Why Tessavox would build the reading, and how it would find out if it was wrong.

The business case

A streaming service would build this as a feature: it already holds each listener's recent listening, and a readable, arguable reading that ends in a prompt for the service's own licensed creation tools keeps listeners who want to make music inside the service.ASSUMED

• Tessavox already holds what the reading needs: every listener's recent plays. Nobody is asked to describe their taste, which the interviews behind this project found almost nobody can do.
• The prompt the reading ends in is pasted into Tessavox's own creation tool, trained on music it has licensed. The reading is the way in; the tool is what it feeds.
• What a listener rejects and what they choose is a signal a recommender does not collect: a person correcting a description of their own listening.

What it would measure

NORTH STAR

• Readings that end in a creation start

  Of the readings a listener opens, the share that reach the creation screen and start a track.

  It is the whole claim in one number: a reading worth having is one somebody carries into something they make.

SUPPORTING

• Kept-line rate

  Lines kept, over lines shown.

  A reading whose lines are mostly rejected is describing somebody else.

• Chosen-reading rate

  Kept lines where the listener picked one of the two offered readings, over kept lines.

  It says whether the offers land. A line kept with "neither" every time names a real pattern and misreads it.

GUARDRAILS

• Rejection rate, per template

  For each line template, rejections over showings.

  High means the template is wrong. Near zero is its own warning: a line nobody ever rejects may be true of everyone, which is the Barnum effect the receipts exist to prevent.

• Carve-out hits

  Rendered sentences that match the carve-out patterns.

  The target is zero, always: nothing on any surface asserts anything about trauma, abuse or mental health.

• Opt-out rate

  Listeners who turn the reading off, over listeners shown it.

  A feature that reads your listening back to you has to be easy to refuse, and a rising opt-out is the first sign it is unwelcome.

How it would test it

An A/B test at Tessavox's creation entry. Half of the visitors who open the creation tool see the reading and the prompt it ends in; the other half see the plain prompt box they see today.

Primary metric: creation starts per visitor who reaches the creation entry, compared between the two arms.

PLANNING ASSUMPTIONS

Every number below is a planning assumption or computed from one. None is a result.

Creation starts per visitor at the creation entry, with the plain prompt box: 20%PLANNING ASSUMPTION

We found no published rate, so nothing measured stands behind it. It is a round number chosen to make the arithmetic visible, and the sample size is shown for two others beside it.

The smallest lift worth detecting: 2 percentage pointsPLANNING ASSUMPTION

The reading adds screens before creation, and screens cost attention and maintenance. The assumption is that a lift smaller than a tenth of the baseline would not pay for them, so the test is sized to see one that size and not smaller.

False-positive rate, two-sided: 5%PLANNING ASSUMPTION

Convention, kept because a reviewer can check it at a glance; nothing about this feature argues for another.

Power: 80%PLANNING ASSUMPTION

Convention: a one-in-five chance of missing a real lift of the minimum size is the usual trade against run length.

Visitors needed in each arm, from the assumptions:

n per arm = (z₁₋α/₂ · √(2·p̄(1−p̄)) + z₁₋β · √(p₁(1−p₁) + p₂(1−p₂)))² ÷ (p₂ − p₁)²

20% → 22%: 6,510 visitors per arm

If the baseline were different:

• 10% → 12%: 3,841 visitors per arm
• 30% → 32%: 8,394 visitors per arm

THE RESULT THAT WOULD KILL IT

• The upper end of the 95% interval for the lift falls below the smallest lift worth detecting. The reading is then not worth its screens, however much people enjoy it.
• Any carve-out hit in the rendered copy, at any point in the test.
• A template rejected by more than half the listeners it is shown to is pulled, and the test continues without it.

What five departments would ask