Changelog - 2026-03-23_02

Assure Consistent Workflow and Codify Docs — Fact-Check Pipeline & Source Catalog

Introduces a three-step LLM-powered fact-check pipeline (extract → verify → correct) and a per-section comprehensive source catalog. The mechanical-only fact-checker is upgraded to use Perplexity Sonar Pro for independent claim verification and Claude for surgical corrections.

Why

The existing fact_checker.py used regex to find claims and string matching to check if they appeared in research data. It could detect "this sentence mentions $50M" but couldn't assess whether the source actually said $50M, whether the number was about the right company, or whether a better source exists. Claims were flagged but never acted on — the 4-fact-check.json artifact was a dead end.

Investment analysts need:

1. Verified claims — not just "has a citation" but "citation supports the claim"

2. Corrections with traceability — old value → finding → new value → new source

3. Complete source lists — every source the pipeline found, including excluded ones

What Changed

New Agent: fact_verifier.py (LLM-powered verification)

Reads 4-fact-check.json, filters to suspicious/critical claims, sends batches to Perplexity Sonar Pro for independent verification.

Pipeline: fact_check → fact_verify → fact_correct
                         ↑
            Perplexity Sonar Pro
            "Is this claim accurate?"

Each claim gets tagged:

Output: 4-fact-check-verified.json + 4-fact-check-verified.md

New Agent: fact_corrector.py (surgical section updates)

Reads verified results, groups corrections by section, uses Claude to update specific sentences. Does NOT rewrite sections wholesale.

# Safety guard: reject if corrected content is wildly different length
len_ratio = len(corrected_content) / len(original_content)
if len_ratio < 0.5 or len_ratio > 2.0:
    print("Correction produced suspicious length change, skipping")

Output: 4-corrections-log.json + 4-corrections-log.md

Traceability chain:

claim found → checked for accuracy → found better source →
updated claim with citation → logged change with before/after

New Agent: source_cataloger.py (per-section source lists)

Compiles comprehensive source lists in 3-source-catalog/ by reading all pipeline artifacts:

1-research/*.md          → "found in research"
2-sections/*.md          → "included in final draft"
source-validation-log-*  → "excluded-invalid", "hallucinated"
4-fact-check-verified.*  → "verified by LLM"
4-corrections-log.*      → "added by correction"

Each source is tagged with a status:

Output: 3-source-catalog/{section}-Complete-Source-List.md + README.md

Modified: remove_invalid_sources.py (validation logging)

The cleanup gate now persists full URL validation results as source-validation-log-cleanup_sections.json. Previously, validation details were only printed to console and lost.

Updated Pipeline Sequence

... → validate_citations →
  fact_check (regex extraction) →
  fact_verify (Perplexity Sonar Pro) →
  fact_correct (Claude surgical fixes) →
  source_catalog (compile source lists) →
validate → scorecard → ...

Canonical Orchestration Document

Created context-v/reminders/Ideal-Orchestration-Agent-Workflow.md as the single source of truth for agent execution order. Documents all 31 nodes across 9 phases, known bug patterns, and the golden rules for edge ordering.

Verification (Metabologic v0.2.3 → v0.2.4)

Files Changed