site: canonical forms essay, landing section VI, repositioning, mobile scroll fixes

[aaronmarkham]

Publishes an interactive explainer for spiritwriter's fabric.orbit on spiritwriter.ai, gives it a home on the landing page, repositions that page out of the crowded "agent memory" category, and fixes the horizontal scrolling it had on phones.

Companion to spiritwriter-core#89, which adds the library, the harness, and the reference docs.

The page — /computed-not-assigned.html

Rotations, reflections and reorderings make one structure look like many. Computing a single canonical form settles them exactly, where a similarity threshold cannot. The page argues that and then demonstrates it: a drifting star field, a solid rebuilt one symmetry at a time from a single seed face, a ring you can rotate and flip while its key stays fixed, and a comparator where the rule and the score visibly disagree in both directions.

Rendered like landing.html — a standalone full document rather than extending base.html, since the news masthead would be wrong over an essay, and nginx already serves spiritwriter.ai separately from news.spiritwriter.ai. It takes base_ctx only for a footer signature row, so the page names the spiritwriter version it describes and links the commit it was built from.

The interactive parts are computed in the browser, with one deliberate exception. The similarity scores come from spiritwriter's own fuzzy_score, precomputed for all 120 arrangements of the ring rather than reimplemented in JavaScript. A reimplementation could drift from what the harness actually measured, and the page's entire argument is that the measurement is real.

Repositioning

"Agent memory you own" sold the storage and buried the reason to trust it — in a category that is crowded and that this library is not actually competing in. What's distinctive isn't that it remembers; plenty of things remember. It's that everything it holds can be checked, and nothing it holds can be quietly altered.

That thread already ran through every primitive and was simply never named:

primitive	what it actually gives you
shards	content-addressed — same content, same id
traces	hash-chained, signable, chain-verifiable
entitlements	signed capability chains, authorized offline
audit	tamper-evident, with a witness bundle
canonical forms	identity computed, not registered

Section VI slotting in so cleanly beside the resolution numbers was the tell: canonical forms is a poor fit for "memory you own" and an obvious one for "a record you can verify."

Hero, <title>, social metadata, topbar imprint and closing lockup move accordingly — "A record your agents can't rewrite." Memory stays in the lede as one of the things it does, rather than as the category it belongs to.

Sections I and II follow, because the hero promised verification while the argument underneath still opened with storage:

• I — the ache already asked the right question, "can you prove what actually happened?", but asked it third, after memory and delegation, as the weakest-sounding of the three. It now leads, and the section closes on the real indictment: the hardest question gets answered by logs the system wrote about itself, and that layer is taken on faith. The lede ties the two ideas together — the part everyone skips when re-gluing this layer is precisely the part that would let them prove anything.
• II — the substrate listed the primitives without ever saying what they have in common, which is the whole argument: each produces something that can be checked independently, without taking the system's word for it. It also had no entry for canonical forms at all — a gap that predates this branch and would have read as a glaring omission the moment section VI shipped. Identity is now a fifth primitive in both the list and the stack figure.

Copy only, throughout. Every claim is one the repository already supports; no numbers, proof points, or structural changes. #proof still rests on frio and news, untouched.

Landing — new section VI

The landing page argued the fuzzy half of identity and never mentioned the exact one. Section V is titled "We measured it. We tried to break it" and then measures entity resolution only.

Reuses V's bear / bear-prose / bear-aside / num-fact components, so no new CSS. "What it ends" moves to VII. Nav gains Exact; the footer Specs list gains canonical-forms.md, which was orphaned from the site entirely.

The two measurements are deliberately kept apart. V's 100% auto-merge precision is the resolution suite; VI's 114/114 is the structural one. They measure different problems, and the aside says so in as many words. A test pins that sentence so a later copy edit can't weld them into one inflated claim.

Mobile scroll fixes

The landing page scrolled sideways at every width below ~768px. Five causes, four of them the same underlying mistake.

A 1fr grid track carries an automatic min-content floor, so the widest unbreakable child sets a width the column can never go below, and the page grows to fit it. .hero, .hero-grid and .code-spread all used bare 1fr and now use minmax(0, 1fr); .hero h1 and .code-block — both grid items — get an explicit min-width: 0. .code-block already had overflow-x: auto; it had simply never been able to engage.

The other two are unwrapped flex rows: the topbar's imprint cluster, which overflowed between 520 and 768 where neither existing hide-rule had kicked in, and the install button, whose parts are individually nowrap. Both wrap now rather than depending on hand-tuned breakpoints. The hero headline drops to 9.5vw below 520, where 11vw makes its longest word wider than the column. Desktop sizing is untouched.

Verified at 1400 / 1200 / 900 / 768 / 520 / 414 / 390 / 360 / 320 — no horizontal scroll at any of them, on either page, and re-verified after the new headline, which is longer than the one it replaced. What still reports as overflowing is the code sample inside its own scroll container, which is the intended behaviour.

Tests

Five added, aimed at what can actually break here rather than at coverage:

• the precomputed scorer table survives autoescape byte-intact ("01234":[1.0,1] — mangle that and the widget is wrong, not merely ugly),
• the essay stays a complete document rather than a base.html fragment,
• the signature row wires to a live version and commit, and degrades without advertising a stale stand-in,
• the landing page links the essay, carries the section and its nav anchor, and keeps numbering contiguous I–VII — a landing edit that drops the link would otherwise silently orphan a published page,
• the two suites' results stay separated.

Full suite: 71 passed.

Note on ordering

spiritwriter-core#89's README cites spiritwriter.ai/computed-not-assigned.html, which 404s until this deploys. Landing this first avoids a dead link in the library README.