# Design direction and original prompts

Reference reviewed: https://motionsites.ai/ and https://motionsites.ai/backgrounds on 14 October 2026. Their public gallery suggests an authored focal moment and animated backgrounds. These prompts are original, adapted to this site's established identity; no paid prompt text, assets or unseen animation is claimed to have been copied.

## Preserve the existing site

Keep the photographic slate-blue lake, droplet, ripples, display type, serif article, light/dark themes, menus and older detector chapters. Change the homepage's research question and primary reading action. Add one chapter for the learning experiment and visibly label the previous results as the detector foundation.

Tokens: paper #f6f7f9; ink #14171f; research emphasis #c42a68; water #273e4d; hero text #f8fafc. Archivo supplies headings and controls; STIX Two Text supplies prose. Reuse the existing dark-theme tokens and detector colours. Article text remains left-aligned in a readable column; the training replay and results table have more width.

Layout: existing water hero > honest research status > learning replay > held-out evidence > method and limitations > preserved detector chapters.

## Original homepage prompt

“Evolve Detecting the Unseen into a journal about what an AI trader learns when rewarded for execution quality. Preserve the existing lake, navigation, palette and typography. Keep the title, and put the actual research question below it. Make the primary action Read the learning experiment. Treat the old detector work as the foundation. Present negative results with the same prominence as positive ones. Never label a strategy novel or manipulative solely because it earns money or triggers an alert.”

## Original motion prompt

“Design a replay of the learner's changing choices, using measured categorical probabilities at each training round. Preserve the site's restrained editorial feel. Let visitors select the learner and reward condition, scrub through rounds, or press Play. The single focal motion is probability flowing between ordinary trading choices; bars share an exact zero-to-one scale and show numeric percentages. State-dependent updates use 180 ms deceleration, never invented values or simulated live training. Pause when the tab is hidden or the replay leaves view. Start paused, respect reduced motion, and retain a complete exact table if scripting fails.”

## Original results prompt

“Let readers compare calm, volatile and non-reactive markets without losing which population each result describes. Pair each learner's performance with its no-extra-order twin and a benchmark selected before testing. Show uncertainty beside each effect. Separate an alert rate from evidence of manipulation. Keep a semantic static table, raw episode download and run manifest available. A filter changes the evidence shown; it must never hide an inconvenient learner or replace missing evidence with zero.”

## Review against the brief

The lake already supplies the visual identity and ambient motion. Adding another decorative scene would dilute it. The new replay instead makes the learning process inspectable. It uses the same visual system as the old website, with no added framework or animation dependency. Motion is user-controlled and bounded to the replay. The findings remain readable without JavaScript.