The system's knowledge is organised into three dossiers. This is the second. Every abbreviation is spelled out on first use — Entity ID (Meta's label for a visually distinct ad), DTC (direct-to-camera), CTA (call to action), CPL / CPBC (cost per lead / cost per booked call).
When an ad wins, there are several valid ways to multiply it — and no shared way to decide which to do first. This framework fixes that: it scores every iteration path and returns a ranked action list. Source: the founder's Winner Iteration Scoring Framework (March 2026).
A winning ad's reward is not "run it harder" — it's 5+ iterations across different Entity IDs. Meta now gives near-identical ads one ticket to the auction, so the only way to scale a winner is to re-make it in genuinely distinct visual forms. The matrix decides which form to make first, by score.
Before scoring paths, the system asks one question: how often do our ads actually work? That single number decides whether we spend production on finding new winners or multiplying the ones we have. Two versions are tracked:
Amelia's hit rate is exceptionally high (early-stage, low-saturation niche). Re-calculated monthly; sliced by format and by Pioneer (new concept) vs Iteration.
| Portfolio Hit Rate | New concepts | Iteration | What it means |
|---|---|---|---|
| Below 15% (1 in 7+) | 75–80% | 20–25% | Winners are rare. Spend capacity finding them. |
| 15–33% | 50–60% | 40–50% | Winners at a decent rate. Balanced. |
| 33–50% | 40–50% | 50–60% | Strong. Shift toward iteration. |
| 50%+ (Amelia: 57%) | 30% | 70% | Drowning in winners. The bottleneck is multiplying them fast enough before they fatigue. |
Every iteration path is scored 0–10 on three axes. The composite is their average — higher means do it first.
How sure are we this iteration reads as a new Entity ID? Cross-world or cross-style change = 10 (guaranteed). Substyle = 7. Environment-only = 5. Same person/setting, new script = 0.
How many Entity IDs per unit of effort? Composite edits from existing footage score high; refilming a new day scores lower.
How fast can it be live? Composites & AI hooks 1–2 days; Skillshow 2–4 days; a refilm depends on the filming schedule.
| Priority | Path | EC | YPU | STS | Score |
|---|---|---|---|---|---|
| 1 | Composite Edit — splice hook/body/CTA across proven styles | 10 | 9 | 9 | 9.3 |
| 2 | Skillshow Transposition — winning script as an iPad slideshow walkthrough | 10 | 8 | 7 | 8.3 |
| 3 | AI Hook Injection (Kling) — AI-avatar 3-sec hook on a proven body | 8* | 8* | 8* | 8.0* |
| 4 | Different-Style Refilm — winning script in a new validated format | 10 | 6 | 4 | 6.7 |
*AI Hook Injection scores are provisional — they rise once the Kling workflow's output quality is validated. This is the path Gelo's 15 delivered IDA hooks sit on. Refilm scores lowest on speed but carries the highest long-term value: every new validated format compounds the footage library for all future winners.
Footage exists in four styles (Selfie, Studio, FAQ, Podcast) but only Selfie has a confirmed winner with spend behind it — so Gate 2 is unmet and composites stay blocked. We're in Mode 2. The ranked action list the matrix returns:
The matrix decides what to make. The rubric judges whether what we made is good enough to ship. It scores craft, flags the no-nos, and returns one of four verdicts — and you make the final call.
A niche-agnostic quality bar. Two layers: the no-nos (the patterns that make an ad bad — trip one and the verdict caps below green) and the quality bar (eight weighted dimensions, 0–100). Nothing auto-rejects on a market guess; the system flags, you decide.
Stress-tested blind against 10 real Amelia ads (5 winners, 5 losers): it caught 4 of 5 losers and the 3 lowest scores were all genuine losers — a strong filter against bad ideas. It over-rejected some winners, which drove two design choices (proof is a scored deduction not a kill-switch; format-execution fit carries real 15% weight). Weights stay Provisional until more performance data lands.
→ Open the full Great Ad Rubric (Consolidated V1) — the eight dimensions, the no-nos, the worked example, the back-test, and the glossary.
Both engines stand on two foundational strategy memos. They're the "why" behind the scores.
Meta's Andromeda delivery system clusters ads by visual similarity and gives each cluster one ticket to the auction. Above ~60% similarity, ads compete with each other instead of reaching new audiences. So the unit of scale is the Entity ID, not the ad. The Bible's multiplication matrix — 6 creative worlds × 8 styles — turns one proven concept into 19–22 distinct Entity IDs. AI-generated hooks are World 6: 10–15 fresh-face hook variations in an afternoon, spliced onto a proven body.
From the Jeremy Haynes mastermind: the old game of one "majority hook" milked across 100 iterations is dead. The new game is 25–50 genuinely distinct ads, each speaking to a specific "minority hook" — the algorithm matches each message to its small, relevant audience pocket. Campaign structure stays simple (one cold campaign, a few ad sets, 25–30 ads each); fatigue is managed with the 50/50 test then a full refresh.
Agent-facing versions of these live in standards/creative-strategy-framework (hit-rate methodology, path scoring) and standards/style-system (14 format profiles with Entity-ID signal classification). This dossier is the human-facing read.