That distinction is critical. A model answers. A Machineforce Unit decides — inside a boundary, with a memory of what it decided before, and a named human it answers to.
That the underlying model gets smarter without retraining. It doesn't. The base model is stateless. Weights don't change between calls. Anyone telling you their model learns on the job without retraining is selling you something that isn't true.
That the unit as a whole system gets measurably better at its job over time — through the layers around the model, not the model itself. The model does the language. The system does the judgment.
For most enterprise work, the performance-relevant behavior lives in the wrapper layers, not the raw model.
The model is the one part that never changes between calls. Everything else is where a unit's experience accumulates.
Before the model is called, the unit assembles what happened the last several times a situation of this shape arrived — and what the operator did about it. The model is unchanged. What changes is what it is looking at when it decides.
Same model · better inputsWhen a pattern holds long enough to be trustworthy, the unit proposes an adjustment to its own thresholds — and a human ratifies it. Next time that condition appears, behaviour differs.
Same model · new ruleRepeated successful handling gets promoted into a named procedure. When the pattern recurs, the unit follows the playbook — often without calling the model at all. Faster, cheaper, less to go wrong.
Same model · used lessWhen an operator marks an action wrong, that judgement attaches to that specific decision and its context — not to a vague global score. The unit can then tell how accurate it actually is on this class of work, and calibrate when it escalates.
Same model · better calibratedJob: process vendor invoices for accounts payable.
The standing rule is conservative: escalate any invoice over $5,000, or from a vendor the unit has not seen. An ACME Corp invoice arrives — $8,200, matching the purchase order. The unit escalates. The operator approves. The whole event is recorded with its context.
Same base model. Zero retraining. Zero weight updates.
The pattern was noticed, surfaced to the operator, ratified by a human, and written into the unit's standing configuration — with every step of that sealed to an audit record.
We'd rather you know the boundary of the claim than discover it later.
The second column is where enterprise task performance actually lives.
A good analyst doesn't rewire their brain. They build up experience, adjust their rules of thumb, accumulate patterns worth repeating, and calibrate their confidence against feedback from the person they report to.
That's exactly what a Machineforce Unit does — and unlike a person, every step of it is recorded, attestable, and bounded by a policy it cannot widen.
Five units are live right now. Talk to one, then try to push it out of scope.