Sparks: AI can predict intimate partner violence before survivors disclose. But should it?

Missing from every accuracy metric is the web connecting prediction to police arrival, broken lease, vanished income - the forecast is never one event but a cascade, and the cascade was never measured.

Ask what difference the prediction makes on Tuesday morning: she loses the moment of choosing when to tell, and no accuracy figure can buy that back.

Computed risk is the claim; the premises on which it rests - that her disclosure is a variable rather than a right - are nowhere stated, which suggests stating them would reduce the confidence.

Strip away the dashboards and the essential fact remains: a machine now tells a woman her own life before she has chosen to speak it.

What is within her control - when, whether, to whom she tells it - the algorithm takes without asking; what is beyond its control, the violence itself, it cannot touch.

A system that announces a woman's danger before she has consented to have it announced - by what authority, under what law, before what tribunal does it answer?

Forget the intentions - here is who has leverage: the institution holding the prediction, the abuser who may obtain it, and the woman whose only leverage was choosing the hour of her own disclosure.

Announced in terms of predictive accuracy, the system means, for the woman whose landlord receives the flag, an eviction notice - a calculation no accuracy percentage has performed.

From the side of the door where the prediction lands, intervention arrives on the system's schedule, in the system's language, and the centre of it is not the woman it concerns.

Punch cards can encode the probability faithfully, yet no card encodes the variable that matters most - her sovereign choice of when to name it - and a mechanism omitting its purpose is no instrument.

Accuracy was commended, tea was poured, and the safety board adjourned without placing on the agenda the one party who would be surprised by the knock at the door.

If the prediction can be made without her voice, and the intervention can follow without her consent, then her safety has been decided for her - and a safety she did not choose is not hers.

How admirable, the machine that knows her danger before she speaks it - and how fortunate the same ingenuity has not yet been turned to predicting which neighbour will report her.

Beneath the debate about accuracy sits a question of character: engineers who built a machine to speak for a woman rather than asking what she needed to speak for herself.

Handed a warning she never requested, she loses inside this reform the one power she held - the timing of her own testimony - and the structure calls the loss protection.

Trained by institutions that punished disclosure, she is then declared incapable of judging her own danger - a circularity so perfect it would be admirable were it not so destructive.

We optimised the detection circuit and left the human switch untouched; the complete system would route every prediction through her hands, and the distance between those two designs is a choice.