Evidence for AI-generated reports that a reviewer can check later.
Elvane Ledger is an early-stage library that records the claims in an AI-generated report together with the exact passages and preserved sources behind them. It runs locally and stores evidence as files in a location you choose.
A citation points to a source. It doesn’t keep the evidence.
Weeks after a report is written, a reviewer may need to know which passage a claim relied on and whether the source still says the same thing. By then the tool, the model and the original files may all have changed.
Claim, passage, source, fingerprint.
- Claim“Customer concentration is 41%.”A statement from the AI-generated report.
- Passage“Customer concentration is 41%.”The exact text the claim relies on.
- SourceA preserved copy of the file the passage came from.
customer-summary.txt - FingerprintChanges if a single byte of the source changes.
sha256 2b0e0234…cc1fe2b5
A matching fingerprint shows that recorded evidence hasn’t changed. It doesn’t show that a claim is true, or that an AI actually used the source. Ledger keeps those questions visible instead of answering them for you.
What exists today.
| Form | A local library. Evidence is stored as files in a location you choose. It is not a hosted service. |
|---|---|
| Integrations | A local adapter for MCP-compatible tools, and a bundle format for exchanging evidence. |
| Providers | Not tied to a single model provider. |
| Stage | Early. Being validated on real report-review workflows with a small number of design partners. |
Two demos, both synthetic.
- Atlas Components: a fictional acquisition reviewClaims, evidence, verify, change and review
- The reservoir demoAn earlier, separate demonstration
Prerecorded Ledger operations and scripted model output. No live AI runs in your browser.
Evidence assessments and design-partner pilots.
We look at one AI-assisted reporting workflow with you: where the sources come from, what a later reviewer would need, and what is currently lost. You receive an evidence and gap assessment with explicit limits. If Ledger looks like a fit, a pilot is scoped separately, time-boxed, and any hosting or data handling is agreed with you rather than assumed.
Not included: fact-checking, proof that an AI used a source, compliance assurance, or production readiness.