# 04 — Clip Mining and Organization

## 1. Clip first principles

A clip is not just a short time range.

```text
CLIP
= bounded source interval
+ complete or deliberately incomplete meaning
+ story role
+ performance
+ visual/audio state
+ context dependencies
+ usage rights
+ provenance
+ handles
```

The system mines clips at several scales because the best short soundbite, complete answer, and full anecdote often overlap but are not identical.

## 2. Clip scales

Durations are project-configurable; these defaults are starting ranges, not laws.

| Scale | Typical range | Function |
|---|---:|---|
| Micro moment | 0.25–3 s | reaction, breath, gesture, transition |
| Hook fragment | 3–12 s | curiosity, surprise, high-density statement |
| Short soundbite | 8–30 s | one complete insight or claim |
| Contextual clip | 20–60 s | claim plus reason/example |
| Exchange | 20–120 s | question and answer, challenge and response |
| Story fragment | 45–180 s | setup, turn, partial payoff |
| Complete anecdote | 90–600 s | beginning, change, meaning |
| Chapter | 3–20 min | coherent topic or narrative movement |

## 3. Candidate generation

Create boundaries from:

- sentence and clause endings;
- breaths and pauses;
- speaker turns;
- question/answer edges;
- change in claim or story role;
- topic boundaries;
- emotional peaks;
- laughter and reaction;
- camera-safe edit points;
- music phrase boundaries;
- visual action completion.

For every promising nucleus:

```text
1. create minimal candidate;
2. test context completeness;
3. expand left/right to satisfy dependencies;
4. add source handles;
5. create alternative boundary variants;
6. score and deduplicate;
7. assign bins and graph relationships.
```

## 4. Clip score

A configurable base score:

```text
ClipScore
= 0.15 HookStrength
+ 0.14 Clarity
+ 0.13 ContextCompleteness
+ 0.11 Novelty
+ 0.11 EmotionalPower
+ 0.10 EvidenceValue
+ 0.08 SelfContainment
+ 0.07 VisualPotential
+ 0.05 PerformanceQuality
+ 0.03 Brand/AudienceFit
+ 0.03 TechnicalQuality
- RiskPenalty
- RedundancyPenalty
- RepairCost
```

The score is multidimensional. Preserve sub-scores so editors can sort by different goals.

## 5. Story-role bins

```text
00_UNREVIEWED
01_HOOKS
02_PROMISES_AND_QUESTIONS
03_ORIENTATION_AND_DEFINITIONS
04_PROBLEMS
05_GOALS_AND_STAKES
06_ORIGIN_STORIES
07_ATTEMPTS_AND_PROCESS
08_OBSTACLES_AND_FAILURES
09_EVIDENCE_AND_EXAMPLES
10_CONTRASTS_AND_OBJECTIONS
11_REVEALS_AND_REVERSALS
12_INSIGHTS_AND_MODELS
13_RESOLUTIONS_AND_RESULTS
14_TAKEAWAYS
15_CALLS_TO_ACTION
16_REACTIONS
17_BRIDGES_AND_TRANSITIONS
18_BROLL_AND_VISUAL_ONLY
19_AUDIO_ONLY
20_PICKUPS_NEEDED
21_FACT_CHECK
22_RIGHTS_REVIEW
23_REJECTED
```

Also create dynamic bins by:

```text
speaker / topic / entity / emotion / camera / source / chapter /
duration / aspect fitness / evidence state / approval state
```

## 6. Clip card

Each candidate displays:

```text
thumbnail
waveform
speaker
source and timecode
verbatim transcript
short summary
story roles
topics and entities
claim/evidence state
context-completeness score
performance score
visual suggestions
camera/audio quality
usage history
risk flags
handles
alternatives
human notes
```

## 7. Boundary variants

For a nucleus `N`, create:

```text
V0 = exact sentence
V1 = question + answer
V2 = premise + sentence
V3 = sentence + consequence
V4 = full exchange
V5 = full anecdote
```

Each boundary variant inherits the same semantic nucleus but has different:

- context completeness;
- duration;
- hook strength;
- pace;
- platform fitness;
- edit repair cost.

## 8. Context expansion algorithm

```text
WHILE completeness < threshold:
    find highest-severity unresolved dependency
    IF dependency exists immediately before:
        expand left to natural boundary
    ELSE IF dependency exists immediately after:
        expand right to natural boundary
    ELSE IF bridge can supply context without distortion:
        attach bridge
    ELSE:
        mark not-standalone
        break
```

High-risk statements require explicit human review even after automated completeness passes.

## 9. Deduplication

Two clips may be duplicates when they share:

- nearly identical words;
- same proposition;
- same anecdote;
- alternate camera views;
- repeated answer in different wording;
- generated versions of the same shot.

Store relations:

```text
DUPLICATE_OF
NEAR_DUPLICATE_OF
ALTERNATE_BOUNDARY_OF
ALTERNATE_CAMERA_OF
ALTERNATE_TAKE_OF
PARAPHRASES
```

Do not delete alternatives automatically. Rank them by the intended use.

## 10. Clip fitness profiles

### Hook fitness

Prioritize:

```text
unanswered question
+ meaningful surprise
+ high consequence
+ specificity
+ compressed conflict
+ strong performance
```

Penalize:

```text
false suspense
+ missing context
+ sensational distortion
+ generic claim
```

### Educational fitness

Prioritize:

```text
clear concept
+ mechanism
+ example
+ correct qualification
+ visual explainability
```

### Evidence fitness

Prioritize:

```text
source linkage
+ clear claim
+ dates/units/population
+ low ambiguity
```

### Emotional fitness

Prioritize:

```text
authentic performance
+ complete emotional turn
+ room for afterbeat
+ low interruption
```

### Short-form fitness

Prioritize:

```text
rapid orientation
+ standalone meaning
+ one dominant payoff
+ vertical visual viability
+ caption viability
```

### Audio-only fitness

Prioritize:

```text
no unresolved visual reference
+ intelligible speech
+ audible speaker identity
+ clean transitions
```

## 11. Bridge taxonomy

When a strong clip lacks context, add a bridge.

- host voice-over;
- on-screen text;
- included question;
- chapter card;
- date/location card;
- name/title lower third;
- document or chart;
- previous quote fragment;
- visual establishing shot;
- sound bridge;
- short generated illustrative shot, disclosed when necessary.

A bridge cannot be used to hide a material qualification.

## 12. Pickups queue

Create a pickup when no existing source can perform a required story function.

Pickup types:

- missing definition;
- unclear pronoun/reference;
- absent transition;
- unasked follow-up;
- factual correction;
- source citation;
- pronunciation;
- legal/disclosure line;
- narration;
- room tone;
- clean plate;
- insert/detail;
- establishing shot;
- reaction;
- safety/alternative framing.

## 13. Rejection reasons

- meaning incomplete;
- materially misleading without omitted context;
- weak or redundant;
- poor technical quality with no repair path;
- unverified high-risk claim;
- consent or rights unavailable;
- private information;
- defamatory framing risk;
- generated artifact;
- continuity conflict;
- unsuitable for audience or accessibility;
- no story function;
- better alternative exists.

## 14. Handles

Store at least:

```text
semantic handle — enough context before/after to preserve meaning
performance handle — breath, gesture, and reaction space
transition handle — frames for dissolves or speed changes
audio handle — room tone and crossfade material
generation handle — start/end stability for synthetic shots
```

Handle length is a property, not a fixed global constant.

## 15. Selection set logic

A sequence should not be assembled by choosing the individually highest-scoring clips only.

```text
SequenceScore
= sum(ClipUtility)
+ CoverageOfRequiredRoles
+ CausalCoherence
+ EmotionalArc
+ RhythmDiversity
- Redundancy
- Contradiction
- ContextDebt
- DurationOverflow
```

Use constrained set selection:

```text
REQUIRE one hook
REQUIRE orientation before specialized terminology
REQUIRE evidence for high-impact factual claims
REQUIRE payoff for each retained open loop
LIMIT repeated propositions
PREFER performance and visual diversity
```

## 16. Organization views

The UI should offer:

1. **Bin view** — familiar folders and selects.
2. **Transcript view** — words linked to media.
3. **Graph view** — claims, questions, setup/payoff, and dependencies.
4. **Storyboard view** — beats and coverage.
5. **Timeline view** — ordered output.
6. **Matrix view** — clips × story roles × quality dimensions.
7. **Map view** — locations and geographic evidence.
8. **Version view** — reuse across outputs.

## 17. Usage tracking

For every clip, record:

```text
where used
which source range was used
which words were omitted
which version used it
what graphic/audio covered it
whether it was reframed
whether it was generated or transformed
which human approved it
performance after publication
```

This creates a reusable media memory rather than a one-off project folder.