.claude/skills/video-script-production/references/principles.md

# Principles: why the rules exist, and how much to trust each

Every rule in `scripts/rules.json` traces to one of four lessons (A to D) from the Moser Baer v3 review, or to one of two follow-ups (E, F) from the v4 review. This file says what each lesson rests on, and **how strong that support is**. Read the grades: several popular "retention laws" are craft lore or contested science, and the pipeline should not pretend otherwise.

Grades: **A** replicated or experimental, narrow scope. **B** established theory with supporting studies. **C** craft doctrine, widely used, little controlled evidence. **D** contested or failed replication: do not build on it. **Ours** = our own corpus, small n, hypotheses only.

## Lesson A: no naked data; cognitive load is a budget

> Raw data gives value, but only after the viewer has a human reason to care. A number's meaning comes from stakes or contrast.

| Support | Grade | What it says for us |
|---|---|---|
| Cowan (2001) | A, debated (3 to 5 chunks; some say 7) | A viewer holds about four things at once. Each new figure, term and name is a chunk. Budget them. |
| Mayer's multimedia principles: coherence, signaling, segmenting (via secondary summaries; one small 2022 EEG study favoured signaling) | A for learning tasks, unverified for entertainment explainers | Cut interesting-but-irrelevant material; point at the one thing that matters; deliver in segments. |
| Loewenstein (1994); Golman and Loewenstein: curiosity rises when the information seems *important* | B | A fact must be made important before it is stated. That is the "stake" step. |
| Muller (Veritasium thesis, via secondary sources): clear explainers can leave misconceptions untouched; viewers who think they already know disengage; a video that confronted the misconception felt "confusing" and taught more | B/C, secondary sources | Start from the belief the viewer holds, then break it. Some effort is fine if the effort is the point. |
| Slovic and others: "psychic numbing", identifiable victim effect | **D** | A 2024 pre-registered replication found no support with hypothetical donations; publication bias is suspected. We use human stakes as a **design choice**, not as a proven law. |
| Ours: numbers in the first 60 s went the wrong way in both "Why X died" cohorts (rho -0.32, -0.51) | Ours | Keep the hook nearly numberless. |
| Ours: both HinExp winners corrected a belief about the viewer's own life | Ours, post-hoc | Prefer stakes that land on the viewer. |

Rules that follow: hero-number budget, stake line before every hero number, at most four quantities in any 30 seconds, jargon glossed at first use, human-scale translation of big figures.

## Lesson B: one core driving loop, a destination without a direction, and a promise that out-delivers

> Find the real loop that respects the packaging promise and is the one we deliver. Hint where it ends, not how it gets there. In the first 60 seconds, escalate, so the viewer expects the video to give more than the title promised.

| Support | Grade | What it says for us |
|---|---|---|
| Loewenstein (1994) information-gap theory; Golman and Loewenstein (2022): people seek information more when the gap is important and recently opened, and avoid it when it threatens a belief they like | B | A loop needs (1) a prior the viewer already holds, (2) a visible gap against it, (3) importance. "Complete ignorance triggers no curiosity" is the usual secondary-source gloss of the 1994 paper; we could not verify the wording. |
| Brewer and Lichtenstein (1982), structural-affect theory: suspense, curiosity and surprise come from the *order* in which events are told. Curiosity = outcome shown first, route withheld. A critique says showing the outcome is neither necessary nor sufficient. | B | Destination without direction is a recognised discourse structure, not just a slogan. Show the outcome shape; withhold the causal route. |
| Hitchcock's bomb under the table (Hitchcock/Truffaut) | C | Tell the audience what is coming and let the ordinary scenes carry dread. Suspense beats surprise unless the twist is the point. |
| MrBeast production memo (leaked 2024, authenticity confirmed by two former producers; we only read secondary summaries) | C | The first minute is the main loss point; the video must prove the packaging promise early; re-engagement points around minutes 3 and 6. No source we found says "re-hook every minute". |
| Vendor blogs (vidIQ, others): confirm the promise, deliver a first payoff, chain small hook/payoff cycles | C | Same advice, no controlled evidence. Used only as agreement. |
| Zeigarnik effect ("open loops stick in memory") | **D** | Replicates inconsistently. Do not cite it as the reason loops work. Our reason is information-gap importance plus unfinished causal chains. |
| Ours: cold-opening on the collapse was CONTRADICTED in both cohorts (rho -0.47, -0.45; ledger H-005) | Ours, winners-only cohorts | v4 keeps the lockout scene under 25 seconds because the thumbnail ("LOCKED OUT") must be proven early, opens in the viewer's own world first, and leaves the cause untold. Treat this as the main tension between a lesson and our data; the next cohort should test it (ledger). |
| Ours: three high-view Moser Baer videos opened on a human scene (3 of 3 vs 1 of 4 small ones) | Ours, n=7 | Consistent with a human-scene opening; could be channel size. |

Rules that follow: a declared core driving loop with a destination hint and a withheld list that must not leak into the hook; three-rung promise ladder (title promise, a stranger answer, a use the viewer can take); at least one unannounced bonus; first payoff inside 30 seconds.

## Lesson C: a causal chain: BUT and THEREFORE, never AND THEN

> If the link between two sections reads "and then", the story is flat. It should read "but" or "therefore".

| Support | Grade | What it says for us |
|---|---|---|
| Parker and Stone, NYU class (2011): replace "and then" between beats with "but" or "therefore" | C | Craft rule from working writers. Primary source is the class clip, which we did not open. |
| Frank Daniel (USC, 1986), via secondary sources; Randy Olson's ABT template (And, But, Therefore; TEDMED 2013; *Houston, We Have a Narrative*, 2015) | C | "And" may set shared context (the first beat). Between beats, the link must be a complication (BUT) or a consequence (THEREFORE). |
| Our own rubric already had K3 ("more than half of transitions are therefore/but") | Ours | v4 tightens it to every transition and makes it checkable from a declared chain. |
| Ours: causality scores were not a differentiator among winners (range restriction) | Ours | Do not expect to prove this on winners-only cohorts. |

A third link type, **BECAUSE** (the next beat explains the previous), is allowed because it is a causal link read backwards.

Rules that follow: every junction declared and typed; hinge line present in the text; no additive opener; the swap test; the dependency (shared noun) test; a BUT must reverse an expectation the previous beat planted.

## Lesson D: Understand, Care, Change, Compel

> Show something to understand. Make them care why it matters. Change what they now believe. Give them a reason to continue: force it by closing a loop with a bigger one.

| Support | Grade | What it says for us |
|---|---|---|
| Information-gap theory (above): importance, recency | B | Compel works when the next question is *created by* the answer just given. |
| Harmon's Story Circle: the "need" step carries the stakes (secondary sources) | C | Establish what is at stake before the action. |
| Narrative transportation, self-reference effect (already in our taxonomy) | B | Care is easier to build with a person or the viewer's own life. |
| Ours: 4.4 loops open at once in winners (observational), but loop-timing raters disagree (0.08 to 0.39) | Ours | We cap *major* live loops at three because of Cowan, and declare loops in the script so they can be checked without rater judgment. |

Rules that follow: each beat declares its four parts with quotes; the last two sentences of a beat open a question that the next beat answers; the spine loop stays open until the last quarter; no uncovered stretch longer than 45 seconds.

## Lesson E: a pull between the questions (implicit open loops)

> The open loop does not have to be explicit. Something implied can keep the viewer watching, as long as it is there for the whole script.

See `implicit-pull.md` for the types, the evidence and the placement method. Short version: suspense research supports concern for a person under threat and a shrinking set of ways out (Gerrig and Bernardo, grade A for fiction; Brewer and Lichtenstein, B); dramatic irony and the Zeigarnik effect are craft lore and a contested effect (C and D). Rules that follow: at least four declared pulls of at least three kinds, each touched in the text at least twice, no stretch over 45 seconds with nothing pending.

## Lesson F: plain spoken voice

> Scripts that read as machine-written lose the listener even when the facts are right ("no X, no Y, just a Z").

See `voice.md`. The catalogue comes from Wikipedia's *Signs of AI writing* and two public skills (blader/humanizer, hardikpandya/stop-slop); Wikipedia itself says the signs are not proof and that humans use them. Grade **C**: a craft judgement about repetition, not a tested effect. Rules that follow: budgets (not bans) on negation pivots, staged run-ups and script-talk, stock phrases, drumroll rhythm, and unspeakable written English.

## Lesson G: a personality that travels, and figures that explain

> The voice should be both earnest (follows leads, researches, is honest about gaps) and aloof (unbothered by who wins, finds the case for and against on every side), and it should explain with metaphors, similes and parables.

See `texture.md`. The supported parts are one early, novel, apt metaphor (Sopory and Dillard 2002, grade B), saying where an analogy stops (Gentner; Brown and Salter, B) and two-sided refutational treatment (Allen 1991, B). The persona itself has no experimental support in explainer channels (grade D) and the English winners we measured use almost no figures of speech, so this is a deliberate bet. Rules that follow: both registers present early and late; contested calls told both ways; exactly one governing metaphor that returns and states its own limit; glosses by analogy for hard words; a parable with no figures in it; a budget on comparisons, stock figures and verdict words.

## Lesson H: rhythm, hard moments and easy ones

> Data insight moments should alternate with easy-to-follow storytelling moments.

See `texture.md`. Attention and encoding rise with pacing and arousal until overload and then fall (Lang's LC4MP, grade B, mostly lab and TV); the "anecdote and reflection" and "what is / what could be" structures (Glass, Duarte, grade C) describe the same alternation. We measure it in 30-second windows and calibrated the thresholds on eight English winners (a small sample). Rules that follow: a cap on consecutive number-heavy windows, a floor on quiet windows, no two dense beats in a row, a landing (a person, the viewer or a picture, no figure) after each hero number, and a quiet human close.

## Lesson I: the human element

> Personify. Find the personal stake, the viewer's (implied) or a character we care about.

See `texture.md`. Narrative moves beliefs and intentions a little (Braddock and Dillard 2016, B) and "you" lines aid memory (self-reference effect, A for memory). The identifiable-victim effect is close to zero in the largest studies and replications (Lee and Feeley 2016; later work) and the mirror-neuron explanation of empathy is contested (grade D), so the design relies on a **want, a concrete detail, a stake and presence**, not on a single victim. Rules that follow: a protagonist, a witness and the viewer; the protagonist named in the hook with a want; no hinge line that leans on a pronoun nobody has named; nobody and no "you" absent for more than a minute; viewer stakes in the hook, the middle and the close.

## Lesson J: tell the whole story

> The v3 and v4 scripts truncated the story: the founder, the city, the partner, the rise and the human nuance were cut. Tell all of it.

The user's request that produced v5 (made after reading v4). It is craft, not evidence (grade C): a story that skips its origin, its mechanism and its people gives the viewer little to care about (Lessons A and I) and nothing to be surprised by. Rule that follows: each beat declares the part of the story it tells (`beats[].arc`: origin, rise, turn, response, fall, people, blame, lesson, plus hook, peak and verdict); all the required parts must have a beat, the origin within the first two beats after the hook, the lesson in the last two, and the parts in a sensible order. The default list suits a company-collapse story; a different video sets `chain.arc_required`. The rule forces the author to plan for completeness; it cannot judge whether a beat is rich.

## Lesson K: nothing the viewer has to decode

> "Who is Ratul?" "That gate is not set up." "No one talks like that." "Why is Xidex here?" Every person, term, count and sentence should be understood the first time it is heard.

From the Hinglish v6 review (2026-10-25, `review-round-7.md`). The mechanism is cognitive load: working memory holds a handful of items (Cowan, already cited under A), a new unlabelled name takes a slot and gives nothing back, and signalling the structure of what comes next lowers the cost of following it (the signalling principle: Alpizar, Adesope and Wong 2020, d = 0.38 across 29 learning experiments, abstract-level; advance organisers are positive on average with wide variation, a 1992 update, with earlier reviews disagreeing). Those are studies of instructional media, not of entertainment explainers, so the grade is **B-** for the mechanism and **C** for this use. The rules are craft built on it. Parts:

1. **Cast tags.** Everyone named gets a role or a relation the first time ("Ratul Puri, Deepak's son and a company director") and again when they return after about a minute or a beat boundary. The protagonist is exempt after the first tag. (K1, K2)
2. **Name rent.** A new person or organisation per beat is a cost (K3); a name that appears once must have a declared reason or be cut (K4). Xidex was a billion-dollar trivia with no tie to the story.
3. **Counts are backed.** A spoken count ("three pivots", "three questions", "three problems") is either a promise whose items come after it or a tally whose items came before it, and the number matches. (K5)
4. **A named idea is defined once, in plain words, at its first use, and used again** (pivot, scale, the treadmill). A label used once is decoration. (K6)
5. **Do not gloss what the audience knows** (crore, CBI for an Indian business audience). The audience is declared in `chain.audience`; a different audience moves the term back to jargon. (K7)
6. **Plain, concrete, speakable.** No vague referent ("whatever was running out"); the verb next to its subject, not behind a long aside; a concrete noun for a category ("the pirated copy", not "free"); the word people say (feeling, busy, pivot). A demonstrative has its antecedent in the sentence before. (K8 and the checklist)

## Lesson L: stakes at the turns

> The fall has weight only if the rise was felt; the turn must be heard; where a person pays, the line should say what it cost.

From the same review. Craft (grade C) with the support of the contrast and two-sided evidence already cited under G; the claim that a stated cost raises emotion is a design bet. Parts:

1. **The mirror shows the real association**, including the slightly guilty one (a pirated movie on a blank disc). (checklist 20)
2. **The prize, then a contrast word.** The hook states the winning in the audience's own terms ("made serious money") and the fall opens on a contrast ("But one morning"). (L1)
3. **Where a person pays, an intensifier of the cost** ("a full 17 years", "part of their own pay"). (L2)
4. **Character through a deed.** The hook's last line is a puzzle about a person's decisive act ("the day Deepak Puri walked away from a business that was working. Why would he?"), not about a place; it claims no manner (calm, cold) that no source gives. (checklist 22)
5. **Each beat is a small arc: doubt, the edge, the proof**, then the next problem. A beat that has the doubt and the success but not the person's edge gives the viewer nothing to understand. The edge must be one the sources support. (checklist 23)
6. **A limit carries the want** ("only goes so far, and he wanted something new").
7. **A BUT is spoken.** A junction typed BUT has a contrast word within two sentences either side. (L3)
8. **Restate the height before the fall point** ("while the company was still a champion, with its fate in its own hands"). (checklist 24)
9. **Aloof about people, plain about mechanisms.** A neutrality disclaimer about something the record supports ("I have no favourite theory") leaves the viewer with nothing to change their mind about. State what the record supports and what the popular story skips; keep the disclaimers for verdicts on people and for what is unproven. (L4)
10. **A cliffhanger lands in the viewer** ("you'd do a double take"), not in an unnamed crowd, so no unsourced reaction is claimed. (checklist 25)

## Lesson M: density, a headline first

> Sometimes the script can become too dense. Give a one-line summary of the events to follow (it attracts curiosity, opens a loop, gives something to understand), then use the following lines to fill in the details, close the loops, make people care about the human story and the stakes, and give them a reason to continue.

From the same review (the user's own strategy). It uses the signalling and advance-organiser evidence cited under K (B- for learning, C for story video) and the loop mechanism of Lesson D (the headline opens a loop that the details close). Parts:

1. **A dense beat opens with a headline**: one sentence, at most 20 words, within the first two sentences, no figure in it ("What happened took half a year: the company went from saying it was well-funded to asking for time."). A beat is dense at seven or more figures or four or more new names. (M1)
2. **The hook ends in a trailer, not a menu.** The tease is a list of fragments with the strongest last (at most 30 words); the takeaway is one short sentence that keeps the "you" (at most 14 words). From the moment the story has been told, a viewer will not wait for fluff. (M2)
3. **No lab voice.** The narrator does not announce a test ("if that is right, you would expect..."); the story shows the sign ("But the first signs were starting to show"). (M3)

## What this does not prove

None of lessons A to M has been tested in our corpus on weak versus strong performers. They are design constraints, not findings. The pipeline's job is to make them cheap to follow and cheap to test: the linter enforces them, `script-cohort-analysis` rubric v3 measures them blind, and `references/learnings.md` records what each test shows. If a rule is contradicted in two cohorts, change `rules.json`; do not defend it.