Darwinian Creativity.
PROTOCOL v0.1Read as Markdown

Design Assay Protocol v0.1

Vocabulary

Term Meaning
Target The thing being designed: a screen, flow, component set, or brand system.
Genome The design system encoded as data: tokens, component recipes, composition rules.
Gene One mutable parameter in the genome, with a bounded range.
Variant One rendered candidate produced from a genome.
Generation One batch of variants produced in a single round.
Assay The full measurement pass over a generation using the fitness functions.
Gate A hard pass/fail check. A variant failing any gate scores zero.
Hit A variant that passes all gates and beats the baseline composite score by the hit threshold (default 10%).
Hit rate Hits divided by variants assayed.
Baseline The existing design, or the best human-made option, scored with the same functions.
Run record The reproducible log of one assay run.

Stages

Stage 0. Target profile

Define before generating anything.

  • What the target is and who uses it.
  • Fitness weights for this target (see 02-FITNESS-FUNCTIONS.md).
  • The baseline, rendered and scored.
  • Budget: max generations, max variants, max compute cost.
  • Stop condition: default is a plateau, meaning composite gain under 1% across 2 consecutive generations.

Output: target.json

Stage 1. Genome

Encode the design system as data.

  • Tokens: color, type scale, spacing, radius, elevation, motion.
  • Component recipes: structure plus which tokens each slot accepts.
  • Composition rules: layout grids, section order, density.
  • Each gene gets a type, a range, and a lock flag. Locked genes never mutate. Brand-critical values such as the logo and primary hue are locked by default.

Output: genome.json, validated against schema/genome.schema.json

Stage 2. Variation

Produce a generation.

  • Mutation: change 1 to N genes within their ranges.
  • Crossover: combine genes from two parent variants.
  • Exploration: a fixed share of each generation (default 15%) is mutated with wider ranges, so the population does not collapse to one local optimum.
  • Generation size: 200 to 2,000 for screens, 50 to 200 for full brand systems.
  • Mutation is cheap work and runs on local inference by default.

Output: gen-{n}/variants/*.json

Stage 3. Render

Every variant becomes real output, not a description.

  • Rendered as production code (Next.js, Tailwind) in an isolated harness.
  • Screenshots at 320, 768, 1280 and 1920 px widths, in light and dark.
  • A variant that fails to render is logged and discarded.

Output: gen-{n}/renders/{id}/

Stage 4. Assay

Run the fitness functions as a funnel, cheapest first.

  1. Gates (seconds per variant): accessibility criticals, overflow, contrast floor.
  2. Measured functions (seconds to minutes): token fidelity, rhythm, performance, responsiveness.
  3. Judged functions (most expensive): rubric scoring by model judges, run only on the top 20% after step 2.

Output: gen-{n}/scores.jsonl

Stage 5. Selection

  • Keep the top 5% by composite score.
  • Add a diversity bonus so near-duplicates do not fill the survivor set.
  • Survivors become parents for the next generation.
  • Repeat Stages 2 to 5 until the stop condition is met.

Stage 6. Human review

The designer reviews the finalists (default: top 10).

  • Every accept or reject is logged with a one-line reason.
  • Repeated reasons become candidates for new fitness functions. This is how taste gets encoded over time.

Output: review.jsonl

Stage 7. Field assay

Ship the top 2 or 3 finalists to real users when the target allows it.

  • A/B or multivariate test, with the baseline as control.
  • Pre-register the metric and sample size before launch.

Output: field.json

Stage 8. Report

Publish using 04-CASE-STUDY-TEMPLATE.md.

Run record

Every run directory contains:

run-{date}-{slug}/
  target.json
  genome.json
  protocol-version.txt
  seed.txt
  gen-0/ ... gen-n/
    variants/
    renders/
    scores.jsonl
  review.jsonl
  field.json        (if Stage 7 ran)
  summary.json

summary.json holds the headline numbers: variants generated, hits, hit rate, best composite, baseline composite, gain, generations, wall-clock time, compute cost.

A run is reproducible when the seed, genome, protocol version and model versions regenerate the same variant set.

Versioning

  • Protocol versions follow semver. Breaking changes to stages or scoring bump the major version.
  • Every published result states the protocol version it ran on.