# Score and select moments

Use this when a decision model scores many candidate moments. A decision model is any model that answers a fixed question with one answer from a set list and a confidence from 0 to 1. Any model that does this can be the judge. Keep the questions in a rubric file so every run asks the same thing.

Every number on this page is a starting default. Tune it to your own goal and your own picks. Change a weight only after a logged test or a calibration pass gives you a reason.

## Three kinds of question

| Kind | Answer set | What it does |
| --- | --- | --- |
| Gate | Yes or no | Drops the moment when the answer is the drop answer. A gate never adds points. |
| Quality | Yes or no, or a three-level scale | Adds points with a weight. A negative weight takes points away. |
| Category | A label from a list | Sorts moments for filters and test reports. A category never adds points. |

Keep accuracy and honesty checks as gates. A true clip is not a better clip. It is the minimum.

Write each three-level scale as three situations, from worst to best. Describe what is present in each one. Do not use "weak", "medium", and "strong". Score the levels 0, 0.5, and 1. Score yes as 1 and no as 0.

### Read a gate

- Drop the moment when the model gives the drop answer with a confidence of 0.6 or more (default).
- When the model gives the drop answer with a lower confidence, send the moment to a person.
- When an optional input is missing, the gate answer is "no fault". For example, a clip with no caption copy cannot invent a number in its caption.

### Clip gates

Drop a moment when any of these is true:

- A new viewer needs the sentence before the cut to understand the first sentence.
- The point needs a screen or demo that a vertical crop cannot show.
- The cut makes the speaker seem to mean the opposite of the full context.
- The first line uses "that", "he", "they", or "it" with nothing to point to.
- The central thought stops before its payoff.
- The cut starts inside the edited cold open.
- The cut starts or ends inside a sentence.
- The payoff refers to an event that only the earlier context names.
- A stated fact conflicts with the verified source notes.
- The cut drops the actual answer only to save time.
- The caption or label shows a number, a result, or a list that the clip does not give.

## Core and secondary scores

Pick five core scores. Each one is a three-level scale. The defaults are:

1. **One idea.** The cut asks the viewer to follow one idea from start to result.
2. **Reach.** The want or consequence matters beyond people who practise the guest's craft.
3. **Stands alone cold.** A viewer who saw nothing earlier can follow the premise.
4. **Payoff lands.** The last sentences deliver what the start set up.
5. **Surprise or tension.** The point goes against a likely belief, or the stake is clear.

All other quality questions are secondary. Examples are a want in the first line, a stake for the viewer, a named object, and a verified number. Others ask whether the clip explains how the result happened and whether a viewer can retell the ending in one sentence. A product walk-through takes points away.

## Composite score

1. **Adjust for low confidence.** When the model answers with a confidence under 0.5, move that answer halfway to the neutral value of 0.5. A "yes" (1) at confidence 0.4 becomes 0.75.
2. **Core.** Take the geometric mean of the five core scores. Multiply them and take the fifth root. One low core score pulls the whole core down, and a core score of 0 makes the core 0. An average would let four strong scores hide one failure.
3. **Secondary.** Add each secondary answer times its weight. Scale the sum to 0 to 1: subtract the lowest possible sum, then divide by the range between the lowest and highest possible sums.
4. **Final.** Final = 0.75 x core + 0.25 x secondary (default weights).

### Worked example

These numbers are invented to show the math.

| Moment | Core scores | Core | Secondary | Final | Result |
| --- | --- | --- | --- | --- | --- |
| A | 1, 1, 0.5, 1, 0.5 | 0.76 | 0.60 | 0.72 | Passes the core floor |
| B | 1, 1, 1, 0, 1 | 0 | 0.80 | 0.20 | Fails: the payoff never lands |
| C | 0.5, 0.5, 0.5, 0.5, 1 | 0.57 | 0.80 | 0.63 | Fails the core floor |

Moment B shows why the core uses a geometric mean. A high secondary score cannot save a clip with no payoff.

## Select the clips

1. Remove every moment that fails a gate.
2. Keep moments with a core of 0.65 or more (default core floor).
3. Of those, keep moments with a final score at or above the 60th percentile of the gate-passing moments in this episode (default cut). The cut adapts to each episode, so a strong episode does not flood the slate and a weak one does not pass weak moments.
4. Keep the best moment from each group of overlapping spans.
5. Send moments that are mostly screen to the wide format only.
6. Cap the picks for each topic so one subject does not fill the slate.

Never fill below the floor. If three moments pass, return three. You may keep the next few below-floor moments on a bench for tests and calibration. Do not post them to reach a count.

## Long clips

A horizontal clip of 1 to 12 minutes needs more than a short clip. Add these gates. Each one drops the clip when the answer is "no":

- The clip has a setup, a development, and an answer to its promise.
- Each part adds context, evidence, a useful step, or a result.
- The end delivers the promise without the full episode.
- A useful detail comes early. The defaults are about 10 seconds for a 1 to 3 minute clip and about 20 seconds for a longer clip.
- No dead stretch runs too long. A dead stretch adds no new context, evidence, or result. The defaults are about 8 seconds for a 1 to 3 minute clip and about 10 seconds for a longer clip. A pause alone is not a dead stretch.
- The support fits the type. A tactic needs an action. A story needs a result. A claim needs a named source and a reason.

Give double weight (default) to three secondary checks. Each minute gives a reason to keep watching. The clip moves clearly from one point to the next. A new viewer can follow any demo in the full frame.

## Editorial checks before release

A person answers these for each picked clip and its title. Each one must be "yes". "Not sure" does not pass.

- The source supports every claim in the title and hook.
- A new viewer can tell the subject.
- Names, conditions, negations, and uncertainty stay as the speaker said them.
- The clip delivers its promise, and the ending holds a result.
- The frame keeps every fact the viewer needs.
- The clip is different from the others in the slate.

## Tune the scores to your goal

Each goal needs different picks. Change the weights, add questions, or move a question from secondary to core.

| Goal | Weight up | Weight down or add |
| --- | --- | --- |
| Views from strangers | Reach, surprise or tension, a stake in the first line | Add a negative weight for craft-only detail |
| Comments from the right buyers | A problem the buyer has, firsthand experience, a specific example | Weight reach down. Add "does this matter to the buyer you want?" |
| Sign-ups or full-episode starts | A payoff that opens a next question, an action the viewer can take | Add "does the next step follow from this clip?" |

Write a new question the same way as the others: one thing to judge, set answers, and situations instead of degrees.

## Calibrate against your own picks

1. Take 20 to 30 candidate moments from two or three episodes.
2. Mark each one Keep or Skip yourself before you look at the scores.
3. Run the model. Compare its top picks with yours.
4. Read every disagreement. Find the question that caused it.
5. Fix the question wording first. Change a weight only when the wording is clear and the disagreement stays.
6. Run again. Stop when most of the model's top picks match yours.
7. Check the result on an episode you did not use for tuning.

Repeat a short pass when your goal changes or when test results stop matching the scores. Log each change and the reason for it.