Northwestern University · GAIN Agentic AI Investigative Journalism Challenge
A single Agent Skill orchestrates a full investigation against a million U.S. lobbying records
and produces nine verified findings, every claim grounded in a primary source.
Findings · Report · Replication · Audit chain · Interaction traces · Acknowledgments
/data-detective
That's the launcher. The orchestrator takes the brief, designs a methodology, spawns subagents, invokes nine sub-skills across seven phases, runs an adversarial fact-check, gates on user approval, and produces three deliverables.
- 📄 Findings report:
report.html— public-facing artifact ·findings-report.md— narrative audit - 📋 Audit ledger:
evidence-map.json— every claim → cards → query hashes → URLs - 🔁 Full audit chain:
case-trace/— anomalies, evidence cards, fact-check verdicts, OSINT cycles - 🌐 Web landing: https://buriedsignals.com/gain-2026
The data-detective package is one self-contained unit. Mirrors the Spotlight layout convention: agents at the package root, all skills as siblings under skills/ (orchestrator named after the package), references + pipeline driver at the package root.
data-detective/ ← drop this in your agent's skills dir
├── agents/
│ ├── investigator.md ← PLANNING + EXECUTION subagent
│ └── fact-checker.md ← adversarial verifier subagent
├── skills/ ← all skills siblings; runtime resolves by bare name
│ ├── data-detective/SKILL.md ← orchestrator (the launcher)
│ ├── ingest/SKILL.md ← P1 · corpus → DuckDB
│ ├── resolve/SKILL.md ← P2 · entity resolution
│ ├── detect/SKILL.md ← P3 · detector catalog (D1-D12)
│ ├── evidence-cards/SKILL.md ← P3 + P5 · record → audit card
│ ├── external-data/SKILL.md ← P3 · FARA, Congress.gov, USAspending
│ ├── similarity-search/SKILL.md ← P3 · vector NN (optional)
│ ├── report-drafting/SKILL.md ← P5 · synthesis (direct path)
│ ├── spotlight-handoff/SKILL.md ← P6 · → Spotlight (handover path)
│ └── vault-ingest/SKILL.md ← P7 · Obsidian write-back
├── references/ ← orchestrator-level docs
└── run_pipeline.sh ← end-to-end pipeline driver (CWD-agnostic)
Every sub-skill validates against agentskills.io spec: name matches dir, description ≤ 1024 chars, body < 500 lines.
| Phase | What | Sub-skill invoked | Subagent spawned |
|---|---|---|---|
| P0 preflight | Skill discovery, profile pick | — | — |
| P1 ingest | Profile-driven ETL to DuckDB | ingest |
— |
| P2 resolve + methodology | Bridge-key entity resolution; investigator plans | resolve |
investigator (PLANNING) |
| P3 cycles + fact-check | Detectors + drill + external joins + evidence cards + adversarial verification | detect, evidence-cards, external-data, similarity-search |
investigator (EXECUTION), fact-checker (adversarial) |
| P4 Gate 1 | User reviews findings, picks finishing path | — | — |
| P5 synthesis (direct path) | report.html + findings-report.md + evidence-map.json | report-drafting |
— |
| P6 spotlight-handoff (handover path) | Hand chosen leads to Spotlight for OSINT amplification | spotlight-handoff |
spotlight:investigator, spotlight:fact-checker (external) |
| P7 vault-ingest | Obsidian write-back, --sensitive aware |
vault-ingest |
— |
Gate 1 is binary: direct report (data-detective drafts the final report) or spotlight handover (Spotlight runs its own pipeline and produces the report on its end, in hybrid mode that walks both orchestrators' phases). Both paths converge at P7.
data-detective reuses Spotlight's orchestration discipline (brief → methodology → cycles → fact-check → gate → ingest) but on a different medium: structured records instead of web OSINT. At Phase 6 handover, the two skills compose — Spotlight's web-side personas pick up where the data-side personas leave off, with the upstream audit chain preserved as context.
Nine verified findings. Eight high-confidence, one medium. Five were promoted to data-detective from the six formal Spotlight handoff cycles (OS-001 through OS-006; OS-006 updated C-004), and five of the handoffs surfaced material new evidence. Corrections passes ran 2026-05-25 and 2026-07-09 — see the Corrections log at the end of findings-report.md.
- C-001 — A self-styled "sovereign government" filed $180M of LDA records via Rev. Dr. Christina Loren Clement in 2024-2025
- C-002 — Anchor & Arrow (HASC / Sen. Cotton / Navy-comptroller alumni) signed 13 AI-defense startups; client Apex Technology won a $22M Air Force STRATFI during the engagement
- C-003 — 682 LDA registrants disclosing foreign entities have no FARA counterpart; D10 narrows the gap to 200 named candidates
- C-004 — Brownstein's Senate-Finance alumni lobby Apollo while Warren and Grassley attack it from that committee; Apollo followed Mark Warren to Capitol Tax Partners in June 2026
- C-005 — The FEC flagged Clement's $50.25M campaign self-loan (RFAI, April 28, 2026) and moved it to its false/fictitious repository; Quinnipiac Prof. John Martin on record
- C-006 — Akin Gump's Adani retainer is effective the day after the EDNY indictment was unsealed (LD-1 effective 11/21/2024); DOJ's move to drop the case awaits court approval
- C-007 — Vertex Pharmaceuticals (TDY client) sued HHS-OIG, lost at district court, appeal pending; TDY founders were Part D architects at HHS
- C-008 — $5M RAI Services → MAGA Inc (C00892471) → FDA flavored-vape guidance timeline, via McFaul/Ballard
- C-009 — Lazarski (Cornerstone) extends C-004's pattern to Senate Armed Services via Apollo portfolio company Arconic
Full claim-by-claim narrative + replication paths in report.html.
# 1. Clone
git clone https://github.com/buriedsignals/gain-2026.git
cd gain-2026
# 2. Install — copies all 10 skills + 2 agent personas into ~/.claude/
./install.sh
# 3. Place the corpus where the profile expects it
# (see data-detective/skills/ingest/PROFILE-FORMAT.md).
# For this run: U.S. Senate LDA JSON 2008-2025, House LDA XML 2008-2025,
# congressional press release JSONL.
# 4. Launch in a fresh Claude Code session
/data-detectiveThe installer mirrors what Spotlight's marketplace handler does at install time: flat-copies each sub-skill from data-detective/skills/<name>/ to ~/.claude/skills/<name>/ so the runtime registers each individually, and copies the agent personas to ~/.claude/agents/. Override destinations with CLAUDE_SKILLS_DIR=... CLAUDE_AGENTS_DIR=... ./install.sh for non-Claude-Code harnesses.
The orchestrator handles the rest: preflight, brief, methodology, cycles, Gate 1, finishing path.
- Submitter: Tom Vaillant · Buried Signals · tom@buriedsignals.com
- Repository: https://github.com/buriedsignals/gain-2026 (will be tagged
final-submissionat the submission deadline) - License: MIT (skills) · CC-BY-4.0 (findings + report)
- Conflicts of interest: None. No relationship to any registrant, client, member, or lobbyist named in findings.
- Future development: This repository is frozen. Productized
data-detectiveevolves at github.com/buriedsignals/skills.
See Acknowledgments for the full categorized list of data providers, trackers, and tooling this investigation stands on.
The human-judgment moments in the run (brief approval, methodology approval, Gate 1, finishing-path choice, OSINT lead curation, post-publication verification pushback) are documented in interaction-traces/ — see interaction-traces/README.md for the manifest mapping each transcript to the findings it produced. The directory contains the full raw Claude Code session transcripts in JSONL form: the main investigation (01-data-detective-investigation-20260522.jsonl, 9.7M), the post-publication verification + corrections pass (02-verification-and-corrections-20260525.jsonl, 1.7M), and five parallel spotlight:fact-checker subagent traces (one per finding cluster).
This investigation stands on public-interest data infrastructure and open-source tooling. A genuine thank-you to every provider below — the findings simply would not exist without this work. (Listing here does not imply affiliation or endorsement; always respect each provider's terms of service.)
| Category | Projects & services we're grateful to |
|---|---|
| Challenge & corpus | Northwestern University — GAIN (Generative AI in the Newsroom) Agentic AI Investigative Journalism Challenge — Senate LDA JSON, House LDA XML, and congressional press-release corpus |
| Federal primary sources | Senate LDA database + REST API · House Clerk lobbying disclosures · DOJ FARA eFile · FEC / openFEC API · FEC docquery · USAspending.gov API · Congress.gov · govinfo |
| Open-data mirrors & trackers | OpenSanctions (DOJ FARA bulk mirror) · unitedstates/congress-legislators · OpenSecrets · LegiStorm |
| Court records | Free Law Project — CourtListener / RECAP · PacerMonitor |
| Corporate registries | Delaware iCIS · Florida Sunbiz |
| Journalism cited | The New York Times · ProPublica · Law360 · The Hill · Boston Globe / Boston Herald · KFF Health News · Politico · Reuters |
| Tooling & runtime | Anthropic — Claude Code (Agent Skills) · DuckDB · PyArrow · fastembed + DuckDB VSS · Firecrawl · Internet Archive Wayback Machine · Astral — uv · Spotlight (Buried Signals) |