All capabilities
04Categorising

Categorising.

Where one category ends and the next begins is a rule people agreed on, not a fact in the data, so a classifier between two classes has nothing left to learn from. A person applies your rule to the item in front of them.

  • Your taxonomy, appliedYou send the categories. The answer is one of them.
  • Reads what the words meanCoded language and context a keyword model takes at face value.
  • Priced by the mistakeA mislabelled ticket and a missed counterfeit are not the same cost.
  • Only the uncertain onesYour classifier keeps the clear cases. People take the tail.
Your agent sends
“Designer-inspired watch, AAA quality, looks identical to the original”Marketplace listing, £45
Categories
Allowed · Counterfeit · Needs more information
Quorum’s analysis

What happened

The classifier was splitAllowed 0.52, Counterfeit 0.48.
Amara read the listingCounterfeit: “AAA quality” is how replicas are sold. 0.926.
Joel was asked independentlyCounterfeit. Together, 0.990.
Counterfeit, at 0.990.

The listing comes down under your counterfeit policy, and the decision carries its confidence.

How it works

Step by step.

The boundary cases are where a model is least sure and where the policy matters most.

01SendThe item and your categories.
02ReadSomeone proven at your kind of item.
03WeighA second look if one is not enough.
04AnswerOne category, with its confidence.

Real-world example

A replica sold in plain sight.

A moderation agent reviews new listings. The words are polite and the photos are clean, so the classifier calls it a coin toss. People who have seen the phrase before know exactly what “AAA quality” means, and the listing is handled under the policy that fits it.

Classifier: allowed, 0.52Not used
Answer: counterfeit, 0.990Removed
Trust and safetySupport triageCompliance labels

For crypto agents

Where this saves a transaction.

An on-chain payment cannot be taken back, so the moment before the agent signs is the moment to ask.

Payments agent“Our wallet has changed”A long-standing vendor emails: pay our new address from now on. Fluent, polite, the right names and invoice numbers. A routine update, or the classic payment-redirection scam? A person puts it in the right category before the next payment goes out.
DAO agentRoutine proposal, or a treasury drain?A governance proposal is labelled a parameter update but moves funds to a new address. A person decides which category it really falls in before the agent votes.
The agent holds the transaction, not the market. It asks before it signs and waits about six seconds for the answer. Set deadline_ms to the window you actually have: if nobody answers in time you are refunded, and the agent takes its safe path, which is not to sign.