What a Defensible Trigger Evaluation Has to Record

Parametric structures are attractive partly because they are objective: a threshold is crossed or it is not. That objectivity only survives contact with a dispute if the evaluation was recorded well enough to be reconstructed.

Here is the minimum we think an evaluation record needs, and why each field earns its place.

The source, named specifically

Not "NOAA" but which product, from which endpoint. Agencies publish multiple series that disagree with each other legitimately — operational versus reanalysed, provisional versus final. "The NOAA value" is not a citation. A URL that resolves to the specific record is.

This is the same standard applied to every ingested record in our own data: source, source URL, ingestion timestamp. A row that can be traced to a script but not to the document it came from is an assertion, not evidence.

The observation timestamp, distinct from evaluation time

When we evaluated and when the measurement was taken are different facts, and the gap between them is exactly the latency component of basis risk. Collapsing them into one timestamp destroys the ability to audit that gap afterwards.

The value and the threshold, together

Storing "breached: true" discards the margin. Storing the value alongside the threshold preserves how close the call was, which is what anyone reviewing a contested evaluation actually wants to know. It also makes approach-to-threshold monitoring possible at all — you cannot alert at 80% of a threshold you did not record.

Not-breached and not-evaluated, kept apart

This is the field most often missing, and the most consequential.

If a data source is unreachable, a system that records "threshold not reached" has stated something it does not know. The two states look identical downstream: a quiet monitor and a broken monitor produce the same dashboard. And an outage is most likely during a major event, because that is when load is highest — so the failure and the condition it conceals share a cause.

An evaluation record should be able to say "unknown", and any system consuming it should treat unknown as materially different from negative. We would rather return an explicit error than a confident null.

Why record all this if it is rarely read

Because the cost is trivial in advance and unrecoverable afterwards. The moment an evaluation is questioned is the moment it can no longer be improved. Everything above is cheap to write at evaluation time and impossible to reconstruct later from a boolean.

It is also, in practice, what a compliance review asks for. Not the model. The trail.