A note before the work
Thank you to Northwestern's Knight Lab and the GAIN organisers for hosting this challenge. Building agentic skills that hold power accountable is exactly the kind of work the next decade of journalism needs. We're grateful to be part of it.
The submission
An agentic skill that reads a million lobbying records and finds the story.
We built data-detective — a ten-skill orchestrator for structured-records investigations — and ran it against U.S. federal lobbying disclosures. Nine verified findings. Every numeric claim cites a SQL query by SHA hash; every record has an evidence card; every external source is archived.
Deliverables
- A
Findings report
Public-facing journalism artifact. Per-finding replication paths, archived primary sources, phase-by-phase methodology walkthrough.
- B
data-detective Agent Skill
Orchestrator + nine sub-skills. All ten validate against the agentskills.io spec. Portable across structured-records corpora.
- C
Audit trail
Every detector output, every evidence card, every adversarial fact-check verdict, every Spotlight OSINT investigation output. The full audit chain for replication.
- D
Interaction traces
Session structure and human-judgment moments. Brief approval, methodology approval, Gate 1 review, finishing-path choice, OSINT lead curation.
Method
Deterministic where possible. LLM only over already-filtered candidates.
- Full corpus ingest
- ~4 min pyarrow bulk insert · 60× speedup
- Entity resolution
- 99.9% deterministic bridge keys
- LLM tokens on raw documents
- 0 exploration is SQL-only
- SQL-only detectors
- 12 SHA-tagged per query
- Agent Skills
- 10 orchestrator + 9 sub-skills
- Findings verified · OSINT promoted
- 9 / 6 spotlight-handoff cycles