Offline document intelligence for assurance and engineering teams
Audit controlled documents before the wrong revision travels.
Cortessa cross-checks an approved document set for conflicts, superseded content and broken references. It runs locally inside your boundary, shows engineers the cited source and leaves the decision with them. Cortessa does not certify.
- Audit leads
- Analysis stays on the engineer box
- Human authority is retained
Cross-check the document set before a wrong revision travels
Audit examines relationships across an approved set of controlled documents. It surfaces cited candidates for an engineer to inspect; it does not make the decision for them.
-
01 · Cross-check
Review the approved set
Cortessa checks for supersession, broken references, stale content and terminology conflicts across documents.
-
02 · Show the source
Open the evidence
Each candidate finding points back to the cited passages so the issue can be checked in context.
-
03 · Human decision
Review and decide
The engineer records the disposition and retains approval authority. Cortessa never certifies.
A prompt is an instruction. Cortessa adds a separate check.
A general-purpose language model can be told to cite sources and refuse when unsure. That instruction is still interpreted by the same model producing the answer. Cortessa checks the output separately before presenting any claim as Verified.
-
The prompt
The model interprets it
A request to cite or refuse guides the answer, but it is not an independent acceptance rule.
-
The Cortessa gate
A deterministic rule checks it
A versioned rule outside the model decides whether each claim is eligible for Verified, Draft or Refused.
-
The human decision
The source stays open
Verified attests that the claim occurs in the cited source; it does not certify correctness. A person decides.
What exists now, what is in development, and what is not claimed
The local product and governed revision-exchange foundations exist. The complete site network and production rollout do not.
- Implemented locally On-box document ingestion, indexing, resident model, Audit and Ask; human review; and governed revision-exchange foundations.
- In development Author; authenticated peer transport; key provisioning and recovery; delivery scheduling; backup and restore; migration; controlled document promotion; and production site rollout.
- Not claimed No deployed customer network, production accreditation, automatic synchronisation or replacement of every document- and records-management function.
For the people responsible for controlled evidence
Cortessa is built for regulated organisations where the current, correct, cited fact matters and cloud AI may be barred at the boundary.
-
Daily users
Assurance, quality and systems engineers
Review cited findings, inspect the source and record the engineering disposition.
-
Decision makers
Programme and engineering leaders
Evaluate a controlled workflow against agreed evidence and success criteria.
-
Boundary owners
Security and accreditation teams
Assess a local, no-cloud-dependency design inside the organisation's own boundary.
Local intelligence. Site-wide document continuity.
Each engineer box runs its approved documents, ingestion, index, model, Audit and Ask locally. The target site architecture connects those boxes through departmental nodes to an on-site hub, entirely inside the site boundary.
Only explicitly approved, signed revisions are designed to move. An AI answer does not become site truth, and receiving a revision does not silently add it to another engineer's searchable document set.
-
Engineer boxes
- Approved local documents
- Audit · Ask
- Human approval
-
Departmental nodes
- Permission-scoped exchange
- Signed revisions
- Visible history
-
On-site hub
- Permitted information only
- Conflicts remain visible
- No central approval
Verified, Draft or Refused — never a hidden confidence score
The source check records what can be established from the cited material. It does not decide whether the source itself is correct.
The claim is provably a near-verbatim span of the cited source.
The claim cannot be proven against the source and is not passed off as fact.
The required fact is not present, so the system refuses rather than guess.
The evidence, with its test context
- ~97%retrieval hit-rate across a 480-question blind evaluation over a 700-document, 29,506-chunk public corpus. Raw artifacts hash-manifested.
- 76.9%strict factual accuracy in the shipped configuration on the same sealed evaluation; 91% on single-fact questions and 92% on entity disambiguation.
- 16 / 16model fabrication attempts held back by the source gate in that run; none reached a user as Verified.
- 156real, public NASA documents in the live demonstration set — every citation opens a document anyone can download and check.
Internally executed blind protocol on real, public document sets: the engine never saw the sealed key; a separate source-only oracle authored it and a different model family judged the answers. Independent external replication is a pilot objective. Figures are from interim development hardware.
Cross-checking controlled documents by hand?
Begin with an agreed, limited evaluation of Audit and Ask. Use the evidence to define the first departmental workflow and the route towards the target site architecture.