Reading the Ontology
Three questions put to a working estate, with their reasoning traces, their sources and their cost. What the answers look like when the ontology does the work instead of search. And why every one of them inherits the quality of the ontology underneath.
A question lands on an analyst's desk at 9am. The lender wants a letter on where the permitting stands, and the credit committee sits after lunch.
In a lot of firms the next two days go like this. Pull the project file. Work out which of seventeen approvals are actually in hand. Chase the ones that aren't. Read the tracker, notice two cells disagree with the folder, ask the person who filed it, wait. Then write the letter, and sign it, and hope the tracker was right about the four items marked pending.
This is the fourth piece in a series on what Coherence actually is and does. Part 1 covered what Coherence is: the layer that models what is required to exist, not only what does. Part 2 covered how the source material gets in. Part 3 was the ontology itself: the kinds of thing, the connections, and why the grammar is fixed. This part is what the same day looks like when the domain is already authored.
To make it concrete: we modelled a construction-permitting practice end to end. The legal framework, a shared canon of permit types, the issuing authorities and the dependency rules between them, two complete project files with their documents and internal correspondence, and each authority's real processing-time history. The material is synthetic but faithful: real cadastral numbers, real deadlines, no client data.
Then we asked it three questions a partner would ask. Each answer carries its full reasoning trace, its sources and its cost. In all three, watch what does the work: the ontology, not search.
One: it refused the wrong document, and said why
"What do we know about this project's land?" Ten steps, about six minutes, around two and a half cents.
It produced a seven-section answer (cadastral number, area, owner, encumbrances) with every attribute carrying the document it came from, and the parcel's identity confirmed by two independent records.
The interesting part is what it left out. The file also contained a land-register extract belonging to a different property. It excluded that extract explicitly, and gave four reasons: area 31,640 against 84,320 square metres, rural against urban zoning, a different owner, and a legal mortgage that does not exist on the real parcel.
A retrieval system would have cited it. It is the closest thing to "land register" text in the folder. The report would have come out wrong on four attributes and carrying a phantom encumbrance, with your name on it.
Search matches text. The ontology knows which parcel is the project's.
Two: it told a structural gap from a simple delay
"What is missing from this file?" Twenty-five steps, twenty-five-plus sources.
It returned four absent permits. For one of them it made the finding no tracker can make: the request is not pending. It is structurally rejected: filed without two permits it depends on, so the clock was never running. It is back at zero.
It reasoned on dependencies rather than documents, which is why it also caught something else. An earlier approval in the file rests on a permit that expired months ago, so its basis is void. A keyword search would have found that approval and read it as valid, because on its face it is.
And it read one authority's own history, 96 days against a statutory 30 on every file, as a systemic bottleneck rather than an isolated delay.
None of those three findings is in any single document. They are all consequences of how the documents relate, which is the only place that kind of finding can come from.
Three: it wrote the letter a committee can act on
"Write the response to the project's lender." Forty-five steps, six tools, about five minutes, around two and a half cents.
The complete letter: fourteen of seventeen approvals obtained, three missing, and a realistic filing date that appears in no document, derived from each authority's measured pace rather than the statutory deadlines nobody meets.
Then it found the thing that decides the financing. That filing date lands eleven days before the planning certificate expires. It proposed the fix: file the extension a month earlier. And it exposed the false greens: items ticked off in the tracker that expire before the filing date arrives.
Every concept the letter cites was machine-verified against the ontology before it shipped. Twenty of twenty real, zero fabricated. The check re-caught the foreign land register from the first question.
A dashboard shows four items marked "pending". The ontology shows one request structurally rejected with its clock at zero, a systemic bottleneck at a named authority, and a collision between two dates that no single document contains. That is the difference between reporting a status and being able to act on it.
So which of the three does the analyst open?
The three answers above are the three products, and it's worth naming which was which.
Governed answers. Question one. Ask the estate anything; the structure decides what a complete answer contains, every claim cites its source, and where the evidence runs out it says so instead of improvising. Your own applications and your own AI can query it the same way.
Generated deliverables. Question three. Specifications, assessments, impact analyses, letters, chaptered, every chapter naming its sources, gaps declared rather than filled. A draft your expert edits and signs, never a blank page.
Governed change. Not shown above, because finding something is not fixing it. Modernization run as chartered initiatives through two human gates to cutover, data proven correct to the cent, and what ships writing itself back so the ontology cannot go stale as the estate moves.
Question two sits underneath all three. It isn't a product; it's the property that makes them work. The ontology holds what the domain requires to exist, so the difference between that and what your material actually provides is computable rather than noticed.
The answer doesn't go stale
A compliance report is a fossil the moment it's saved. It shows the estate from the day somebody assembled it, and it still shows that next quarter.
A read on the ontology behaves differently. What gets saved is the question rather than the answer, so opening it runs it again against the ontology as it stands. When last night's re-reading changes which permits are in hand, the answer changes with it. One that came through drops off without anyone editing anything.
And the currency has a mechanism behind it rather than a promise. When a change ships through the platform, the same step that records it writes what it built back in. So the finding that was red turns green on the screen where it was first found.
There's a companion rule that matters more than it sounds. A decision remembers what it was based on. When one of those things later changes, the platform opens a finding for a person rather than quietly revising the decision, and it clears only when somebody re-attests it. What you approved is never rewritten by what the ontology concluded afterwards.
The precise claim, since it's easy to overstate: the ontology is kept current by re-reading, and divergence shows up as a finding. It isn't a live mirror of your systems. Drift is flagged, never silently repaired.
An analyst can't cite what they can't see
The clearances from Part 3 don't stop at the thing in the ontology. They follow it into every answer. The same question, asked by two people, can come back with two different answers, and the regional owner sees their own scope without setting a filter for it. Inside a project, commercially sensitive material stays hidden from anyone not cleared for it, in the letter exactly as in the ontology.
Nothing has to be configured for that to hold. Every read inherits the rules already there, so there's no route around access control, and material somebody isn't cleared for never reaches their draft in the first place.
Asking in plain language carries the same discipline plus one more. Every reference in an answer has to appear in what the tools actually returned, or the answer is replaced with an honest one and its confidence capped. Where the ontology holds nothing, it says so rather than inventing something.
What this changes day to day
The archaeology goes away. The work that used to fill those two days (pulling the file, reconciling the tracker against the folder, chasing which approvals are real) lands with whoever actually has the question, and takes minutes.
The definition of an obtained approval is the one the ontology carries, so two analysts asking the same thing start from the same answer, and the letter one of them signs points at material the other can open and read.
And every finding ends in a decision rather than a chart. A short fixed list of things a person can do with it, each recorded with their name on it. Some stop the finding coming back. One of them opens a piece of work.
Four things we won't soften. What the ontology knows is bounded by what was read into it, and material nobody supplied is a declared gap rather than an inference. Nothing here measures people. We show that a file is sitting with an authority and for how long, not what the official does next. And the judgement stays yours: what the asset is worth, whether the deal is sound, what you advise. We change what your expert spends the week on, not what they decide.
One caveat, the same one Part 3 closed on. Every answer here inherits the quality of the ontology under it. Point them at a domain nobody has authored and the analyst is back in the folder, reconciling by hand. Good answers are evidence that the ontology is good. They aren't a way around authoring one.
Finding it isn't fixing it
By the afternoon committee the analyst has the letter, the three missing approvals, the realistic date, and the collision that decides the financing. All on current material, inside her own clearances, in about five minutes.
Naming the extension to file is not the same as filing it. That's a different job and it needs a governed path rather than the analysis behind it. Part 5 is that path: a sentence becomes chartered work grounded in the real finding, scope gets proposed by the machine and confirmed by a person so it can't creep quietly, and two different people sign anything that changes reality. The AI drafts and never signs.
The permitting estate is modelled on synthetic material: real cadastral numbers and real deadlines, no client data. Step counts, timings and costs are from that estate and are not presented as transferable. Effort figures in this series are design targets validated per engagement, not warranties.
This site uses cookies
We use essential cookies for the site to function and analytics cookies (Google Analytics) to understand how you use it. Analytics cookies are only activated with your consent. We do not track you across other websites. Your data is stored in the EU and processed in accordance with GDPR. Read our Privacy Policy
