Point us at a slice of your real data — one business unit, one document class, one year. Real records with real ambiguity in them. Curated sets make every system look good and tell you nothing.
We run the domain probes and show you, against your own records, what can already be answered and where your corpus has gaps. You get the gap list whether or not you proceed.
Every deployment runs this in order. We will not take a deployment past step three until all three tests pass.
Read-only access to where your knowledge already lives — mail, document libraries, structured exports, scanned archives. Nothing is migrated. Nothing is moved.
We run the domain probes before configuring anything.
Corpus integrity: what proportion of ingested documents actually have readable text (below 80% is an ingestion problem, and no amount of search tuning will fix it); what proportion of fragments carry a valid entity identifier; whether the searchable representation is internally consistent.
Entity resolution: what proportion of inbound records resolve to a known entity, and what the top unresolved references are.
Obligation patterns: sampling correspondence for obligation density and the vocabulary your domain actually uses for commitment.
Document structure: which classes have to be split along their own hierarchy, and where the load-bearing information sits.
The output is a written gap list.
Three acceptance tests, run against your data, in front of you.
State interrogation. Ask what is outstanding for a given entity. The answer must be synthesised from extracted obligations across correspondence and documents — not a search for documents containing the word.
Grounded citation. Ask a domain question. Every factual claim must be traceable to a specific source, fragment and entity. An uncited claim means ingestion or synthesis is wrong.
Active learning loop. Approve one correction. Ask its exact topic again. The correction must appear in the retrieved set. If it does not, the loop is inert regardless of what the records say.
Scheduled ingestion. Accumulating obligations. Compounding corrections. Your team’s corrections are the asset from this point forward.
A system that retrieves your documents but cannot answer a question about your state is a search tool. We would rather tell you that at week three than sell it to you as something else.
A real slice, not a sample set: one business unit, one document class, one year.
Alongside it, five questions your team currently answers from memory, and the name of the person they ask when they cannot. Those five questions are the actual specification. The person they ask is the risk you are trying to retire.
We do not need to hire your industry before we start. On a foreign public bench, failure diagnosis moved a clean 77.5 to a labelled 90.0 in a working day, without a hired domain expert. That is how we enter a new ontology. It is not a claim that your operators are optional. They still approve the corrections. The expert is not gone. The six-month workshop is.
It is whether your business generates unstructured records that people have to remember.
The engine is the same every time. The vertical is the ontology: the taxonomy, the obligation types, the document classes, the language of the answer.
Property is the furnace. Insurance and fire are structured-layer pilots. A paying second vertical, in a workflow that is not adjacent, is what converts portability to proof. If that is you, say so in the first line of the email.
The full technical brief — architecture, ingestion, governance, deployment, and the limitations section — is available on request from [email protected].