Why no AI grades anyone here.
A named person attests every rating in Vontos. Here is what we let machines do, what we never let them do, and why the line sits where it sits.
The rule is one sentence.
AI assists; humans attest. No machine output becomes evidence about a person by itself. A named human reviews it, owns it, and seals it — and anything a machine helped draft is labeled as machine-assisted in the record, permanently. The lists below are the rule, not a feature tour — some of it is live, some is still being built, and all of it obeys the same line.
What machines may do here
- MAYDraft translations of framework text between English and Filipino — the author confirms before anything publishes.
- MAYTranscribe and propose — planned, with the rule set before the feature. A teacher's voice note becomes proposed entries the teacher edits and confirms — nothing records until they do.
- MAYSummarize what already exists so a busy reviewer starts oriented, never so a reviewer can skip reading.
What machines may never do
- NEVERRate a person. A rating is a human judgment with a name on it. There is no autopilot for judgment here.
- NEVERWrite to the record. Unconfirmed machine output cannot touch evidence, milestones, or the roll-up. Ever.
- NEVERDecide placement. Readiness calls belong to the people who know the learner.
And the roll-up is not AI at all.
The number on a learner's radar is a decaying average — disclosed arithmetic in which recent evidence outweighs old. It can be recomputed by hand from the record. There is no model, no inference, and nothing to explain that the arithmetic doesn't already say.
Where we got it wrong once.
[retracted:] Because a human attests every rating, systems built this way sit outside the high-risk tier of AI regulation. [end retraction]
The struck sentence above came from this entry's draft. It was wrong, and we're publishing it visible rather than deleting it. Human attestation is right on the merits — it is not a classification escape. Under the emerging rules, a system that evaluates named learners is treated as high-risk however the signature is arranged. Our current arithmetic isn't an AI system at all, which is the honest defense; the loophole we implied was not.
The honest position.
We hold a stricter line than the regulations require, because trust in a child's record cannot ride on a loophole. If a machine ever rates your learner in Vontos, that is not a feature — it is a bug report we want the same day.