Who decides what better means when AI ranks public projects?

The record knows everything except the point

It is June 12, eighteen days before the fiscal year closes.

The supervision dashboard for one road contract shows six facts: physical progress 61 percent, funds disbursed 78 percent, eighteen days to fiscal close, three approved variation orders and one pending, five weeks to the rainy season, and a contractor carrying two other public jobs.

On the desk: a payment certificate and the pending variation. Approve?

There is no correct answer yet, because the decisive question is hiding inside that one: what is this quarter for?

If this quarter is for closing the year, approve and release. Unspent funds return to treasury, and next year's ceiling drops with them.

If this quarter is for protecting the record, hold the certificate until the site verifies the 61 percent. Disbursement is seventeen points ahead of the road.

If this quarter is for building a capable contractor market, split the scope, keep a stretched contractor liquid, and move the risky section past the rains.

The board did not change. The definition of better did. And the signature is not an abstraction: whoever signs the certificate answers to the Auditor General.

The June 12 board is an illustrative composite of infrastructure supervision dashboards. The numbers teach the mechanism. They do not report a specific contract, contractor or country.

A decision is the facts, times a definition of better that someone owns

Data can say what has happened, what is happening and what may happen next. A capable model can find patterns across more of the record than any one engineer can hold in mind.

But prediction is not purpose. A forecast does not decide whether absorption, audit integrity or market capability deserves more weight this quarter. The facts can be shared. The definition of better cannot be assumed.

That middle step is the decision layer. It is not a component you can buy. It is work: choosing, out loud, which legitimate goal rules this quarter, and owning what follows.

Any ministry can procure the same dashboard, the same model and, increasingly, the same data standards. What no procurement can deliver is a definition of better chosen here, this quarter, by someone who answers for the outcome.

This is not a public-works quirk. Swap the ministry for a restaurant at 7:02 p.m. deciding between covers, quality and training a new cook, or a classroom in exam season deciding between this month's scores and a child's confidence: the same trap repeats wherever goals compete. Chess flatters the machines because chess supplies its own objective; checkmate is checkmate. A road programme has to choose its objective again every quarter, and someone bears every tradeoff.

Name the design mistake plainly: treating an intelligence system as if it were already a decision system.

The site knows things the portal does not

The people closest to the work hold context that is distributed, unwritten or simply newer than the record.

The supervision engineer can tell which variation order is engineering and which is appetite; on the portal the rows look identical. The 61 percent is last month's certificate, and the culvert at kilometre 14 washed out on Tuesday. The contractor is competent and stretched, and one late certificate tips all three of their sites into arrears. The escarpment section cannot be compacted wet, and missing the five-week window means waiting until February.

Some of this can and should be written into the record. Some changes faster than any record. Some is learned only by standing on the site.

F. A. Hayek described relevant knowledge as dispersed, incomplete and sometimes contradictory. Michael Polanyi argued that part of what people know cannot be made fully explicit. Local knowledge and tacit knowledge are not identical ideas, but both expose the same mistake: treating what a central system can store as the whole of what a decision requires.

The decision layer is not a box you own. It is a loop you run.

Run it on the June 12 board: the record gives 61 built, 78 paid, one variation pending; the site adds what the engineer, the accountant and the contractor know; this quarter's better is said out loud; one person approves, holds or splits and puts a name to it; and the outcome goes back into the record for the next call.

Station five feeds station one. The return stroke is the whole game.

Two guardrails, because a loop run badly is worse than none.

First, local judgment is not automatically right: it can preserve bias, folklore and obsolete practice just as faithfully as wisdom. The loop needs evidence, review and outcomes that can correct it.

Second, and more binding: an owned objective ranks the options that are already lawful and contractually admissible. It cannot authorise a payment the contract does not support, a variation the rules forbid, or a scope change outside the delegation of authority. The decision layer chooses among permitted options and answers for the choice; it does not manufacture permission. Preference is not authority.

Why this is the moat

Models get cheaper every quarter. Records and dashboards can be copied in a procurement cycle. The decision loop is a practice that compounds: each turn can leave the record truer and the judgment sharper. For a public institution, that is the capability that remains when the vendor leaves.