← All posts

2026-06-15 · Nebbos

Run the cheap check before the expensive one


title: Run the cheap check before the expensive one description: Deterministic detection runs continuously and costs nothing to watch; only a firing pattern ever spends a model call. date: 2026-06-15 author: Nebbos

The lazy default for an AI product is to put a model call on every event. A ticket arrives — reason about it. A deadline moves — reason about it. It looks intelligent and it scales badly: most events are unremarkable, and you have just paid a premium to be told so.

Most events are not worth a thought

An operation emits a constant stream of small signals. The overwhelming majority mean nothing on their own. Spending a model call on each is like hiring an analyst to read every line of a log. The skill isn't reading everything — it's noticing the handful of moments that have actually changed shape.

Six detectors that cost nothing to watch

Nebbos runs deterministic detection continuously, underneath any reasoning. Six detectors watch the shape of work as it moves:

  • Deadline risk — a commitment drifting toward a date it won't make.
  • Capacity crunch — load arriving faster than a team can clear it.
  • Velocity drop — throughput that fell off and hasn't recovered.
  • Handoff stall — work passed across a boundary that nobody picked up.
  • Absence signal — an expected event that simply didn't happen.
  • Cascade risk — a single slip positioned to take others down with it.

These run on patterns, thresholds and timing — arithmetic, not inference. They are effectively free to keep running, which is the whole point: you can afford to watch everything when watching is cheap.

Only a firing pattern spends a thought

A model call is a deliberate act, not a reflex. It happens when a detector fires — when something has crossed a line worth reasoning about. At that moment the question is no longer is this interesting but what does this mean, and what should the lead do about it. That is work worth paying for, and only then do we pay for it.

The right model at the right price

When reasoning is warranted, governance-tiered routing sends it to the model that fits the call: a fast model for a routine read, a stronger one where the stakes and ambiguity justify it. The tier is a governance decision, not a hardcoded vendor choice. Nebbos is model-agnostic — there is no lock-in to a single provider, and the routing can change as the frontier does.

The effect is restraint by design. Detection is constant and cheap; reasoning is rare and earned. You get the early read without paying to think about everything that was never going to matter.