TECHNOLOGY
Weights and retrieval learn different things.
This is not a choice between two ways of doing the same job. One teaches the model the task. The other teaches it this operator. A system that only does one of them fails in a way the other would have caught.
Weights — fine-tuning
changes monthly- Learns
- the shape of the job: vocabulary, the categories, the output contract, house tone
- Time to effect
- a training run — hours to days
- Auditability
- none at the row level. You cannot say which example produced this answer
- Reversibility
- retrain, or roll back to a previous adapter
- Scales
- well — fifty thousand examples all contribute, at no context cost
- Fails when
- you have eighty examples. It memorises noise, and a thirty-case eval cannot tell
Retrieval — exemplar memory
changes in seconds- Learns
- this operator's judgements: €40 not €75, waive the detention, settle small claims in full
- Time to effect
- the next request. Two seconds
- Auditability
- complete. "Informed by these three past decisions" is a screen you can show
- Reversibility
- delete the row and the behaviour is gone
- Scales
- badly — you can only fit three or four decisions in a prompt
- Fails when
- the model does not understand the domain well enough to use what it retrieved
Why you need both
The moment that convinces a room is someone correcting a decision and the next similar case answering differently. No amount of fine-tuning does that live — you would need a training run mid-meeting. That behaviour requires retrieval. And retrieval on a model that doesn't know the domain just quotes the manual back at you. Neither half is optional.
Act. Measure. Human feedback. Improve.
Human approvals and overrides are inputs to retraining, not a separate governance layer. This is why “people stay in charge” and “it gets better over time” are the same sentence at MoveCore rather than two claims in tension. Every override is a labelled example. Every approval is a confirmation. The corrections your customers' staff make during their working day are the training set.
- 01
Notice
watches every event as it happens
- 02
Decide
chooses the next best action
- 03
Act
does the work, or asks a human first
- 04
Check
did it actually work?
- 05
Learn
adjusts so next time is better
How we know it got better.
Every Domain ships with code-based evaluations covering the decisions it is permitted to make. Changes to weights or memory are measured against them before promotion, and continuously after. A model that improves on average but regresses on a safety-relevant category does not ship.
Safety before compliance. Compliance before cost. No exceptions.
If anything fails, the system falls back to safe mode. It never guesses.