Draft and explain
Produces an answer, summary or proposal from the supplied context.
What still needs checkingCheck facts, sources, omissions and the checks actually performed. An answer alone is not an execution log.
An answer is displayed. A decision must be substantiated.
Facts, rules, checks and approvals: examine what each system provides, then test the traces of a single fictional dossier.
Retrieve and contextualise
Adds retrieved documents to the generation context. The answer can cite sources.
What still needs checkingCheck retrieved documents, their versions and their interpretation. Enriched context guarantees neither completeness nor a correct answer.
Check and retain the trace
Executes explicit conditions on identified facts; retains the versions, results and approvals provided for by the system.
What still needs checkingCheck rule correctness, scope coverage and fact quality. Reproducible does not automatically mean legally correct.
These components can be combined. A product using an LLM or RAG can also include deterministic checks and a log. This compares functions, not audited competitor products.
Change the dossier. Examine what changes.
Fictional internal procedure: before closure, an owner must be identified and two documents present under version 1; three under version 2. This is not a legal rule.
Context being checked
- Procedure version : 1
- Identified owner : yes
- Documents present : 2 / 2
The buttons change this mini-scenario’s inputs. The displayed conditions are evaluated locally.
Fictional check result
Ready for human approval- ✓ Owner provided
- ✓ Required document count met
The result is neither a legal compliance verdict nor automatic permission to close the case.
Original document fingerprint
The JSON document declares the original owner. ‘Alter a document’ changes that owner in a copy, separately from the decision context above.
Calculating SHA-256…Copy verification
Checking…SHA-256 comparison computed in your browser against this demo’s original. It does not establish content truth, authorship or date.
Educational demonstrator: condition and fingerprint calculations in the browser, with no call to an LLM, a RAG system or the production NELSE engine. Historical reconstruction covers only the fictional data preserved here.
Understand the limits and references
NIST describes the risk of erroneous generative AI answers. Google’s research on sufficient context shows that RAG can still produce incorrect answers. Neither determinism nor a fingerprint certifies a dossier’s legal correctness.
NIST · Generative AI Profile · Google Research · Sufficient ContextNELSE Proof
Examine the package.See exactly what the check establishes.
The public verifier works locally in your browser. It recalculates the internal consistency of a NELSE demonstration package instead of asking you to infer integrity from a screenshot or a dashboard status.
Browser-local verification
Recalculate it yourself.
Load one of the public fictional cases or drop a compatible JSON package. The browser recalculates fingerprints, chain continuity and root without uploading the file to NELSE.
No download required: click to verify
Cataract surgery
No statutory reflection period is declared for this procedure in the demonstration corpus. A professional-practice delay can still be recorded separately and its provenance remains visible.
Cosmetic surgery
The demonstration package records a 15-day period after delivery of the estimate (French Public Health Code, Art. D. 6322-30), with the resulting timing constraint represented explicitly.
The file never leaves your browser. Fingerprints are recalculated locally. NELSE does not receive the package, its contents or the result.
Download package « Cataract surgery »·Download package « Cosmetic surgery »
Establishes · does not establish
Proof is stronger when its boundary is explicit.
Control
Establishes
Does not establish
Structure
A journal and declared root are present in the submitted package.
That the package really originated from a particular source.
Fingerprints
Recalculated SHA-256 fingerprints match the presented events.
That the data contained in those events is factually true.
Chain
Each event references the recalculated fingerprint of the preceding event.
That no other internally coherent version of the file ever existed.
Root
The recalculated root matches the root declared by this package.
That it existed at a particular date without independent external evidence.
Corpus metadata
A module, version and effective date are declared.
That the corpus is legally complete or correctly encoded.
External anchor
The package can declare an anchoring state; external evidence can strengthen the check.
Authenticity of an external anchor when its independent receipt is not provided.
From integrity to replay
Verification is one layer of the trust chain.
A public package can demonstrate internal consistency. Stronger assurance can additionally depend on source provenance, fact evidence, human review, external continuity evidence and the preserved historical execution context. Those layers must not be collapsed into one marketing claim.