Mirror

A shopper with an event does not have a product in mind; she has a date, a dress, a venue, a budget and a fear of getting it wrong. Every store makes her do the translation first. Fit is the primary axis, not an edge case: about half of apparel returns are size related. She needs the system to say what it knows, admit what it does not, never assert a size the data does not support, and volunteer the answer it earns nothing from when she already owns it.

The original brief was an agentic Amazon assistant for "a quiet mechanical keyboard under $100," a CLI with a cart link. Five research documents (about 6,300 lines, every figure tagged by source reliability, vendor statistics re-sourced before use) came before any product code. Then the brief was replaced outright and the old one left in git history. Keyboards became the event; Amazon became one source among six; frictionless checkout stopped being a virtue. The one idea that survived, reading one and two star reviews for hidden flaws, became the Evidence Lens. The CLI agent was kept as a working reference with its eleven case golden set and, stated plainly in the README, shares no code with the app.

The first spec said if confidence is at least 0.9, show one product. I rejected it in writing as a magic constant dressed as intelligence. Confidence is now derived from two things the system genuinely knows: the score margin between the top candidates, from the same function that ranked them, and the count of unresolved dimensions. It returns an ordinal band, so "94% match" is a compile error rather than a code review habit. On the real catalog that rule bit back: mass tier Amazon inventory is attribute uniform, so the top candidate never beat the runner up and "I'd start here" could not fire at any maturity. Review evidence, weighted by volume, became a bounded ranking term; two of five missions now answer with one product on the real catalog, and the page says "The closest I have" when they do not.

Maya's own shoes rank on merit against retail through the same scorer. Three of four lose and the interface says why; the black slingback legitimately places at $0. One owned answer maximum. Nora, the second shopper, owns nothing that qualifies, and the honest outcome is that the closet pass returns nothing and says so rather than surfacing a near miss to keep the feature visible. Same code, same rule, opposite result: a better test than a second winning $0 answer.

Three passes went looking for my own failures. The interaction pass hunted claims that outlived what made them true and controls that looked like affordances and were not: a comparison screen reassuring her about constraints nothing had applied, a fit verdict contradicting its own evidence, a "Forget this" control with no handler, 18px touch targets on the fastest route through the app. The first run audit found the empty start flow built, marked done, and unreachable: the default profile opened a brand new install into six months of someone else's history. A new shopper could answer the fork honestly, forever, and never be shown a recommendation. The polish round found "black loafers" leading with a gold shoe because the router kept the noun and dropped the adjective. Each is written down with its fix.

App

• Next.js 16, React 19, TypeScript. One XState machine with two branches: the Maya mission and the new shopper journey. Journey position is state; surface complexity is derived.
• Motion animates state, GSAP animates time. Base UI primitives, Tailwind 4 over a token file that Figma and code both read; a sync check gates the build.
• Editorial Performance Minimalism: image dominance, monochrome interface, 46px to 11px scale contrast, near zero radius, no shadows. Three anti patterns enforced: no fake precision, no generic AI visual treatment, no conventional ecommerce chrome.
• Deterministic intent router and parser as the floor; an optional Claude route for phrasings they cannot map. A 501 without a key is a working deploy.

Catalog

• Canopy API over Amazon search, listing and reviews; each listing judged by Claude for attributes, fit notes and a quality tier; cached to disk with a build report recording every rejection and its reason. The free tier died 399 requests in; a cached rebuild with an explicit tier floor took 23 products to 79.
• Outerwear modelled as a discriminated union, not optional fields, so the compiler named all 70 places that assumed shoes.
• A self test route with 97 checks, each one a failure mode found while wiring the real catalog in.

Imagery and proof

• Every person, outfit and look generated with Higgsfield: a fictional shopper, fictional brands. Product photos are the listings' own.
• A Playwright capture pipeline photographs every screen off the running app by visible text; nothing is restaged. The reel and every frame on this page came out of it.
• The reference engine: a Claude SDK tool loop with typed review extraction, a prompt injection boundary, and an eleven case golden set. Kept for its architecture, not connected.

The core path is a deterministic parser, not a model; the point to prove is that conversation changes the workspace, offline and identically every run, and the reference engine slots behind the same interface next. Profile states are fixed snapshots you switch between, not learned from you. Purchase is simulated at the final step. The public deploy runs without a model key, so it understands fewer sentences than the local build with one. Three sample shoppers, no user testing yet.