adScientist / instrument v4.06 — syd
Live · AEST
120 240 360 480 600 720 840 960 1080 1200
[ Experiment № 04.16.26 / Rapid Split Testing ] /var/log/meta-ads.csv

Run your
Meta ads
like a laboratory,
not a
guessing game.

adScientist is the autonomous test bench for paid social. Bulk-generate creative with AI, run them through Rapid Split Testing — auto-pause at 2,000 impressions, promote winners under $1.00 CPOLC, and let the rule engine kill what doesn’t earn its keep. No dashboards to babysit. No ad fatigue you missed at midnight.

Tests run / 30d
14,308
▲ 12.4% week
Avg. winner CPOLC
$0.74AUD
▼ 38% vs manual
Spend automated
$2.4MAUD
▲ 6 figures / wk
fig. 01 — specimen chamber
SPECIMEN CHAMBER
— campaign c-0917 / set a — interest:dtc-skincare
TESTING
Variant 03 / 12 angle: social_proof
creative_03_v2.png
"We stopped guessing. Our CPA dropped 41% in three weeks."
— Hana K., founder, Loop Skincare. 2,103 ads tested last quarter, fully on autopilot.
Shop Now ⟶
Impressions
1,847
Outbound clicks
7
CPOLC (AUD)
$0.82
Recall Lift / 60s +11.8 pts
§ 02 / The Foil filed under methodology · rev. 04.16.26

Meta says ten thousand.
The textbook says seven.
We pause at two.

The conventional wisdom says you need 7,000–10,000 impressions to be 95% confident a variant is a winner. That answer is correct — for the question "did this beat the control?" But that’s a research question. You don’t have a research question. You have a business: is this ad earning its keep, today? Different question. Different threshold. Different math.

Fig. 02 — impression threshold by authority x-axis: impressions per variant · 0 → 12,000
adScientist rapid split testing 2,000 DEFAULT · LIVE
The textbook stat. sig. @ 95% 7,500
Meta recommends per-variant minimum 10,000
0 2K 4K 6K 7.5K 10K 12K

The standard method asks

"Did this variant
beat the control?"

That’s a research question. The right tool is statistical significance, the right threshold is ~95% confidence, and the right sample size is 7K–10K impressions.

Useful for academic papers. Catastrophic for a $100/wk ad budget — at that spend, the standard method gives you one swing per week. That’s not a testing program. That’s a prayer.

vs

Rapid Split Testing asks

"Is this variant
making money?"

That’s a business question. The right tool isn’t a confidence interval — it’s a performance gate. The right threshold is whatever proves the ad pays its rent.

Run five variants where the textbook runs one. Each gets 2,000 impressions. The ones that clear the gate get promoted. The ones that don’t get killed before they bleed you dry.

THE GATE winner criteria · auto-promotion ruleset file: rules/winner.yml
> 5 Outbound link clicks
∧ AND
< $1 CPOLC (AUD)

Statistical significance,
replaced by economic significance.

A variant that produces five real outbound clicks at under a dollar each has demonstrated something more valuable than statistical separation from baseline: it has demonstrated economic viability.

The gate does the work the impression count would otherwise do. We don’t need a research paper. We need a winner. The 2,000-impression auto-pause isn’t a corner cut — it’s a portfolio optimization made possible because the gate is doing the heavy lifting.

safety rails: budget cap $50/day · circuit breaker armed ruleset valid
STANDARD METHOD $100 / wk · 10K impr per test
test 01 1 swing per week
52 swings / year. One winner discovered roughly every 4–6 weeks. Most of the budget pays for confidence you didn’t need.
RAPID SPLIT TESTING $100 / wk · 2K impr per variant
v.01
v.02
v.03
v.04
v.05
260 swings / year — same budget, 5× the discovery rate. Losers die early. Winners get found before the textbook is done warming up.

Meta is asking “is this statistically true?”
You should be asking “is this profitable?”
They are not the same question.

Lab Log · Live
14:02:11 winner promoted — act_8596 · CPOLC $0.61 14:02:08 variant 07 auto-paused @ 2K impr · CPOLC $1.48 14:01:54 18 new creatives generated — angle: case_study 14:01:31 rule fired: kill_underperformer · saved $24.10 14:01:02 recall lift detected · +9.4 pts on variant 05 14:00:48 bulk launch — 12 ads / 5 ad sets · budget $100/wk 14:00:11 winner promoted — act_2214 · CPOLC $0.84 14:02:11 winner promoted — act_8596 · CPOLC $0.61 14:02:08 variant 07 auto-paused @ 2K impr · CPOLC $1.48 14:01:54 18 new creatives generated — angle: case_study 14:01:31 rule fired: kill_underperformer · saved $24.10 14:01:02 recall lift detected · +9.4 pts on variant 05 14:00:48 bulk launch — 12 ads / 5 ad sets · budget $100/wk 14:00:11 winner promoted — act_2214 · CPOLC $0.84