> ## Documentation Index
> Fetch the complete documentation index at: https://docs.highailabs.com/llms.txt
> Use this file to discover all available pages before exploring further.
# The 100 — Test Design
> One hundred scripts built as a designed experiment rather than a content dump: four arms, fourteen formats, five controlled pairs, and the rules for what gets killed and what gets funded.
## What This Is
One hundred new scripts, and **not one of them is an original idea.**
Every format here is lifted from something that already worked at scale and is documented in the [video teardowns](/social/teardowns), the [landscape](/social/landscape), or the [mascot playbook](/social/mascot-playbook). The invention is in the *arrangement*: enough repetitions of each format to tell a real result from a lucky one, split across four arms that are scored on **different metrics**, with ten scripts spent on controls that make the other ninety interpretable.
**The point is not one hundred videos. The point is a readable answer.**
A hundred scattered one-offs produces a hundred anecdotes. A hundred structured reps produces a ranked list of formats, a set of library rules, and a defensible decision about where the next thousand go.
## The Four Arms
The [teardowns](/social/teardowns#the-ten-patterns) contain the finding that shapes this whole design:
> **Shares and likes are different products.** Cheap + relational = shares. Dense + quotable = likes. **Pick one and design backwards.**
So the slate does not have one goal. It has four, and each arm is judged only against its own.
```mermaid theme={null}
flowchart TD
S["THE 100"] --> A1["ARM 1 — REACH
45 scripts
metric: views + shares"]
S --> A2["ARM 2 — PARTICIPATION
25 scripts
metric: comments per view"]
S --> A3["ARM 3 — TRAFFIC
20 scripts
metric: saves + site sessions"]
S --> A4["ARM 4 — CONTROL
10 scripts
metric: the difference between twins"]
```
| Arm | Scripts | Optimised for | Judged on | Failure looks like |
| --------------------- | ------- | --------------------------- | ----------------------------- | ------------------------ |
| **1 — Reach** | 45 | Being seen and sent | Views, share rate | Nobody sends it |
| **2 — Participation** | 25 | Being argued with | Comments per view | A dead comment section |
| **3 — Traffic** | 20 | Being reopened and acted on | Saves per view, site sessions | Saves that never convert |
| **4 — Control** | 10 | Nothing | Difference between twins | An ambiguous result |
**Do not judge an arm on another arm's metric.** The Counter Kit will lose on views to every comedy pack in the library and that is not a failure — it is education grammar, which the teardowns show is [the suppressed lane](/social/teardowns). If it earns saves at a materially higher rate, it has done its job. Ranking all one hundred scripts on a single number would erase exactly the finding this design exists to test.
## The Fourteen Formats
Every pack traces to a documented finding. The right-hand column is the sourcing check: if a pack cannot name the evidence it came from, it should not have been written.
| # | Pack | Scripts | Arm | Comes from |
| -- | ------------------------------------------------------------------ | ------- | ------------- | --------------------------------------------------------------------------------------------------- |
| 1 | [Edible Ain't Shit](/social/scripts/the-100/edible-aint-shit) | 10 | Reach | The format the teardowns name as our **signature bit** — two drawings, infinitely re-skinnable |
| 2 | [The Nostalgia Lab](/social/scripts/the-100/nostalgia-lab) | 8 | Reach | Pattern 1 — *"the product got better and the culture got worse"*, three independent high performers |
| 3 | [The Paranoia Ward](/social/scripts/the-100/paranoia-ward) | 8 | Reach | Pattern 7 + the seven-year accumulation curve on **"Weed Makes You Scared"** |
| 4 | [Formal Systems](/social/scripts/the-100/formal-systems) | 9 | Reach | Pattern 8 — the **45.2M** video is a police procedure, not a stoner joke |
| 5 | [Immovable Punchline](/social/scripts/the-100/immovable-punchline) | 5 | Reach | Pattern 6 — the **2.8M** fixed-line specimen |
| 6 | [Seven-Second Cards](/social/scripts/the-100/seven-second-cards) | 5 | Reach | The **11% share rate** on 7 seconds and two cards — highest in the library |
| 7 | [The Taxonomy](/social/scripts/the-100/the-taxonomy) | 8 | Participation | Pattern 2 — claimable labels ran **3,900-16,900 comments** vs **52-370** |
| 8 | [The Rotation](/social/scripts/the-100/rotation) | 6 | Participation | The highest-participation documented format — a casting exercise with no punchline |
| 9 | [Outsider Confession](/social/scripts/the-100/outsider-confession) | 5 | Participation | **7.7M** — the decode outperformed the entire format it decoded |
| 10 | [The Thought](/social/scripts/the-100/hits-blunt) | 6 | Participation | A **falsification test** — the research says this is a caption format, not a video format |
| 11 | [Sections](/social/scripts/the-100/sections) | 6 | Traffic | A timeline made spatial — *"a joke and a product feature wearing the same clothes"* |
| 12 | [The Counter Kit](/social/scripts/the-100/counter-kit) | 7 | Traffic | Pattern 10 — **utility earns saves; jokes do not** |
| 13 | [Look It Up](/social/scripts/the-100/look-it-up) | 7 | Traffic | The [funnel's](/social/funnel#the-traffic-engine) own mechanic: a question only a lookup answers |
| 14 | [The Control Pack](/social/scripts/the-100/control-pack) | 10 | Control | Five matched pairs, one variable each |
## Why Reps, Not Variety
The single most common way a content test fails is spending one video per idea and reading the winner as a finding.
TikTok distribution is extremely high-variance at low follower counts. A format can produce a 400-view video and a 40,000-view video **from the same account in the same week** — so a one-shot comparison mostly measures which video happened to catch a pocket of the feed, not which format is better.
```mermaid theme={null}
flowchart LR
A["1 video
per format"] --> B["You learn which
video won"]
C["5-10 videos
per format"] --> D["You learn which
format won"]
```
**This is why the slate has fourteen formats and not forty.** Forty formats at 2-3 reps each would look more ambitious and would tell us nothing. Fourteen at 5-10 reps produces a median per format, and a median is a thing you can act on.
The two smallest packs — [Immovable Punchline](/social/scripts/the-100/immovable-punchline) and [Seven-Second Cards](/social/scripts/the-100/seven-second-cards) at five each — are the ones where each rep is cheapest to add later if the early signal is good.
## Release Order
Every pack has a designated **baseline** episode, marked in its own test-variable row. That is not decoration — it is what makes the wave structure possible.
```mermaid theme={null}
flowchart TD
W1["WAVE 1 — BASELINES
weeks 1-3 · 14 videos
one per pack"] --> W2["WAVE 2 — DEPTH
weeks 4-13 · ~60 videos
reps into what cleared the bar"]
W1 --> W3["WAVE 3 — CONTROLS
weeks 5-13 · 10 videos
interleaved, 7 days between twins"]
W2 --> W4["WAVE 4 — CAPSTONES
weeks 14-16 · ~16 videos
the ones needing an established audience"]
W3 --> W4
```
### Wave 1 — Baselines (weeks 1-3)
Fourteen videos, one per pack. The cheapest possible read on which formats deserve depth, before any of them has consumed a production budget.
### Wave 2 — Depth (weeks 4-13)
Reps go into packs that cleared the bar in Wave 1. Packs that did not clear it are **held, not deleted** — a single baseline is weak evidence, and a held pack can be revived cheaply if a neighbouring format succeeds.
### Wave 3 — Controls (weeks 5-13, interleaved)
The five pairs run alongside Wave 2, seven days between twins, never two twins in one week. They occupy slots that would otherwise be reps and that cost is deliberate.
### Wave 4 — Capstones (weeks 14-16)
Sixteen scripts across the slate are explicitly written to run **after** their pack has established something. Posting them early spends a payoff on an audience that has not built the expectation:
| Capstone | Requires |
| --------- | ----------------------------------------------------- |
| `EAS-010` | Nine reps of the reversal, so breaking it lands |
| `PAR-008` | Five comedy episodes, so the straight one converts |
| `TAX-008` | Five axes established, so the chart means something |
| `CTR-007` | Six utility videos to compile |
| `ROT-006` | Real audience rotations to read back |
| `HIT-006` | Five thoughts that turned out real, so "No." reverses |
### Cadence
| Lane | Per week | Packs |
| --------- | ------------- | ------------------------------------------------------------------------------------------ |
| **Main** | 3, fixed days | The animated packs — matches the [mascot playbook](/social/mascot-playbook) operating spec |
| **Cheap** | 2 | Seven-Second Cards, The Thought, Sections — near-free, do not consume the animation buffer |
Five a week across roughly sixteen weeks. The main lane is capped by the operating spec, not by script supply — which is exactly why six of the fourteen packs were designed to be near-free to produce.
## Decision Rules
**We have no benchmarks and we must not pretend otherwise.** Every figure in the research is someone else's account, at a different size, in a different niche. Nothing in the library has been tested against a real audience.
So nothing here is scored against an absolute number. **Everything is scored against our own running median** for the same arm, which begins to exist after Wave 1 and gets more trustworthy every week.
### Per video
Assessed at the **48-hour snapshot**, using the existing [hook thresholds](/social/scripts/hook-library#reading-hook-results). 48h is the earliest checkpoint the [performance log](/social/scripts/testing-framework#the-performance-log) actually records, so every verdict here is reproducible from the source of truth:
| 3s hold rate | Read | Action |
| ------------ | ------------ | ---------------------------------------------------- |
| Above 70% | Hook working | Reuse the family aggressively |
| 55-70% | Acceptable | The hook is not the bottleneck — look at the middle |
| 40-55% | Weak | Re-cut with the B hook before touching anything else |
| Below 40% | Broken | Rewrite. Do not re-cut |
### Per pack
Assessed once a pack has **at least three** posted reps, against the running median of its own arm:
| Result | Verdict | Action |
| ------------------------- | ------------ | ---------------------------------------------------------------------------- |
| Median ≥ 2× arm median | **Scale** | Fund remaining reps immediately. Consider promoting to a recurring franchise |
| Median 1-2× arm median | **Continue** | Run the planned reps. Re-assess at completion |
| Median 0.5-1× arm median | **Hold** | Stop after the current rep. Revisit only if a related format scales |
| Median \< 0.5× arm median | **Kill** | Stop. Record why. Do not quietly keep making them |
**"Kill" means stop posting, not delete the scripts.** A format that fails against a 2,000-follower audience may work against a 50,000-follower one, because the feed is being asked a different question. Killed packs stay in the library with a dated note explaining the result.
### Per arm
| Signal | What it would mean |
| -------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Reach arm dominates on every metric | Comedy is the whole business. Move the traffic packs into captions and stop making them |
| Participation arm out-comments by 3×+ | Claimable labels are the growth engine. The Taxonomy becomes a permanent weekly slot |
| Traffic arm converts despite low reach | **The most commercially important outcome available.** A small audience that acts beats a large one that scrolls, and the library needs a permanent utility lane |
| Nothing separates | The variables that matter are not the ones we tested. Go back to hooks and posting time |
## What Each Arm Would Have To Prove
| Arm | The hypothesis it is testing |
| --------------------- | -------------------------------------------------------------------------------------------------------------------- |
| **1 — Reach** | That formats which worked for live-action human creators still work for an animated mascot with a compliance ceiling |
| **2 — Participation** | That a claimable label beats a good joke for comment volume — the two-order-of-magnitude claim in Pattern 2 |
| **3 — Traffic** | That saves and site sessions can be bought at an acceptable reach discount, given education grammar is suppressed |
| **4 — Control** | That five specific production decisions we currently make on instinct are worth what we assume |
## Honest Gaps
These are stated here so a reader can discount the design appropriately rather than discovering them halfway through.
| Gap | Consequence |
| --------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| **No audience yet** | Wave 1 runs against near-zero followers. Early medians will be noisy and should be re-baselined once the account has scale |
| **Site sessions are not visible in-platform** | Arm 3 produces no readable result unless every Look It Up video ships with its own tracked link. This is the single easiest way to waste twenty scripts |
| **Four roles need casting** | The Old Head, The Clerk, The Outsider and The Witness are uncast. They do **not** need building — any Cantina character matching the [brief](/social/cast) will do — but a pack cannot start until its role is cast and written down |
| **One pair is a coin flip** | Every control pair needs a second run with a different fact before its result becomes a rule |
| **Historical potency figures are unsourced** | Every `[DATA: ...]` token in the Nostalgia Lab must be filled from a real source or those beats do not ship |
| **Sixteen weeks is a long feedback loop** | If Wave 1 is ambiguous, the right response is to extend Wave 1, not to push into Wave 2 on a hunch |
## Related
Every prompt in the library, in one page.
Where every row gets logged.
The evidence every pack traces to.