Claude Opus 4.8 ran eight tasks twice. Identical prompt and identical files both times. The only difference is whether it first read a real Emulo profile mined from my own coding sessions. Then every output was scanned for things the profile explicitly forbids.
| Rule | What the profile says | Opus 4.8 | + Emulo |
|---|---|---|---|
| purple | Taste: black/white (anti-purple) | 9 | 0 |
| em-dash | No em-dashes. | 1 | 1 |
| fabricated-metric | No fake data. Show `--`, never a fabricated number. | 1 | 0 |
| rule-of-three | No forced rule-of-three. | 1 | 0 |
#5b4ef0 (hue 245deg sat 0.67)#7c5cf5 (hue 253deg sat 0.62)#4335cf (hue 245deg sat 0.74)2,000+ teamsNo thread. No poll. Just the time.#5b4bdb (hue 247deg sat 0.66)#3d2fb8 (hue 246deg sat 0.74)#2a2850 (hue 243deg sat 0.50)#231d5c (hue 246deg sat 0.68)#4a3bc4 (hue 247deg sat 0.70)—#a26cff (hue 262deg sat 0.58)–Both runs get the same task text and the same starting files. Nothing in the prompt hints at a preferred answer, and no taste is leaked into either side.
Claude Opus 4.8 both times. One run alone, one after reading the real profile file. That single difference is the whole experiment.
The Emulo run reads the actual mined profile the tool ships. No hand-written instructions were written for the task.
No taste panel. A violation is a hue in the purple range, a fabricated metric, an em-dash. Every hit prints the matched text.
Orbit, a scheduling app neither run had context on.
Emulo broke one of its own rules: an em-dash in the pricing copy. It is in the table above. A result that hid it would be a pitch.
This does not claim Emulo produces better work. What it counts is narrower: how often each output broke rules written in the profile. The plain run never saw those rules, and one run each way cannot tell the profile's content apart from simply being handed a long, confident document. The placebo test was built to check exactly that, and there the mined profile did not beat an invented one. On a separate blind panel judging generic quality, the plain model actually won several tasks, partly by doing the very things these rules forbid, because fabricated social proof and a tidy rule-of-three score well with a neutral judge.
Read it as directional: eight tasks, one operator, one model. The scanner is regex and hue math, so it catches concrete rules and cannot see a vague one. It also missed three purples on its first run until the detector was rewritten to measure hue instead of matching a list of hex codes, which is a fair thing to be skeptical about.