How It Works

Confidence

Every automated categorization carries a confidence between 0 and 1, computed from precedent, policy match, and vendor-name similarity. At or above 0.85 it is applied; below, it is routed to a human.

The scale

Every automated categorization carries a confidence: a number between 0 and 1, stored at the moment the decision is made and never recomputed afterwards. It is a property of a decision, not of a transaction, and not of the organization.

Confidence answers one narrow question: how much support did the evidence available at decision time give this particular classification, relative to the alternatives that were considered? It is displayed as a percentage in the application, but the stored value is the fraction.

How confidence is computed

Confidence is derived from the organization's own accumulated history, not from a general prior about what vendors usually are. Three signals contribute:

Precedent share

Of the prior decisions this organization has recorded for this vendor, what proportion landed on the category now being proposed. This is the dominant term by a wide margin. An organization that has categorized a vendor the same way eleven times supplies far more evidence than one that has never seen it.

Policy match

Whether the proposed category is consistent with the organization's written rules.

Vendor-name similarity

A lexical signal on the vendor string — the weakest of the three, and weighted accordingly. It is what remains when there is no history to draw on.

A volume adjustment adds a small amount as the number of prior decisions grows, saturating quickly: the difference between one precedent and five matters, and the difference between forty and fifty does not.

The floor case is the informative one. Where an organization has never seen a vendor and no rule applies, the result sits near 0.45 — deliberately below the threshold. The system is designed to concede that a first encounter with an unknown vendor is a guess, and to route it to a person on that basis, rather than to dress the guess in a number that would let it through.

Where a classification is produced by a deterministic rule rather than inference, confidence is not the deciding factor: a rule match is applied because it is a rule.

The auto-apply threshold

The threshold is 0.85. It appears in the application as “the 85% auto-apply threshold.”

export const CONFIDENCE_THRESHOLD = 0.85;

// At or above the threshold, the categorization is applied.
// Below it, the item is routed for human review.

At or above 0.85, the proposed categorization is applied and the item does not interrupt anyone. Below 0.85, the item is routed to a person and waits. That is the entire mechanism.

An officer's override always wins, at any confidence. A high score is not a lock, and overriding a decision the system was confident about is an ordinary act that is recorded like any other — as a new ledger entry, never as an edit. See Officer Actions.

Confidence bands

For display, scores are bucketed. The bands are presentational — the threshold above is what actually decides anything:

High — 0.90 and above

Strong precedent, consistent with policy.

Medium — 0.70 to 0.89

Reasonable support. Note that part of this band sits below the threshold and is therefore still routed to a person.

Low — 0.50 to 0.69

Thin support. Always reviewed.

Very low — below 0.50

Little more than a guess, and labelled as one.

What confidence is not

It is not a probability of being correct. It is a score derived from three weighted signals. An organization that has miscategorized the same vendor eleven times will produce a high-confidence twelfth miscategorization, because precedent is what the score measures and precedent is exactly what is wrong in that case. Confidence measures support, and support is not truth.

It is not a trust score or a compliance rating. It attaches to one decision. It says nothing about the organization.

It is not recomputed. The stored value is the one held at decision time. A confidence refreshed later would be a confidence the system never actually held, which would make the record a narrative rather than a record.