Industries

Where a typed judgement actually earns its place.

Each industry below has one candidate judgement written out. Only one is worked through in full. The rest are sketches. The point of listing them is that the shape repeats.

Telecommunications

Worked through

Is this inbound fault report caused by the network, rather than billing or the customer's own device?

32 judgements inventoried across the fault workflow. Only six to eight turn out to be genuine System One tasks, and the specification says so rather than inflating the count.

Includes the gating experiment: an 8.40 USD calibration audit over 100,000 historical decisions that can kill the programme before anyone builds anything.
Open the specification →

Insurance

Sketch

Can this motor claim be settled without sending an assessor?

High volume, a fixed answer set, and a cost asymmetry that is easy to price: an unnecessary inspection costs a known amount, a missed one costs a great deal more.

The hard part is whether the threshold survives a conversation with the claims director.

Healthcare revenue cycle

Sketch

Is this denied claim worth appealing, and what is it likely to recover?

Judgement was rationed here because appeals cost staff minutes, so teams work the queue in the order it arrived rather than by what each case is worth.

Constrained by data residency in several jurisdictions, which decides the architecture before any accuracy question is asked.

Banking and compliance

Sketch

Does this alert warrant a human review, or is it the pattern we have cleared a thousand times?

A very low base rate, which is exactly where accuracy misleads and calibration decides whether anything can be automated at all.

Regulation usually requires a person to sign. That makes it a triage problem, which is a smaller claim than automation.

Procurement

Sketch

Does this supplier bid meet the mandatory criteria of the tender?

A closed answer set, documents that already exist, and a decision currently made by reading every submission in full.

The honest question is whether the rules are already written down well enough that deterministic code would do it.

Logistics

Sketch

Is this shipment going to miss its window, early enough that someone can still act?

The judgement is cheap. The bottleneck moves immediately to physical execution, which no model accelerates.

Here the chain ends at Track 1 for reasons that have nothing to do with the model.

Legal and contracts

Sketch

What kind of clause is this, and does it deviate from our standard position?

Around 100 clause types, long inputs, and a public dataset (LEDGAR) that makes it testable before any client is involved.

Long context is where judgement quality degrades first. Worth measuring before promising anything.
Why only one is worked through. A sketch costs an afternoon and proves nothing. The telco specification took the work of inventorying a real workflow, which is where the finding came from: most of it does not qualify.

← DashboardTalk to Peter →