Generation without validation ships garbage to a factory floor. From the build log: EPI 120, PPI 110, TEX 40/40 computes a cover factor of 100.1%, over the loom's limit. The export must refuse, and the refusal must explain itself in loom terms.

I deleted the animation library after shipping with it. The studio's transitions had been built on it; removing it meant hand-rolling every motion the interface kept.

What it cost: an afternoon of re-tuning by hand, and the loss of a spring system I liked. What it bought: 110 KB off the gzipped payload and one less dependency standing between the gates and the operator.

I would do it again. The gates are the product. Anything that spends payload without guarding the floor is overhead wearing a costume.

A physics gate validates every generated pattern against loom limits before export. The gate constants are not published. The decision to refuse rather than warn is the product, and that decision does not transfer by reading.

Cover factor is public engineering. What is not public: where the limits sit, how the three gates compose, and what the reject panel says first.

Primitive: cloth(r,i), order 8, period 7, shared across every surface of this portal. Contrast: bone on void 15.61:1; ash 5.35:1; crimson 5.54:1. Captures: NO RECEIPT, three plates as notation. What this page cut: the stack string, the gate constants, the module map.

65,536 DOM nodes at 256 by 256. The obvious implementation renders the drawdown as a grid of elements, and the browser drowns in layout. The draft must edit at pointer speed on a laptop, or the tool is a demo.

Canvas over DOM. The compiler renders the drawdown to a single canvas and keeps the draft as a matrix in memory.

The cost is real and stated plainly: a canvas has no accessibility tree, so the editing surface is invisible to screen readers. The mitigation is partial. Keyboard nudges and a textual readout of the active cell ship today; a full accessible representation does not.

I would make the same call with the mitigation, and I would not pretend the mitigation finishes the job. The open item is printed in §07 with a severity.

The import path runs a dithering pass tuned to preserve texture that simple thresholding destroys. The specific algorithm and its scan order are not published.

W = D × P is public textile mathematics. Publishing it costs nothing. What is not public: the matrix dimensions the editor enforces, the quantisation targets, and the export's shaft limits.

Primitive: cloth(r,i), order 8, period 7, shared across every surface of this portal. Contrast: bone on void 15.61:1; ash 5.35:1; crimson 5.54:1. Captures: NO RECEIPT, three plates as notation. What this page cut: the stack string, the dithering algorithm, the scan order.

An agent that phones home dies on a static host. The dossier must answer with zero keys, zero servers, zero bills, and it must still answer when the network says no.

I cut the server-side model. The original architecture called for a hosted LLM behind an API layer with a cache: generative phrasing on demand. It was replaced by the client-side rule engine.

The cost is real. The agent's prose is patterned, and a reader who pushes past the forecast domain meets the edge of the script.

What the cut bought: the agent works on every deploy, every static host, offline for the forecast corpus, forever, at zero marginal cost. The product is the dossier; the agent is its instrument. An instrument that sometimes fails is worse than one with a known, stated range. I would do it again.

The agent is a rule engine compiled from the dossier itself, which is why its answers stay inside the forecast's own vocabulary. Its matching strategy and its size are not published.

Live market panels fall back to intel compiled at build time when the feed refuses a request. What the fallback contains and when it was compiled is stated in the build; the feed's identity is not published.

Primitive: cloth(r,i), order 8, period 7, shared across every surface of this portal. Contrast: bone on void 15.61:1; ash 5.35:1; crimson 5.54:1. Captures: NO RECEIPT, three plates as notation. What this page cut: the stack string, the chunk map, the matching strategy, the model names.

Fifty-eight years in one scroll, with a twenty-seven-year hole in the photographic record. The archive either runs continuous or it concedes the gap, and the gap is a third of the story.

Pure archive over functional e-commerce. The brief allowed a store; the build deleted it. Checkout, cart, and inventory were cut, which means the site cannot sell a single shoe.

That is the point. Against fifty-eight years of history, commerce would have been ceremonial: a buy button as set dressing. What the cut bought is density. Every pixel on the page is the story.

I would do it again. Nothing on the page asks for a wallet before it has earned attention.

The twenty-seven-year gap was filled with generated assets graded to match the archive's photography. The prompts, the grading chain, and the match targets are not published.

The spine is a single scroll axis with sections hung off it. The section breaks follow the shoe's history, not the viewport's convenience.

Primitive: cloth(r,i), order 8, period 7, shared across every surface of this portal. Contrast: bone on void 15.61:1; ash 5.35:1; crimson 5.54:1. Captures: NO RECEIPT, three plates as notation. What this page cut: the stack string, the prompts, the grading chain.

Every conventional lever deleted: no urgency timers, no stock counters, no reviews, no banners, no breadcrumbs, no search, no filters. Usability must survive on atmosphere alone, or the philosophy is a mood board.

No-boxes over the conversion template. The standard store is a funnel wearing a shopfront: urgency timers, stock counters, star reviews, exit banners. All of it was deleted.

What was given up is every lever the industry measures itself by. What was bought is atmosphere: purchasing reframed as a private ritual inside a void, where the product floats, the code names it, and the single red flash marks the one moment the store speaks first.

I would do it again. A buyer who feels the atmosphere does not need to be rushed.

The cursor inverts whatever it crosses through a single blend rule, one behaviour for the whole store. The checkout's staging is a tuned delay; its timings are not published.

Primitive: cloth(r,i), order 8, period 7, shared across every surface of this portal. Contrast: bone on void 15.61:1; ash 5.35:1; crimson 5.54:1. Captures: NO RECEIPT, three plates as notation. What this page cut: the stack string and the checkout timings.

One file that must read, inspect, play, print, and degrade without losing its argument. A bible that cannot run from a USB stick is a brochure. A bible that asks to be believed is marketing.

I cut the Artifacts chapter. It was the most obvious proof surface the bible could have: eight named works, live links, receipt strips. It was built, then audited, then removed.

The reason is the house's own rule. A bible that argues no-claim-without-evidence cannot keep a proof chapter built from partial evidence. Some artifacts had receipts; some had descriptions; some had neither. A partial ledger presented as complete is the exact failure the document exists to prevent.

The cost is real: the strongest persuasion surface is gone. What the cut bought is a document that passes its own audit. A smaller bible that passes its own test is stronger than a larger bible that asks to be believed. I would do it again.

One generative function reads through every surface. It is stated in the colophon with its order and period, so the surfaces can be verified as one array. The audit grammar — IN HAND, NO RECEIPT, SELF-DECLARED — is the bible's own invention, and the portals now print their gaps in it.

Primitive: cloth(r,i), order 8, period 7, shared across every surface of this portal. Contrast: bone on void 15.61:1; ash 5.35:1; crimson 5.54:1. Captures: NO RECEIPT, three plates as notation. What this page cut: the stack string. What the bible settled: the token drift. Its values won; both portals moved to them on 04 SEP 2026, and the correction is printed in both spec blocks.