json is not created or updated.

Accessing verification.json

From the CLI

You don’t need to read this file directly — the audit and search commands surface its data in user-friendly formats.

From Python

import json

with open("data/118-hr9468/verification.json") as f:
    v = json.load(f)

# Summary metrics
print(f"Not found: {v['summary']['amounts_not_found']}")
print(f"Coverage: {v['summary']['completeness_pct']:.1f}%")
print(f"Exact text matches: {v['summary']['raw_text_exact']}")

# Check individual provisions
for check in v["amount_checks"]:
    if check["status"] == "not_found":
        print(f"WARNING: Provision {check['provision_index']}: {check['text_as_written']} not found in source")

# See unaccounted dollar amounts
for ua in v["completeness"]["unaccounted"]:
    print(f"Unaccounted: {ua['text']} at position {ua['position']}")
    print(f"  Context: {ua['context']}")

Related References

• How Verification Works — detailed explanation of the three verification checks
• What Coverage Means (and Doesn’t) — interpreting the completeness metric
• Verify Extraction Accuracy — practical guide for running and interpreting the audit
• extraction.json Fields — the extraction data that verification checks against