We turn operational judgment into dependable software. Rules carry the load; models enter only where ambiguity makes them useful—and every decision leaves a trail.
“We use AI where it earns its place” only matters if the test is visible. This is ours—and it is the first thing we invite you to challenge.
THE DEFAULT
Use deterministic automation when…
The rule can be written down.
The same input has to give the same output, every run.
The answer is checkable by arithmetic or a lookup.
Being wrong would be expensive and silent.
Someone may need to reconstruct the decision years later.
THE ROUTING QUESTIONCan the decision be stated as a rule?
YESNO
THE EXCEPTION
Invite a model in when…
The input is free text, or a layout nobody has seen before.
The rule exists in people's heads but has never been written down.
Judgment is genuinely required — two careful people might disagree.
The output is a draft that a person reviews before it counts.
Failure would be visible and cheap to correct.
02 / THE BUILD DISCIPLINE
Four layers. Nothing hidden.
Each layer makes the next one safer. The method is deliberately heavier where the cost of being wrong is highest.
01
Boundary contract
Validation before anything else
Input is checked at the boundary — shape, type, range, and whether the thing it references actually exists. Work that fails validation stops there and is queued with the reason attached, rather than being partially processed.
DESIGNED TO PREVENTBad data entering a pipeline and being discovered three steps later, or three weeks later.
02
Readable workflow
Deterministic automation carries the load
The bulk of every system is ordinary code: rules, lookups, transforms, state machines. It is readable, testable, and it behaves the same way on Tuesday as it did on Monday. This is the part that does the work.
DESIGNED TO PREVENTOutcomes nobody can reproduce, and therefore nobody can debug or defend.
03
Versioned rubric
Judgment is governed where it happens
Where a model is warranted, the call runs server-side against a rubric held in version control, with its inputs and outputs logged. The rubric can be read and argued with. Changing it is a change you can see in a diff.
DESIGNED TO PREVENTA quiet prompt edit changing results for everyone with no record that anything moved.
04
Decision ledger
An audit trail that outlives the project
Every decision keeps what it decided, what it saw, and which rule or rubric version applied. Extracted values keep a reference back to their source. The trail is part of the system, not a logging afterthought.
DESIGNED TO PREVENTThe answer “we don't know why it did that”, which is the answer that ends trust.
03 / THE BOUNDARIES
No black boxes. No theatre.
A method is only credible when it costs us something. These are the shortcuts and engagements we will walk away from.
WHERE WE ARE
We are early and deliberately small. Engagements are in build and testing rather than finished stories we can point at, and we would rather say so than assemble a wall of logos. In practice: you talk to the people writing the code, and scope stays small enough to actually finish.
01
Put a model in the path of a decision a rule would make better.
02
Leave you with automation that only we can maintain.
03
Model an operation whose data cannot yet support a model.
04
Treat a slide deck as a deliverable.
HAVE A PROCESS IN MIND?
Bring us the messy version.
We will come back with questions before we come back with a proposal.