Patterns & backtest

Winner pattern ledger

9 patterns · click a row for its evidence panel · patterns cite tests, no vibes

Strength

Status

Scope dimension

Provenance

Sort

Search statements

Id	Pattern · scope	Strength	Status	Evidence (live query)	Half-life
WP-0004	Scale follows the advertorial. High-copy winners only exist behind "N-reasons" pre-sell pages. lpall personasLP01-04 vs PDP-direct	proven	active	seed: 187- & 111-copy winners both behind reasons pages, ~80% of paid traffic baseline · TTM Meta-CAC $36.81-40.10 advertorial vs $43.76 PDP-direct seed est. · live: CON-0007 graduated through LP01	confirmed 06-26 · decays 09-24	evidence
Cited: seed decode (298 of 400 ads routed to reasons LPs) + TST-0001 (LP path held constant in all W25-26 tests) + CON-0007 read (LP01 read-through 34% ≥ 30% gate). n: 400 seed creatives · 48 live · ~$1.41M seed spend baseline. CAC vs baseline: advertorial-routed −9% to −16% vs PDP-direct (TTM, Meta basis) seed est. Counter-evidence: none recorded. Briefing default: hard LP gate in the EV scorer — a cell whose offer has no matching advertorial is BLOCKED from queue entry and auto-fires an LP-build request (currently: O03 motor-skills → ISS-016, 3 cells blocked). Spawned: ISS-016 LP build; LP02 claim-stack module (via WP-0007). Provenance: seed-decode, live-confirmed W26.
WP-0005	The stacked hook formula is the account spine — category-king headline + rush line + urgency close; every concept ≥118 days live runs a variant. hookaccount-wide	proven	active	seed: all 6 longevity-ranked winners baseline · live: 7/7 decided reads used the skeleton, only cold-opens varied	confirmed 06-26 · decays 09-24	evidence
Cited: Foreplay decode verbatims ("#1 … Kit For Toddlers" / "…rushing to get these like crazy!") across every hero concept baseline + all 7 W25-26 pooled reads. Counter-evidence / epistemic note: the skeleton is held constant by design, so it is never the tested variable — confirmation here is survival, not measured lift. A deliberate skeleton-vs-skeleton test is unscheduled (low EV while it keeps winning). Briefing default: the fixed brief skeleton itself — hooks vary only the first-3s cold-open. Spawned: none (it IS the template). Provenance: seed-decode, live-confirmed weekly.
WP-0006	Faith is the deepest, most under-supplied seam. P03 concepts outlast everything (240d max-days, 417d hook) on a quarter of the copy supply (48 vs 187). persona P03SKU-faith-kitLP01	proven	active	seed: Faith Kit 240d / 48 copies, faith hook 417d baseline · live: CON-0007 SCALE at Meta-CAC $22.60	confirmed 06-26 · decays 09-24	evidence
Cited: seed longevity table baseline + TST-0003 → CON-0007 pooled read (9 creatives · $1.9K · 84 purchases). CAC vs baseline: −37% vs $35.96 account Meta-CAC representative. Counter-evidence (load-bearing): the TTM Bibles campaign sat below contribution breakeven on an incremental basis — P03 TTM Meta-CAC $44.01 ≈ true $62 seed est. Durability ≠ incrementality: the lane wins on FRESH wrappers clearing true gates (CON-0007), not on recutting the 417d creative. Briefing default: P03 holds a standing queue slot; copy-supply target raised toward parity with durability rank. Spawned: CON-0016 Sunday-school-teacher cold-open · CON-0017 grandparent passing-down-faith (cross-spawn with WP-0003). Provenance: seed-decode, live-confirmed W26.
WP-0002	Authority cold-open beats mom-testimonial for P02: +28% thumbstop, −19% Meta-CAC vs control across offers. persona P02wrapper W01/W02hook H03ugc	emerging	active	TST-0002 / 0007 / 0013 · 9 creatives · $2.1K · Meta-CAC $29.11 vs $35.96 · seed prior: kindergarten-teacher spine, Magic Kit 119d baseline	n/a — emerging patterns don't decay, they accrue or die	evidence
Cited: TST-0002, TST-0007, TST-0013 (control = incumbent mom-testimonial in lane) · 9 creatives · 72 purchases · $2,096 representative. Counter-evidence: none yet; the W26 pooled read landed ITERATE at rung 3 (front-end strong, CAC $29.11 in the $25-29 band) — supportive but not yet a clean win. Promotion gate: 6+ tests for proven (3 of 6). Briefing default: P02 briefs open with a credentialed authority cold-open unless explicitly testing against it. Spawned: CON-0014 SLP variant · CON-0015 OT variant (authority-figure ladder). Provenance: live-engine with seed prior.
WP-0003	Grandparent-voiced wrapper (W03) clears G1 at ~1.7× account-median thumbstop on P04 cells. The first creative ever addressed to the audience that's already 25% of delivery. persona P04wrapper W03	emerging	active	TST-0009 / 0012 / 0014 · 11 creatives · $1.5K · thumbstop 41% vs 24% median · pooled Meta-CAC $27.31 (2 reads KEEP, 3rd lands W27)	n/a — emerging	evidence
Cited: TST-0009, TST-0012, TST-0014 · 11 creatives · 55 purchases · $1,502 representative. Grounding (real): 55-65+ is ~25% of spend converting 1.75-1.94 vs core 2.07, with zero creative addressed to it baseline — the highest-EV underserved cell the matrix generator surfaced (queue position 1). CAC vs baseline: −24% Meta basis. Counter-evidence: none; both decided reads KEEP ($25-29 band, strategic cell) — max 2 KEEPs before forced KILL/ITERATE, so the W27 read decides. Briefing default: P04 briefs open grandparent-voice unless testing against it. Spawned: CON-0016 passing-down-faith scene · CON-0017 "for the grandkids' visit" containment. Provenance: live-engine.
WP-0007	Objection-handling converts on the LP, not in the hook. A08 lost-pieces as cold-open: CTR flat; moved into the LP02 claim stack: read-through +9pts. angle A08LP02	hypothesis	active	TST-0011 · n=1 test · 6 creatives · $0.7K · LP-level read (rung 5)	n/a — hypothesis	evidence
Cited: TST-0011 (LP test, page-level read) · 6 creatives · 21 purchases · $702 representative. VOC grounding: lost pieces is the #1 stated purchase killer in the review bank baseline — it converts as reassurance after the click, not as a scroll-stopper. Counter-evidence: none yet; n=1 means no claim strength. Briefing default: none — hypothesis patterns never touch the brief skeleton until emerging (3+ tests). Spawned: LP02 claim-stack module request → Lifecycle Strategist (adjacent to ISS-016). Provenance: live-engine.
WP-0008	Price-reframe (A12, one-time vs Lovevery subscription) cannot run as a cold-open. Cold traffic won't carry comparison math in 3 seconds. angle A12offer O05	hypothesis	active	TST-0006 · KILLED at G2 (CPC $1.71 · read-through 19% · $288 · purchases under guard)	n/a — hypothesis	evidence
Cited: TST-0006 · 5 creatives · $288 · gate verdict G2 (deterministic — CPC > $1.40 with read-through < 30%), CAC never displayed because the cell sits under the minimum-data guard representative. Counter-evidence: none. Registry effect: A12 saturation_state UNCHANGED — the angle survives, the placement was wrong. No resurrection without a new hypothesis (logged). Briefing default: none yet; constraint logged on A12 (retargeting + LP comparison module only). Spawned: retargeting concept + LP comparison module, both queued with source = pattern-spawn. Provenance: live-engine.
WP-0009	Kitchen-selfie "spark of learning" opener (W06 × H08) clears G1 for P05. Seed-proven off the account's longest-running opener family. persona P05wrapper W06hook H08	proven	decaying	seed: longest-runner opener family baseline · last confirmed 2026-03-12 (seed reconstruction) · 90d elapsed 2026-06-10 → auto-flagged	flagged 06-10 · revalidation TST-0016 queued W27	evidence
Cited: seed decode (spark-of-learning = longest-running cold-open family baseline); last gate-clearing date reconstructed to 2026-03-12 seed est. Decay pressure: the adjacent recycled-lane re-test (TST-0004) KILLED at Meta-CAC $41.27 — same wrapper family, weak signal. Process: proven + unconfirmed 90d → decaying → revalidation test (fresh W09 variant of the opener, TST-0016, W27) → re-confirm or retire per WP-0001 protocol. Until then the EV scorer suppresses H08 — no new spend against a decayed prior. Briefing default (suspended): P05 cold-opens defaulted to H08; suspended while decaying. Spawned: TST-0016. Provenance: seed-decode.
WP-0001	"Recycled 2024 mom-testimonial UGC still converts." The account's #1 historic failure, formalized so it can never repeat silently. wrapper W06ugcaccount-wide	was proven	retired 06-26	seed: carried the bulk of 2025-26 scale; winning ads were recycled 2024 UGC that no longer converts baseline · re-test TST-0004: Meta-CAC $41.27 vs $29 breakeven → failed	flagged decaying AT INGEST (last confirmed 02-08, >90d) → revalidated → failed → retired	evidence
Cited: seed decode + account audit (incumbents at frequency 4.0-5.4, first-time-impression ratio <30% baseline) · counter-evidence that killed it: TST-0004 (4 creatives · $619 · 15 purchases · Meta-CAC $41.27) representative. Why it's the canonical case: a decayed pattern was still being SPENT against — exactly what the half-life automation exists to catch. It was auto-flagged decaying the day the seed history was ingested, revalidated within two weeks, and retired only after its replacement (CON-0010/0011 refresh wrappers) entered test — never killed into a creative vacuum. Briefing default change: the recut lane is closed; every refresh brief requires a NEW wrapper on the same cell. Spawned: CON-0010 (W08 side-by-side) · CON-0011 (W09 day-in-life) refresh concepts. Provenance: seed-decode, live-refuted.

Every evidence figure on this table is a query against fact_creative_day through the concept join, recomputed nightly — if cited rows get re-attributed or a cited test is reopened, the pattern flags itself stale. A decided test with no learning link is flagged incomplete (0 currently). Scope tags are registry foreign keys, never free text. Reconciliation with the weekly scorecard: the scorecard counts week-W26 movements (1 promoted · 3 active · 1 retired); this ledger is the full set including seed-decode patterns ingested 2026-06-15, whose half-life clocks started from their last reconstructed confirmation, not from ingest.

Decay watch — half-life automation

Proven patterns only · 90-day confirmation clock · auto-flag, never auto-delete

Pattern	Last confirmed	Clock	State	Action
WP-0004 advertorial gate	2026-06-26 (CON-0007 graduation)	decays 2026-09-24	healthy	—
WP-0005 hook skeleton	2026-06-26 (7/7 reads on skeleton)	decays 2026-09-24	healthy	—
WP-0006 faith seam	2026-06-26 (CON-0007 SCALE)	decays 2026-09-24	healthy	—
WP-0009 kitchen-selfie opener	2026-03-12 seed est.	elapsed 2026-06-10	decaying	revalidation TST-0016 · W27 · H08 suppressed in EV scorer
WP-0001 recycled UGC	2026-02-08 seed est.	elapsed at ingest	retired 06-26	re-test failed · recut lane closed

Confirmation = a cited cell clearing its gates with fresh spend inside the window; the nightly job moves last_confirmed_at automatically — no human remembers anything. Emerging and hypothesis patterns are exempt: they either accrue tests toward proven or die by counter-evidence. This automation is the direct answer to the recycled-2024-UGC failure: a decayed pattern can be flagged, but it can no longer be silently spent against.

How a pattern lives and dies

The compounding loop, state-machined

born every DECIDED test must link a learning (feed or contradict a pattern) — unlinked decided tests flag incomplete; n=1 enters as hypothesis

promoted hypothesis (n<3) → emerging (3-5 tests) → proven (6+) — strength is a count, not an opinion

acts proven patterns write a briefing_default into the brief skeleton and SPAWN concepts into the queue (source = pattern-spawn) — learn → generate, the loop closes

decays proven + unconfirmed 90d → auto-flag → revalidation test auto-queued → re-confirm or retire

dies retirement requires a failed revalidation AND a replacement in test — never a creative vacuum; the row stays queryable forever

Briefing defaults currently active

4 of 7 eligible patterns drive a default: LP gate (WP-0004) · brief skeleton (WP-0005) · P03 standing slot (WP-0006) · P04 grandparent-voice open (WP-0003) — H08 default suspended while decaying

Leaderboards — hook family & authority figure

Spend-weighted, Meta basis · grey = minimum-data guard or no data · live cols are W25-26 representative

Hook family	Seed provenance	Live n · spend	Thumbstop	Meta-CAC	Saturation / EV state
H06 anti-screen	164d baseline	9 · $1.9K	37%	$22.60	proven · revalidated W26 was suppressed (incumbent fatigue) → fresh faith wrapper won → reinstated
H03 authority cold-open	119d baseline	20 · $3.6K	38%	$28.33	proven-seed · emerging-live (WP-0002/0003)
H05 faith-duty	417d baseline	0 live	—	—	proven-seed fresh test queued (CON-0016)
H04 status-mom	366d baseline	4 · $0.5K	26%	$43.91	watch first live read missed — one read ≠ family verdict
H08 spark-of-learning	longest-runner opener baseline	4 · $0.6K	21%	$41.27	decaying · suppressed (WP-0009 → TST-0016)
H07 small-biz-relatability	164d baseline	0 live	—	—	proven-seed unqueued
H01 category-king · H02 rush-bandwagon	in ~100% of seed ads baseline	structural — skeleton layers, never the tested variable (see WP-0005)			skeleton

A fatigued hook flips its registry saturation_state and is suppressed in the EV scorer until a fresh wrapper revalidates it — H06 is the worked example: fatigued on the P05 Busy Book incumbents, revalidated in one week by a fresh wrapper on P03. Saturation is validation; the wrapper is where differentiation lives. The A08 "proof" opener from TST-0011 has no registry family yet — H09 proposed at the 07-01 monthly matrix audit (registries are append-only, human-approved).

Authority figure	Live n	Thumbstop	Meta-CAC	State
Grandparent (W03)	11	41%	$27.31	challenger — top thumbstop 2 KEEP reads
Kindergarten teacher	5	38%	$28.98	incumbent control seed spine, 119d baseline
SLP	4	33%	$29.39	grade C — small n
OT	queued	—	—	CON-0015
Sunday-school teacher	queued	—	—	CON-0016
Veteran mom-of-five	backlog	—	—	idea

The authority ladder is the wrapper-differentiation engine on the proven fine-motor-anxiety spine — same hook formula, new credentialed voice each rung (07-creative week-one candidates). Confidence grades by sample size: everything here is C until a figure accrues 10+ reads.

Leaderboards — wrapper win-rate & talent hit-rate

Win = SCALE or KEEP on a pooled concept read · 7 reads decided to date · all grades C until n≥10 per row representative

Wrapper	Reads	Verdicts	Win rate	Meta-CAC
W09 day-in-life	1	SCALE 1	1/1	$22.60
W03 grandparent passing-down	2	KEEP 2	2/2	$27.31
W01 teacher-testimonial	1	ITERATE 1 (rung 3 re-hook)	0/1	$28.98
W06 kitchen-selfie	2	KILL 2	0/2	$42.38
W08 side-by-side (kit vs iPad)	1	KILL 1 (G2)	0/1	guard
W07 unboxing-demo	0 (LP-level read only)	TST-0011 page read	—	$33.43
W04 podcast · W05 street-interview · W10 founder	unexplored — concept-differentiation headroom			—

One losing concept never kills a cell — it kills a wrapper (per-cell verdicts need 2+ concepts). Both W06 kills were decayed-lane re-tests, which is the half-life system working, not a wrapper indictment in fresh hands; W06 re-enters only via TST-0016.

Talent / lane	Launches	G1 pass	G3 reach	Roster decision
T02 · affiliate (CON-0007 lane)	12	58%	25%	renew briefed off validated AI prototype
T-AI · in-house Seedance lane	31	39%	17%	expand ~$31/launched vs $52 account · 24h SLA · 2 prototypes promoted to affiliate briefs
T05 · affiliate	9	22%	0%	brief narrower hook drift vs skeleton in QA lint
T07 · affiliate (new)	4	early — n<8	—	hold judgment
T03 · recut lane	8	13%	0%	closed per WP-0001 retirement

Talent is a first-class registry dimension — these rows ARE the roster decisions, derived from the same fact table as everything else. The AI lane's job is prototype-before-brief: soft-test the hook in-house, brief affiliates only on concepts that already show signal. SLA clocks: affiliate 5-7d, AI lane 24h.

Fatigue half-life by concept

Predicts refresh demand for the queue · triggers: frequency >4.0 · CTR −25% vs 14d peak ×3 days · CAC +20% WoW ×2 · first-time-impression <30%

Concept lane	Copies	Age / peak	Frequency	State	Refresh action
Busy Book / Learning & Play incumbents	187 baseline	186d max baseline	4.0-5.4 baseline	fatigued FTI <30%	refresh briefs auto-spawned (ISS-018): CON-0010 W08 · CON-0011 W09 — same cell, new wrappers; incumbents retire only when a replacement wins in test
Educational Kit	111 baseline	118d baseline	3.6 representative	watch CAC +14% WoW (trigger at 20% ×2wk)	refresh brief drafted, unqueued
Faith Kit	48 baseline	240d baseline	2.9 representative	alive at 240d	supply-up instead (WP-0006 standing slot)
Magic Kit	17 baseline	119d baseline	2.4 representative	healthy	—
CON-0007 (graduated W26)	5 clones wk 1	12d	1.8	fresh	copy-clone expansion on hold-status per graduation protocol

First LIVE half-life estimate needs 4+ scaled live concepts (ETA Q3) — until then the honest read is the seed contrast: the 187-copy concept fatigued while the 48-copy concept is alive at 240 days, i.e. copy-count, not calendar age, looks like the burn-rate driver. Provisional planning number: ~2 refresh briefs/week through July, injected into the queue between Mondays (source = fatigue-trigger). Fatigue aggregates creative → concept → hook: a fatigued hook flips its registry saturation_state (see H06/H08 in the hook leaderboard).

Rollup explorer

Any historical question is one GROUP BY against fact_creative_day through the concept join — not a meeting

Group by

Date range

Persona filter

run query

Saved queries: P03 faith statics vs UGC · 2025-Q3 Persona · TTM Landing page · TTM P×A×O cell · live Wrapper · live

Group	Creatives	Spend	Purchases	Revenue	sw-ROAS	Meta-CAC	Note

Calibration — keeping the engine honest

Attribution correction · verdict backtest · versioned thresholds — rules decide, and the rules themselves get audited

Period	Meta ROAS	Shopify-true	Factor	Δ vs ÷1.40	Action
TTM average	1.92	1.37	1.40	—	locked into v1.0 gates baseline
2026-03	1.97	1.38	1.43	+2.1%	within band seed est.
2026-04	1.94	1.37	1.42	+1.2%	within band seed est.
2026-05	1.89	1.37	1.38	−1.5%	within band seed est.
2026-06 MTD	2.05	1.45	1.41	+0.9%	refit lands 07-01 representative

The 1.40 is over-attribution, not signal loss (match quality 9.1/10) baseline — so it's a correction factor, not a bug to fix. Refit monthly against Shopify referrer revenue. Per-concept rule on top of the account factor: every SCALE re-validates on Shopify-true within 7 days of graduation; divergence >15% from 1.40 freezes the ramp. Next check: CON-0007, due 2026-07-03.

Verdict backtest	Status
Quarterly backtests run	0 — first window opens 2026-09-15 (needs 90d of verdicts + 30d truth lag)
Verdicts accrued for re-scoring	7 (1 SCALE · 3 KILL · 1 ITERATE · 2 KEEP) — each will be re-scored against its subsequent 30-day Shopify-true outcome representative
Threshold-move trigger	>20% of SCALE/KILL verdicts wrong in one direction → gates move, change logged below
Human overrides to audit	0 logged — overrides require a written reason and get back-tested alongside the rules
Shadow backtest on seed history	v1.0 fatigue rules replayed over reconstructed seed facts fire on the recycled-2024 lane in early Feb 2026 (frequency + CTR-decay triggers); spend after trigger ≈ $118K illustrative — reconstructed attribution, directional only. That figure is the case for the engine, not a measurement.

KILL wrongness is the expensive blind spot (a killed concept generates no future data), which is why the backtest re-scores kills against lookalike-cell performance, not just the dead concept's own tail. Until 09-15 this panel reports zero — a backtest you haven't run yet is not a number.

Version	Date	Threshold change	Why (citation required)	Status
v1.0	2026-06-15	Initial gate set: G1 $25 / 2K impr (thumbstop ≥20%, CTR ≥ p25, seed floor 0.9%) · G2 $70 cum (CPC ≤$1.40, read-through ≥30%) · G3 $125 · pooled read 10 purchases or $500 · SCALE ≤$25 Meta-CAC ∧ ≥10 purchases ∧ read-through ≥30% ∧ worst-plausible ≤$29 · KILL >$29 or 0 purchases by $250 · correction ÷1.40 · pattern half-life 90d	Derived from real economics: AOV $76.58 · contribution-breakeven true ROAS 1.89 → Meta-CAC $29 · scorecard true CAC $35 → Meta-CAC $25 baseline	active
v1.0.1	W27	Recalibrate G1 CTR floor from seed 0.9% to account-true trailing-30d p25	Seed floor was a placeholder pending 30 days of live distribution — flagged at v1.0 ship	proposed
—	standing	Gate-drift review: per-variant kill rates off the expected 60/25/15 curve by >10pts for 2 weeks → threshold review fires	W26 actuals 58/24/19 — inside band, no review	monitoring

Thresholds are versioned, not sacred — but a change without a backtest citation or calibration drift behind it doesn't merge. This changelog is append-only, same as every registry.

Compounding state