Fable 5.1 built sixty pages. Opus 5 built the same sixty.

Every run is Claude Code headless with the default system prompt replaced by one sentence, no settings sources, no MCP servers, every tool denied, in an empty directory. Nothing on the machine can reach the model. The only thing that differs between the three conditions is the text appended to that sentence: nothing, the profile, or the placebo. The only thing that differs between the two halves is the model id.

The profile is the file Emulo actually ships, mined from my own sessions, – bytes. The placebo is a working profile of a documentation lead who does not exist, – bytes, written to be silent on every dimension the scorer measures. If a placebo moves the numbers as far as the real thing, the real thing has not shown it is the real thing.

The scorer counts rounded corners, mean colour saturation, card-like containers, output length, and for the terminal brief, rendered width and escape-code density. It measures whether the output changed, not whether it is good. Nothing here measures good.

The system prompt every run shares:

Runs were interleaved across models and conditions and executed six at a time, so neither model got a quieter hour. Screenshots were taken by one headless Chrome five seconds after load, with WebGL on. A page that threw a console error or failed a request is marked on its card rather than hidden.

Things this cannot say. One profile document and one placebo document, so it cannot tell "mined" from "this particular text". No equivalence test, so overlap never means "the same". Two models on one night is one comparison, not a trend. And a screenshot is one frame of a moving page, which is why every card opens the live file.

• Emulo, the tool under test. Open source. Its own result is on this page and it is not flattering.
• The August study this replicates, with the placebo that matched the profile on every prediction.
• The pre-registration, the placebo document, the harness, the scorer, the summariser, the screenshot script.
• manifest.json, every run's exact command line, tokens, seconds and list cost. data.json, every measured value.
• Raw outputs: every card above opens one, and the raw file link in the viewer is the file itself, byte for byte what the model returned after a fence strip.

Who made this: Ohad, one person, the night the model came out. Read every number as coming from someone with a product in the comparison.