The economics of noticing late

The quantity that matters

For any belief a system might maintain, there is a cost function over how late you learn it changed. Sometimes it is flat — knowing on Friday is as good as knowing on Tuesday. Sometimes it is a step — there is a decision point, and before it the information is worth a great deal and after it nothing. Sometimes it is roughly linear in elapsed time.

A Large Universe Model is worth building exactly where that function is steep. Where it is flat, a Large Language Model with retrieval is the correct and much cheaper answer.

This is the most useful test to apply before adopting any of this, and it is more honest than a general claim that continuous is better than static.

Steep cases

• A ship date that has slipped. Value falls sharply after the commitment is made externally.
• A vulnerability newly exploitable. Step function at the moment of first exploitation.
• A supplier heading for failure. Roughly linear — every week earlier is a week more to qualify an alternative.
• A withdrawn result underneath a research programme. Enormously steep; the cost is everything built on it since.
• A cohort drifting toward churn. Steep, because reversal cost rises with drift.

Flat cases

• Historical analysis
• Documentation lookup
• Anything decided on a schedule that cannot be moved anyway
• Questions where the answer is stable over years

The mistake is deploying continuity where the curve is flat, which produces cost and alert fatigue with no offsetting gain.

Why the value is usually underestimated

Because the cost of noticing late is invisible in the counterfactual. Nobody records the incident that would have been smaller, the reorder that would have been placed, the position that would have been reduced. The savings from continuity show up as an absence, which is the hardest thing to put in a business case.

The practical approach is to run the model in shadow first and log what it would have surfaced and when — then compare against when the organisation actually found out. That difference is the value, and it is measurable in a way general argument is not.