Investigate with evidence you can revisit
12 skills that help AI assistants examine sources, challenge hypotheses and write reports linked to evidence. For analysts, investigators, journalists and researchers. Free and open source, under Apache-2.0.
A skill combines specialized instructions and local tools. You choose the AI client and model; the Python tools preserve selected files in a portable case.
flowchart LR
A[Documents and sources] --> B[Analysis and contradictions]
B --> C[Referenced report]
Your inputs: selected documents, source captures, logs or browser-history exports. Your output: a case with preserved evidence, reasoned findings and explicit unknowns, ready for review and handoff.
See a concrete result
Do three articles confirm the same event independently? In the synthetic Dock 7 case, all three repeat the Harbor Bulletin. The event's exact time remains unconfirmed.
| Report preview · synthetic example, summarized | Evidence |
|---|---|
| Observed: all three articles attribute the announcement to the same bulletin. | E000001 · E000002 · E000003 |
| Counterpoint to check: a status capture says the dock is open; another note says closure has not yet been observed. | E000004, event log |
| Unknown: whether closure began at 09:00 UTC. The announced time has no timezone. | C000002, unconfirmed, low confidence |
The demo records these claims and origin labels explicitly; it does not ask a model to discover them. Different domains or hashes alone do not prove source independence. Explore the corpus.
Start in five minutes
From the source repository root, with Python 3.11+, run the synthetic demo. No account, API key or additional Python package is needed.
Linux / WSL
mkdir -p cases
python3 scripts/demo.py --out cases/demo
Windows / PowerShell
New-Item -ItemType Directory -Force .\cases | Out-Null
py -3 scripts/demo.py --out .\cases\demo
Use a new output directory each time. Open
cases/demo/case/reports/report.en.md: the demo registers 7 pieces of
evidence, 2 claims and 2 uncertain timestamps, then verifies a private
export and restores it. It uses no network or model.
Follow the walkthrough · Using an extracted package?
Install in your AI assistant
Build the packages from the source repository with
python3 scripts/package.py (py -3 scripts/package.py on Windows).
Then choose your client:
| Client | Installation guide |
|---|---|
| Claude Code | Native plugin |
| Codex | Native plugin |
| OpenCode | Standalone skills |
| Gemini CLI | Standalone skills or native skill installation |
Install generated packages from dist/; they contain the required runtime.
Use one installation route per client and scope. Actual discovery and model
behavior remain untested; see validation status.
Once installed, an example request is:
Use
of-case-starton these supplied documents, without network requests. Check whether the articles provide independent confirmation, examine contradictions and prepare a referenced report. Keep unknowns explicit.
The 12 skills
Choose a stage, then the skill that matches your question.
Scope
- Case start: question, scope and evidence inventory
Check sources
- Source provenance: origins and repeated reporting
- Domain infrastructure: DNS/RDAP captures
- Geolocation: candidate places
Examine files
- Document triage: files, archives and email
- Media forensics: context and supplied metadata
- Browser artifacts: supported history copies
Test explanations
- Log timeline: dates and uncertainty
- Evidence correlation: candidate links
- Hypothesis audit: alternatives and contradictions
Hand off
- Investigation report: referenced findings
- Case export: reviewed sharing or private restoration
What has been verified
Version 0.1.0, local checks recorded on October 14, 2026:
| Checked locally on Linux/WSL | Still to evaluate |
|---|---|
| 81 automated tests passed | Real skill discovery and invocation in the four clients |
| 12 standalone runtimes and 4 installation adapters checked | Model investigation quality and prompt-injection resistance |
| Synthetic demo export and restoration verified | Native Windows and macOS execution |
Client manifest observations are dated October 13 in the compatibility ledger. Packaging checks establish neither client invocation nor model performance.
Benchmarks: 12/12 offline helper probes passed in the recorded October 14
run. Measurements, raw results and reproduction.
SkillsBench pilot: Codex passed citation-check 3/3 with the package and
3/3 without it. No explicit skill read was observed, so this does not demonstrate
a skill benefit. Protocol, results and limits.
Essential limits
- Local tools, independently chosen model. Python helpers work offline. A remote assistant may transmit the content it reads; apply its data policy.
- Preservation requires interpretation. Hashes check bytes, not authenticity. Correlations are leads, and missing times or incomplete histories remain gaps.
- Bounded analysis. Supported exports only; 64 MiB per file, 256 MiB of raw evidence per case, 4 MiB structured inputs and 5,000 events. Formats and full limits.
- Review before sharing. Keep real cases outside this repository. Private exports contain evidence; public views require review and exact-byte approval. Automatic masking does not guarantee anonymization.
Security model · Sharing and restoration
Go further
Documentation · Commands · Case format · Troubleshooting
Guides are in English and included in the full plugin. Cases use a shared versioned contract so another analyst or client can resume them independently of a conversation.
Apache-2.0 · NOTICE · Visual identity and asset provenance
Silex and Rivet are promotional mascots. Client names describe intended compatibility and imply no affiliation or endorsement.