The escalation decision is computed, then it stops at the glass

The seam

Your own post states it:

For our final approach, we built a single multi-task model that predicts three things at once: Escalation decision (3-way classification) Escalation reason (8 categories) Guideline citations (which guidelines to cite)

C-2 · SOURCED · fetched 2026-08-12

The decision reaches the customer as a handoff. The reason and the citation are computed, logged, and as far as your published material goes, they stop there. Reason accuracy is 97% and citation AUC is 98.7%. The outputs are good enough to hand to a colleague, and this memo is about handing them over.

The contract

Five fields, rendered at the moment of escalation instead of a scrollback: what Fin tried, why it stopped, confidence, what the customer has been told, what not to repeat. One of them is yours already.

When escalating due to a guideline match, we cite the specific guideline.

C-9 · SOURCED · fetched 2026-08-12

So the branch and the rule that fired arrive as fact, with an id and a fetch date behind them. The other four are designed, and both prototypes say so on the card rather than in a footnote.

When nothing flags it

A contract that only works when the model raises its hand is a courtesy. The second prototype is a conversation where Fin did not escalate, answered confidently, was wrong, and nothing flagged it. The five fields still have to carry something there, and the one that earns its place is the field reporting an absence: nothing tested the diagnosis, and no turn in a transcript can say so.

What it costs

Filling in why it stopped, beyond the citation you already compute, means writing escalation guidance. Your documentation prices that.

The more escalation guidance you add, the more often Fin escalates.

C-10 · SOURCED · fetched 2026-08-12

Broader guidance leads to more escalations, fewer resolutions, and higher human support volume.

C-11 · SOURCED · fetched 2026-08-12

The corpus is also bounded.

You can have up to 100 active guidance rules at a time. If you try to activate more, you'll see an 'invalid parameters' error.

C-12 · SOURCED · fetched 2026-08-12

So this spends a capped resource and costs resolutions, and it is worth that only if the fields are read. Measure that first: time from take-over to first outbound reply, against a transcript-only baseline on the same queue. If the number does not move, the contract is decoration and should be withdrawn.

What this does not propose

Preview, Batch Test, Procedure Simulations, guidance authoring assistance and post-publish analytics already exist. Those instruments test the rule being written. This is about the conversation after one has fired.

Why a contract rather than an assistant

pass^k is success every time across k repeats, which is a hard metric for anything generative standing at a handover. So this surface generates nothing: every field is a value your model computed or a value the scenario declares as designed, and the card says which.

Allowing an agent to take long horizon decisions bounded only by the availability of tools is a recipe for unreliable behaviour.

C-15 · SOURCED · fetched 2026-08-12

A briefing that summarises is a generation step at the handover. Five named fields, each computed or declared, is a rendering step. The contract page specifies all five, including what happens when each is missing.