Escaping the AI House Style — Taste Trained on the Public Domain

You are a design-systems lead. Do not feed my agent the same Impeccable and awesome-design-md stack everyone uses. Build me a defensible voice from one public-domain archive.
1. Pick one archive (rave zines, arcade flyers, or the Rumsey maps). Scrape 2,000 to 5,000 images politely, and verify each item's license is public domain or permissive. Drop anything unverified.
2. Embed with open_clip, reduce with UMAP, cluster with HDBSCAN. Prove the clusters are real by re-running across min_cluster_size and seeds and reporting the adjusted Rand index. Inspect them in the Embedding Projector.
3. For the 3 to 5 clusters I like, write a DESIGN.md each (palette tokens, type, grid, texture, motion) and a parameterized JSON image-prompt profile. Tag behavior, not adjectives.
4. Build the same landing page three ways: stock toolkit, cluster A, cluster B. Screenshot all three with Playwright.
5. Run a blind rater test: can people tell the archive pages from the stock page better than chance? Report the p-value.
6. Check the outputs do not regurgitate source images by perceptual hash.
Return the corpus manifest, the clusters with stability scores, the DESIGN.md files, the three pages, and the blind-test verdict.

ROLE
You are a design-systems lead who treats a distinctive voice as a measurable, ownable asset, not a prompt trick.

CONTEXT
The saved posts show the problem and the escape.
- @adam_ha_yes and @nateherkai: the standard stack everyone feeds their agent, 21st.dev, getdesign.md, Impeccable, awesome-design-md, a 'taste' skill, Figma MCP and Playwright. When everyone uses it, the output converges.
- @designteamofone: 'we've lost the plot on the exploratory nature of design.'
- @itsdesignertom: AI gets 80% there, the human finishes the 20%.
- @rndyrbrts: turn taste into data and back. Manifest, measure 'palette, contrast, density, spacing, motion', 'tag behavior, not adjectives', cluster into visual DNA channels, generate from the rules.
- @6ab3 archives: the Internet Archive rave-zine and arcade-flyer collections, and the David Rumsey map collection, 150,000+ historical maps.
- @rajistics: the TensorBoard Embedding Projector, where 'cold' pulls in 'Soviet' and 'war'.
- @chase.variantai: parameterized JSON image-prompt profiles.
- @therobertta: StopSlop removes AI tells but 'the missing opinion survives the edit' - deleting tells is not having a position.

HARD CONSTRAINTS
- License first. Verify every image is public domain or permissively licensed, item by item. 'On archive.org' is not a license. Drop anything unverified, and record the license per item.
- Scrape politely: respect robots.txt, rate-limit, and cache.
- Rules from evidence, behavior tags not adjectives. No 'futuristic'; yes 'dense grid, high-contrast, hand-cut edges'.
- Prove the clusters are real before building on them.
- Prove the result is distinguishable from the stock stack, or admit the corpus added nothing.

DELIVERABLES
1. corpus/: a scraper, manifest.csv (url, local file, resolution, license, source collection) for 2,000 to 5,000 images, deduplicated by perceptual hash.
2. embed/: open_clip embeddings, a UMAP projection, HDBSCAN clusters, and stability.md reporting the adjusted Rand index across min_cluster_size and seeds, plus an Embedding Projector export.
3. design/: for 3 to 5 chosen clusters, a DESIGN.md each (palette as tokens, type scale, grid, texture, motion) and a parameterized JSON image-prompt profile.
4. pages/: the same landing page built three ways, stock toolkit, cluster A and cluster B, with Playwright screenshots in light and dark.
5. blind_test/: the rater protocol, the responses, and a chi-square or exact test against chance.
6. report.md.

BUILT-IN GAP ANALYSIS
List the images dropped for unverifiable license, clusters that dissolve under re-clustering, DESIGN.md rules not actually reflected in the pages, and a blind test with too few raters. Fill what you can. Mark the rest UNFILLED.

SELF-QUESTIONS
Is every image actually clear to use? Are my clusters real structure or embedding noise? Do my tags describe behavior or mood? Can anyone actually tell my pages from the stock page?

Also note: @rndyrbrts' own room-color probe and taste-to-data pipeline is the embryonic version of this - he extracts behavior from a moodboard and generates from the rules. You are doing the same thing but with an owned public-domain corpus and a falsifiable distinctiveness test, so the voice is defensible rather than just personal.

STOP
Stop when the clusters are stable, the three pages are built and screenshotted, and the blind test has run. Stop and ask before training any image model on the corpus.

RETURN
Return corpus/, embed/, design/, pages/, blind_test/ and report.md with the sections Corpus, Clusters, Design Rules, Pages, Blind Test and Gaps.

Pick one archive: the Internet Archive rave-zine collection (archive.org/details/ravezines), the arcade-flyer collection (archive.org/details/arcadeflyers), or the David Rumsey map collection (archive.org/details/david-rumsey-map-collection).
Use the Internet Archive's API or metadata endpoints, respect robots.txt, rate-limit requests, and cache raw responses.
For every item, capture its license or rights statement from the item metadata. Keep only items that are public domain or carry a permissive license. 'Hosted on archive.org' is not a license, so drop anything whose rights are unclear.
Download 2,000 to 5,000 images at usable resolution. Deduplicate by perceptual hash.
Write manifest.csv: url, local file, resolution, license, rights source, source collection. Record the archive collection's total size so you can state what fraction you sampled, since a biased sample skews the clusters. Keep a fetch-date column, because archive items and their rights statements can change.

Embed every image with open_clip (ViT-B-32 or ViT-L-14). Save the vectors.
Reduce with UMAP for visualization, and cluster the full-dimensional embeddings with HDBSCAN.
Prove the clusters are structure, not noise: re-run HDBSCAN across min_cluster_size values (for example 15, 30, 60) and across 3 random seeds for UMAP. Compute the adjusted Rand index between runs. Keep only clusters that persist with a high ARI.
Export the embeddings and labels to the TensorBoard Embedding Projector and inspect neighborhoods by eye, the way @rajistics inspects word2vec.
Name the stable clusters by what they visually are, deferring the behavior tags to step 3. Store the reduced 2D coordinates so the Projector view is reproducible. Note how many images land in the HDBSCAN noise bucket, because a large noise fraction means the corpus has no strong structure and the whole thesis is weak.

For each of the 3 to 5 clusters you want to keep, measure the visual behavior across its images: palette (k-means in Lab, as tokens), contrast, edge and element density, whitespace, dominant type characteristics, grid tendency, texture and any motion cues.
Write a DESIGN.md per cluster in the awesome-design-md style, but derived from your corpus, not borrowed. Use behavior tags, not adjectives: 'dense grid, hand-cut edges, high-contrast duotone', never 'retro-cool'.
Write a parameterized JSON image-prompt profile per cluster, in @chase.variantai's format, with the swappable values (palette, density, texture, subject) exposed.
State, per cluster, what the system should always do and never do, the @rndyrbrts constraint.
Keep every rule traceable to measured features of the cluster.

Define one landing page brief: the same copy, sections and content for all three builds.
Build it three ways with the agent:
- Control: the stock stack (Impeccable plus awesome-design-md plus 21st.dev components), the @adam_ha_yes and @nateherkai setup.
- Variant A: driven only by cluster A's DESIGN.md and JSON profile.
- Variant B: driven only by cluster B's DESIGN.md and JSON profile.
Keep the content identical so only the design voice differs.
Screenshot all three with Playwright at desktop and mobile widths, in light and dark.
This is the @itsdesignertom 80/20 test made concrete: does the corpus-derived 20% actually change the result? Keep a shared components and copy file that all three builds import, so identical content is enforced mechanically, not by hand. Record which stock-stack skills the control used, for the audit.

Design a blind test. Show 15 or more raters outside your team the three pages (control, A, B) without labels, in randomized order.
Ask each rater: group these by visual voice; which feels most distinctive; and which two look most alike. Also ask them to guess which one is the 'generic AI' design.
Score it: can raters separate the archive-anchored pages from the stock page better than chance? Use a chi-square or an exact test and report the p-value and effect size.
This is the dive's falsifier, stated up front: if raters cannot tell the archive pages from the stock page better than chance, the corpus added nothing the shared stack lacks.
Also run @therobertta's StopSlop check in reverse: does each page take a visual position, or just avoid tells?

Check that nothing in the generated pages or any generated imagery is a near-copy of a source image. Compute perceptual-hash (phash) distances between every output asset and the corpus. Flag any distance below a set threshold (for example 8) as a potential copy and remove or regenerate it.
Re-verify the license trail end to end: every corpus image used in training or generation inputs has a recorded public-domain or permissive license, and that license permits derivative and commercial use.
If you intend to train a small image model or LoRA on a cluster for hero imagery, stop here and get explicit human approval first, and document the training-data license basis.
Write a one-page rights summary a lawyer could read.

Run the oracle pass:
(a) Blind raters distinguish the archive pages from the stock page better than chance (p-value and effect size).
(b) Cluster stability: kept clusters persist across settings by adjusted Rand index.
(c) Regurgitation: no output within the phash threshold of a source.
(d) License audit: every used image verified public domain or permissive.
(e) The falsifier: if raters cannot tell the pages apart, say plainly the corpus added nothing and the shared stack would have done.
Write report.md with the sections Corpus, Clusters, Design Rules, Pages, Blind Test, Rights, Falsifier and Gaps. State whether the corpus produced a defensible, distinguishable voice. Rank the gaps by impact on that claim.

You are an independent reviewer: a design lead plus an ML engineer. Open report.md, corpus/manifest.csv, embed/stability.md, design/, pages/ and blind_test/.
1. Enumerate every claim: corpus size and license, cluster stability, design-rule provenance, page distinctiveness, and originality. Mark each verified, unverified or false.
2. Sample 20 corpus rows and confirm the license against the live item pages.
3. Recompute cluster stability: re-run HDBSCAN with a different min_cluster_size and compare the adjusted Rand index against stability.md.
4. Check the three pages use identical content, so the comparison is clean.
5. Recompute the blind-test statistic from the raw responses. Does the p-value support the distinctiveness claim? Are there 15 or more outside raters?
6. Run phash between the output assets and the corpus and list any near-copies.
7. List what is missing: dropped-license images, dissolving clusters, rules not reflected in the pages, an underpowered blind test.
8. Rank the gaps by impact. Fix the top three now. Report what you could not fix and why.
Return audit.md with the columns claim, status, evidence and fix, plus 'Remaining gaps' and an A-F grade on whether the voice is defensible.