Dimensions and severity

The six things a check can be about, and the four levels of how much it matters.

Every rule carries two labels: what it is about, and how much a failure matters. Both come from a fixed list in the code, so no rule can invent a third.

What a rule is about

Six subjects on a chat, and every rule sits in exactly one. On a call a further one is scored, Sound, so a call carries Seven.

The name in the first column is the name on every screen. The wire carries a shorter identifier, which is the last thing on each line and is what an API answer holds.

Policy	A rule the agent is held to by someone outside the company: a regulator, a contract, or a privacy law. On the wire: compliance.
Facts	Whether what the agent said is true against a source that can be checked. On the wire: accuracy.
Answers	Whether the customer left with the thing they asked for. On the wire: resolution.
Progress	Whether the conversation moved forward: the questions asked, the steps taken, the booking reached. On the wire: progression.
Patience	What it felt like to be on the other end. Waiting, dead air, repetition. On the wire: experience.
Plain language	Whether the words worked: the right language, plain enough, and an answer to what was actually asked. On the wire: communication.
Sound	On a call, did the agent sound clear and let the caller finish speaking? Scored on a call only. On the wire: audio.

Nothing outside this list is scored. Each area starts at 100 and every failure takes a share of what is left of it, and the score is the average of the areas.

How much a failure matters

Severity names what a failure does to the customer or to the business. It does not record how firmly the source document worded the rule.

The order below is also the order used when two near-duplicate rules are merged into one. The higher level wins.

critical	Someone is harmed, misled, or owed a remedy. Legal, safety or money.
high	The customer did not get what they came for, and would notice.
medium	The outcome held but the experience did not. Worth fixing this quarter.
low	Noticeable to a reviewer reading closely, not to the customer.

The four kinds of rule

How a rule is answered, which decides what it costs to run and how far it can be trusted on its own.

deterministic	Settled by arithmetic, a timestamp or a text match. Runs on every conversation, costs nothing, and gives the same answer twice.
judged	Needs a judge to read the conversation and form an opinion. This is what the sampling share and the nightly pass exist to ration.
end_state	Asks what the conversation ended as, not what happened inside it: booked, abandoned, handed to a person.
guard	A line the agent must not cross. A single breach fails the conversation whatever else went well.

Where a rule stands

A rule reaches a score only once a reviewer has approved it. These are the states it moves through.

shadow	Scored and counted, but nobody has agreed to it yet. Run one this way to see what it would have caught before it can fail anything.
approved	A reviewer agreed to it. It scores conversations and it can fail one.
rejected	A reviewer turned it down. Kept, so the same proposal does not come back as though it were new.
superseded	Replaced by a newer revision of the same pack check. Distinct from rejected, which records a reviewer taking the rule out of what we check altogether; this one records the rule surviving under a newer row.

Note: Where a value has no line beside it, the code carries no description for it yet. We leave that gap where you can see it rather than invent a sentence here.