Docs and articles
R-55
Technology
Confidence Scores Belong on Every Fact
Half the data in a modern application is inferred by an AI model. Treating those rows the same as a manually-entered bank balance is a bug we keep shipping.
PAR2 Labs
August 29, 2026
2 min
A bank transaction is true. An AI's classification of that transaction as “likely fraud” is true with some probability. These two things are usually stored in the same column type, with no way for downstream systems to know the difference. Tessera makes confidence a first-class property of every fact.
01
The mismatch
Almost every database treats data as boolean: it's there or it isn't, it's true or it's null. That made sense when humans entered everything. It stops making sense when:
An LLM extracts an entity from unstructured text — confidence 0.83.
A computer-vision pipeline labels an X-ray — confidence 0.91.
A reconciliation engine matches two records — confidence 0.67.
A KYC service approves an identity — confidence 0.95.
Pretending these are crisp booleans is how you end up with downstream queries that treat a 0.67 match the same as a 0.99 match, until something breaks publicly.
02
How Tessera handles it
Every fact carries a confidence in the range 0.0–1.0. The default is certain (1.0) for hand-entered data; anything inferred carries the upstream model's confidence. Because confidence is a first-class part of the fact — not a property bolted on by convention — queries filter on it directly. A regulatory report can require, say, confidence at or above 0.9 and exclude every inferred row in one clause, while an operational dashboard includes the same rows without a second copy of the data.
If your data layer can't tell “the bank told us this” from “an ML model thinks this,” your data layer is lying to the people downstream of it.
03
Why this is the right move
Audit queries can exclude low-confidence inferences and still find everything an examiner needs.
Operational dashboards can include them without polluting the high-stakes paths.
You don't lose information — the inferred row stays in the store with its confidence intact, available when the right downstream consumer wants it.
If your data layer can't tell “the bank told us this” from “an ML model thinks this,” your data layer is lying to the people downstream of it.
04
What this isn't
Confidence on a fact isn't a substitute for proper model evaluation, calibration, or human review. It's a way to keep inferred and observed data in the same store without conflating them. The model still has to be good. The database stops pretending those numbers don't exist.
Key Takeaways
01
Inferred data is usually stored exactly like observed data, so downstream queries cannot tell a 0.67 match from a 0.99 one.
02
Every Tessera fact carries a confidence from 0.0 to 1.0: certain for hand-entered data, the model's score for anything inferred.
03
Queries filter on confidence directly, so an audit report and a dashboard can read the same rows at different thresholds.
04
Confidence on a fact does not replace model evaluation, calibration or human review.
PAR2 Labs · Technology
Talk to us