Rubric templates

Define what a good call looks like: slots, questions, required and forbidden phrases, case branching, thresholds, detector toggles, scoring weights, versions and AI drafts.

A rubric template (the dashboard calls it a template; the scorecard calls it a rubric) is the definition of a good call for one call type. Every call is graded against the template of its channel, or the workspace default when the channel has none. Editing a template creates a new version; calls keep the version they were graded with.

Anatomy of a template

Part	What it defines
Name and categories	Display name and the slot categories used to group the scorecard.
Slots	Values the agent must capture. See Slot capture.
Questions	Things the agent must ask, detected by phrases.
Compliance phrases	Phrases that are required (a `COMPLIANCE_MISS` if absent) or forbidden (a `FORBIDDEN_PHRASE` if present).
Case branching	Optional sub-rubrics chosen by the value of a discriminator slot.
Thresholds	Agent latency (default 8 s), dead air warning (12 s) and danger (25 s).
Detection	Greeting phrases and reset window for `CONTEXT_LOSS`, hold phrases and the announced-hold limit (60 s) that keep `DEAD_AIR` and `AGENT_LATENCY` from scoring a warned pause, and the list of disabled fault codes.
Scoring	`captureWeight` and `coverageWeight` (0.5 each by default) for the health score.

Case branching

Real call types branch. A telecom support line handles connection problems, billing questions, new orders and account changes, and each needs different slots, questions and disclosures. Rather than one bloated rubric or four channels, a template can define cases:

Pick a discriminator slot, an Enum slot such as intent or productLine.
Add one case per value, each with its own slots, questions and compliance phrases, plus aliases (keywords) that map captured values onto the case.
Cases can nest up to three levels deep, with up to twenty cases per template.

During analysis the discriminator's captured value is matched against each case's key, label and aliases. The matching case becomes active; slots and questions from the other cases are marked Not applicable and excluded from the completion and coverage rates. The scorecard shows the resolved case path.

Built-in presets

Three templates ship with VoiceLens and can be cloned into any workspace:

Preset	What it contains
Real estate enquiries	Slots for intent, property type, bedrooms, bathrooms, storage, elevator, floor preference, price, location and timeline, plus name, email and phone; every slot is `agentMustAsk`.
Attorney intake	Shared contact slots; a consent question; required recording-disclosure and no-guarantee phrases; cases for injury (with nested car and workplace sub-cases), abuse and accident, each with its own slots and questions.
Telecom and ISP support	Contact and account slots; questions to verify identity, recap and ask if there is anything else; required recording disclosure and a forbidden request for card details; cases for connection, billing, new service and account change.

Presets are a starting point: clone one, then rename slots, add your disclosures and adjust thresholds.

AI rubric drafts

Writing a rubric from scratch is tedious. Two shortcuts generate a draft for you to edit:

From text: paste a script, a QA form or a description of the call type (40 to 60,000 characters) and VoiceLens drafts slots, questions and phrases from it.
From a call: pick an analysed conversation and draft a rubric that would grade it.

Drafts are never applied automatically; you review and save them like any edit. Each draft counts against the plan's monthly allowance (3 on Starter, 50 on Pro, 200 on Agency, unlimited on Enterprise) and against a daily cap of 20 per workspace. Cancelling a queued draft refunds it.

Versions

Saving a template creates a new version. Calls record the version id they were graded with, so historical scores do not change when you edit a rubric. To apply a new version to existing calls, use Re-evaluate on a call (Admin role); the golden-set calibration does this for the whole reference set at once, without saving, so you can compare before and after.

Default and assignment

One template can be the workspace default. Channels use it unless they are assigned a template of their own. Assign templates explicitly on channels that must not change when the default does.

Writing good rubrics

Start narrow. Grade the five things that make the call actionable before adding twenty nice-to-haves.
Phrase compliance loosely. Required phrases match on meaning-bearing words; write "this call may be recorded" rather than a full legal paragraph.
Use forbidden phrases sparingly. They are danger faults; reserve them for things that are genuinely never acceptable.
Branch on captured values, not on hope. The discriminator must be a slot the agent reliably captures early in the call.
Calibrate every meaningful edit. Run the golden set before and after and watch the mean error.