Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Investigate datasets and resolve entities using OpenPlanter methodology — cross-reference heterogeneous sources, build Admiralty/ACH confidence-tiered evidence chains, and apply OSINT tradecraft. Triggers on entity resolution, cross-reference datasets, evidence chain, OSINT investigation, structured investigation.
argument-hint
[workspace-path or investigation-query]
OpenPlanter — Investigation Methodology Skill
Epistemic framework for cross-dataset investigation, entity resolution, and evidence-backed analysis. Extracted from the OpenPlanter recursive investigation agent and enriched with professional OSINT tradecraft (Admiralty System, ACH, FollowTheMoney schema, intelligence cycle methodology).
Claude Code already has the tools. This skill provides the methodology.
Any analysis where claims need to trace to cited source records
Quick Start
# 1. Initialize workspace
python3 ~/.claude/skills/openplanter/scripts/init_workspace.py /path/to/investigation
# 2. Drop datasets into datasets/cp campaign_finance.csv lobbying.json corporate_registry.csv /path/to/investigation/datasets/
# 3. Write an investigation plan
python3 ~/.claude/skills/openplanter/scripts/entity_resolver.py /path/to/investigation
python3 ~/.claude/skills/openplanter/scripts/cross_reference.py /path/to/investigation
python3 ~/.claude/skills/openplanter/scripts/evidence_chain.py /path/to/investigation
python3 ~/.claude/skills/openplanter/scripts/confidence_scorer.py /path/to/investigation
# → plans/plan.md (see references/output-templates.md for plan template)
# 4. Resolve entities across datasets
# 5. Cross-reference linked entities
# 6. Validate evidence chains
# 7. Score confidence
Investigation Methodology
Epistemic Discipline
Assume nothing about the environment until confirmed firsthand. These principles prevent the most common investigation failures:
Ground truth comes from files, not memory. Read actual data before modifying it, and read actual error messages before diagnosing. Model memory of data structure is unreliable—reading the file takes seconds, recovering from a wrong assumption takes minutes.
Empty output is ambiguous. If a command returns empty, cross-check with ls -la and wc -c before concluding a file is actually empty, because output capture mechanisms can silently lose data.
Success does not mean correctness. A command that "succeeds" may have done nothing. Check actual outcomes, not just exit codes. After downloading, verify with ls and wc -c. After extraction, verify expected files exist.
Verify round-trip correctness. After any data transformation (parsing, linking, aggregation), check the result from the consumer's perspective—load the output, spot-check records, verify row counts. Transformations that silently drop records are the most common source of wrong conclusions.
Three failures = wrong approach. If a command fails 3 times, change strategy entirely. Repeating an identical command expecting different results wastes context window.
Produce artifacts early. Write a working first draft of findings as soon as the requirements are clear, then iterate. An imperfect deliverable beats a perfect analysis with no output. If 3+ steps have passed without writing any output, stop and write—even if incomplete.
For the full epistemic framework (including data ingestion rules and hard rules from OpenPlanter's prompts.py), see references/investigation-methodology.md.
Adapted from OpenPlanter's prompts.py, enriched with Middesk, OpenSanctions, and ICIJ patterns.
Normalize — Apply canonical key transformation: Unicode NFKD + diacritic stripping, case folding, legal suffix canonicalization (LLC/Inc/Corp/Ltd → canonical forms), punctuation removal, ampersand normalization (& → and). See references/entity-resolution-patterns.md for complete normalization tables and suffix maps.
Block — Reduce O(N^2) comparisons using blocking keys: first 3 characters of normalized name + state/jurisdiction, phonetic key (Soundex/Double Metaphone), token overlap via inverted index.
Compare — Pairwise similarity with cascading checks: real_quick_ratio() → quick_ratio() → ratio(). Use autojunk=False for entity names because strings under 200 characters produce false negatives with junk heuristics. Include token set comparison for word-order variants ("Apple Inc" vs "Inc Apple").
Score — Multi-signal weighted model:
Hard signals: TIN/EIN exact match (1.0), identical phone E.164 (0.8), identical email (0.8)
Soft signals: name similarity (0.5), address fuzzy (0.2), state match (0.1)
Hard disqualifiers: TIN mismatch when both present (-0.5), country mismatch (-0.5)
Cluster — Group via transitive closure using Union-Find. Gate closure so all pairwise scores exceed threshold—this prevents chain errors where A≈B and B≈C but A≉C. Exclude registered agent addresses from triggering transitive closure alone, because thousands of entities share the same registered agent.
Review — Flag by confidence band:
Score >= 0.85: auto-match (confirmed)
Score 0.70-0.84: queue for review (probable)
Score 0.55-0.69: include in wide net (possible)
Score < 0.55: discard
Evidence Chain Construction
Every claim traces to a specific record in a specific dataset. This is what separates investigation from speculation.
Distinguish direct evidence (A appears in record X), circumstantial evidence (A's address matches B's address), and absence of evidence (no disclosure found)
Document every hop in a multi-step chain with source record, linking field, and match quality
Link strength = weakest criterion in the chain (a chain is only as strong as its weakest link)
Track source lineage to detect circular reporting—Source B citing Source A is not independent corroboration
Confidence Tiers
Based on the Admiralty System (NATO AJP-2.1), adapted for dataset investigation:
Tier
Criteria
Required Evidence
Admiralty Equivalent
Confirmed
2+ independent sources with different collection paths; hard signal match (EIN, phone); or official record with verifiable provenance
Independent corroboration required
A1–B1
Probable
Strong single source (official record); high fuzzy match (>0.85) on name + address + state; consistent with known patterns
Single strong source acceptable
B2–C2
Possible
Circumstantial evidence only; moderate fuzzy match (0.55-0.84); consistent but not yet corroborated; requires additional investigation
Hypothesis supported but not confirmed
C3–D3
Unresolved
Contradictory evidence; insufficient data; single weak source; or unable to verify
Cannot determine with available evidence
D4–F6
Sherman Kent probability mapping:
Confirmed ≈ "almost certain" (93-99%)
Probable ≈ "likely" (75-85%)
Possible ≈ "chances about even" (45-55%)
Unresolved ≈ insufficient basis for estimate
Verification Principle
Implementation and verification must be uncorrelated. An agent that performs an analysis introduces systematic bias when self-verifying—it "knows" what the answer should be and unconsciously confirms it. Use the implement-then-verify pattern:
Step 1: Perform entity resolution and cross-referencing (analysis agent)
Step 2: Read the result files
Step 3: Independent verification (separate agent or separate pass with no shared context from step 1):
- Load output files fresh
- Spot-check N random records against raw source data
- Verify row counts match expectations
- Run validation script
- Report raw output only
The verification executor has no context from the analysis executor. It runs commands and reports output, making its evidence independent.
Anti-bias checks:
Confirmation bias: Score hypotheses by inconsistency count (ACH), not confirmation count, because disconfirming evidence is more diagnostic
Anchoring: Do not rank hypotheses until evidence collection is complete
Circular reporting: Track source lineage; verify independence of collection paths before counting corroborations
Satisficing: Require minimum 3 competing hypotheses before scoring—this prevents premature commitment to the first plausible explanation
delegate_to_rlm.py → full OpenPlanter agent (provider-agnostic, session-resumable)
One-command pipeline for any mode: investigate.py /path/to/workspace --phases all
RLM Delegation — Provider-Agnostic
The RLM agent auto-detects the LLM provider from the model name. Works with any provider the agent supports:
# Anthropic (default)
python3 scripts/delegate_to_rlm.py --objective "..." --workspace DIR --model claude-sonnet-4-5-20250929
# OpenAI
python3 scripts/delegate_to_rlm.py --objective "..." --workspace DIR --model gpt-4o
# OpenRouter (any model via slash routing)
python3 scripts/delegate_to_rlm.py --objective "..." --workspace DIR --model anthropic/claude-sonnet-4-5
# Ollama (local inference, air-gapped)
python3 scripts/delegate_to_rlm.py --objective "..." --workspace DIR --provider ollama --model llama3
# Cerebras (model name doesn't contain "cerebras", so specify --provider)
python3 scripts/delegate_to_rlm.py --objective "..." --workspace DIR --model qwen-3-235b-a22b-instruct-2507 --provider cerebras
# Resume a previous investigation session
python3 scripts/delegate_to_rlm.py --resume abc123 --workspace DIR
# List saved sessions
python3 scripts/delegate_to_rlm.py --list-sessions --workspace DIR
# Control reasoning depth
python3 scripts/delegate_to_rlm.py --objective "..." --workspace DIR --reasoning-effort high
# List available models for a provider
python3 scripts/delegate_to_rlm.py --list-models --provider ollama
Provider auto-detection: claude-* → anthropic, gpt-*/o1-*/o3-* → openai, org/model → openrouter, *cerebras* → cerebras, llama*/qwen*/mistral*/gemma* → ollama. For models without a recognizable prefix, pass --provider explicitly.
API keys pass through environment variables: ANTHROPIC_API_KEY, OPENAI_API_KEY, OPENROUTER_API_KEY, CEREBRAS_API_KEY (or OPENPLANTER_-prefixed variants). Ollama requires no API key. Set OPENPLANTER_REPO to override local clone discovery.
Scripts Reference
All scripts use Python stdlib only. Zero external dependencies. External tools (Exa, Firecrawl, OpenPlanter agent) are invoked as subprocesses.
Individual scripts for targeted government and public data sources. All use Python stdlib only, produce JSON + provenance sidecar, and support --dry-run and --list.
Script
Data Source
Auth
Key Linking Fields
fetch_census.py
US Census Bureau ACS 5-Year
Optional CENSUS_API_KEY
Geography (state, county, ZIP)
fetch_epa.py
EPA ECHO Facility Search
None
registry_id, lat/lon, SIC/NAICS
fetch_icij.py
ICIJ Offshore Leaks Database
None
icij_id, entity name, jurisdiction
fetch_osha.py
OSHA DOL Enforcement
None
activity_nr, estab_name, SIC
fetch_propublica990.py
ProPublica Nonprofit Explorer (IRS 990)
None
ein, org name, NTEE code
fetch_sam.py
SAM.gov Entity Registration
SAM_GOV_API_KEY
UEI, CAGE code, NAICS
Usage pattern (all fetchers follow the same interface):
fetch_census.py queries Census ACS 5-Year estimates
Knowledge graph query
Built-in
wiki_graph_query.py reads OpenPlanter wiki graphs
Local RAG over large document corpora
rlama
Create collection from datasets/, query semantically
Parallel investigation threads
minoan-swarm
Elat Research Swarm with domain-split investigators
Academic/legal research
academic-research
Case law, regulatory filings, citations
Twitter/social media OSINT
twitter
x-search for entity mentions, bird for profile data
Daimonic timeline curation
worldwarwatcher-update
Mazkir ha-Milḥamat entity resolution for non-military domains
US Public Records Datasets
Key datasets and their linking keys for cross-reference investigations:
Dataset
Access
Linking Key
Script
Format
FEC Campaign Finance
api.open.fec.gov + bulk CSV
committee_id, contributor name
dataset_fetcher.py
CSV, JSON API
Senate LDA Lobbying
lda.senate.gov/api
Registrant name, Client name
dataset_fetcher.py
JSON API, XML
SEC EDGAR
data.sec.gov + EFTS search
CIK (Central Index Key)
dataset_fetcher.py
JSON, XBRL
SAM.gov Entity Registration
api.sam.gov
UEI, CAGE code, NAICS
fetch_sam.py
JSON API
EPA ECHO Facilities
echodata.epa.gov
FRS Registry ID, lat/lon
fetch_epa.py
JSON API
ProPublica 990 (IRS)
projects.propublica.org
EIN, org name
fetch_propublica990.py
JSON API
ICIJ Offshore Leaks
offshoreleaks.icij.org
Node ID, entity name, jurisdiction
fetch_icij.py
JSON API
OSHA Inspections
enforcedata.dol.gov
Activity number, SIC code
fetch_osha.py
JSON API
US Census ACS
api.census.gov
State/county/ZIP FIPS
fetch_census.py
JSON API
OFAC Sanctions
treasury.gov/ofac (or OpenSanctions)
Name + aliases, identifiers
dataset_fetcher.py
CSV, XML
State Corporate Registries
Per-state (or OpenCorporates API)
State registration number
—
Varies
Property Records
County-level (or ATTOM/CoreLogic)
Parcel ID (APN), owner name
—
CSV, shapefile
Cross-dataset linking challenge: No universal corporate ID exists in US public records. The standard approach: normalize names, fuzzy match, filter by jurisdiction/address, then anchor on known IDs (CIK, committee_id, EIN, UEI, CAGE, FRS ID) when available.
Multi-Agent Investigation
For complex investigations requiring parallel workstreams, use minoan-swarm. Separate the verifier agent from analysis agents to maintain uncorrelated verification—the verifier receives only output files and verification criteria, with no shared context from the analysis phase.
See references/investigation-methodology.md for the full swarm role template (keret, kothar, resheph, anat, shapash).
Structured Analytical Techniques
For complex scenarios with multiple possible explanations, apply Analysis of Competing Hypotheses (ACH) and Key Assumptions Check. These techniques are detailed in references/investigation-methodology.md.
ACH summary: Build a matrix of hypotheses vs. evidence. Score by inconsistency count (fewest I markers wins), not confirmation count. This counteracts confirmation bias because disconfirming evidence is more diagnostic than supporting evidence. Identify linchpin evidence whose reclassification would change the conclusion.
Deep References
references/investigation-methodology.md — Full epistemic framework, ACH procedure, Key Assumptions Check, multi-agent swarm template
