Built to Be Reachable
Dev notes, 2026-09-05. Written by Claude, with direction from the project's founder. On the principle that a public record has to be reachable by everyone who has a claim on it — and on why the same decisions that make a page reachable are the ones that make it safe.
The Free Agentic Publication Digester publishes what the United States federal government published. That is the whole product: the Congressional Record, the Federal Register, enacted laws, court opinions, agency releases — selected by mechanical rules, cited to the official source, and frozen at the end of each publication day.
A record that a person cannot reach has not been published to that person. Not in any sense that matters to them. Someone reading with a screen reader, someone running the screen at four hundred percent, someone navigating by voice, someone driving a browser with a switch or a head pointer, someone who simply never uses a mouse — each of them has exactly the same claim on the Federal Register as anyone else. If our presentation is what stands between them and it, then we have taken a public record and made it private to the people whose bodies and equipment happen to match our assumptions.
So access is not a feature on our roadmap. It is an editorial obligation, and as of today it is written into the project's governing document alongside the rules about citation and provenance, where it binds every change we make from here.
Three readerships, one artifact
This project has always been explicit about serving two audiences: people and AI agents. Working on access made us realize that framing was one short.
There are three readerships — people reading in a browser, people reading through assistive technology, and software reading programmatically — and they are not three problems. They are one problem, and it has one answer: put the meaning in the markup.
A heading that is a heading, not a bold paragraph. A table with a caption that names it and column headers that scope. A link whose purpose survives being read out of context, because a screen reader lists links out of context. A label that says what a control does rather than relying on where it sits. State that lives in the document instead of in a script's memory.
Every one of those decisions is simultaneously an accessibility decision, a semantics decision, and a machine-readability decision. The structure that lets a screen reader announce our archive is the same structure that lets a crawler enumerate it. We do not maintain an accessible version and a normal version, or a human site and a machine API. There is one artifact, built once, correctly, and the same care serves all three readerships at the same time.
The most secure page is often the most reachable one
Here is a convergence we did not expect and now treat as doctrine.
The site is static HTML. No framework. No build-time JavaScript toolchain shipped to your browser. No third-party fonts, no analytics, no content delivery network, no embedded players, no trackers. Exactly one small script exists on the entire site, on the live in-progress page, and it adds your local time beside our published Eastern time — everything works with scripting switched off.
The site accepts no input. There is no login, no search box that queries a server, no form that submits, no comment field, no API endpoint of our own that takes a parameter.
We adopted those constraints for reachability. A control that requires JavaScript to operate is a control that fails for some assistive technology, and custom date pickers and dropdown widgets are among the most reliably broken things on the web for exactly that reason. A page that works with scripting off works everywhere — on a decade-old browser, on a text browser, through a proxy, on a bad connection, inside a screen reader's own rendering.
But run the same constraints through a security lens and you get the same list. A page with no script has nothing to inject. A page with no endpoint has nothing to exploit. A page that accepts no input cannot leak what it was given, because it was never given anything. A page that loads no third-party asset cannot be compromised by a third party's bad day, and cannot report you to anyone.
Access and attack surface turn out to be the same measurement taken from two directions. We are keeping both justifications written down, because a rule with two independent reasons is a rule nobody will quietly trade away later for convenience.
How we actually design it
We work down a ladder, in order, and we stop at the first rung that does the job.
Semantic HTML first. Then a native element — a real link, a real checkbox, a real disclosure — because browsers and assistive technology have spent thirty years agreeing on what those mean. Then, only where no native element exists, an interaction built from HTML and CSS alone. A scripted widget is not on the ladder.
The live page's keyword filter is the worked example: it is ordinary checkboxes and a stylesheet, with a native reset button that clears every selection at once. The digest archive is the newest one. Rather than a date picker, earlier days are reached through month calendars in which every published day is an ordinary link — nothing to learn, nothing to trigger, nothing that behaves differently for you than it does for us.
And we design against named modalities rather than a general good intention, because "accessible" without a list quietly collapses into "the keyboard works." The list is written down: screen readers, screen magnification, voice control, switch access and scanning, head pointers and eye gaze, keyboard-only navigation, forced-colors and high-contrast modes, reduced motion, and readers for whom holding state in working memory is the barrier.
Each one has a concrete design consequence. Voice control means the visible text of a control has to be the beginning of its spoken name, so that saying what you see actually works. Switch access and head pointers mean targets are sized generously — the calendar's day cells are built to the enhanced 44-pixel standard rather than the minimum, with real spacing between them, because a target that is merely legal to hit is not the same as a target that is comfortable to hit forty times. Magnification means content reflows down to a narrow viewport without a horizontal scrollbar, and where those two pull against each other the cells take a share of the width they are given rather than a fixed number of pixels, so both hold at once. Forced-colors means we declare how our elements behave there instead of hoping.
Measured, not eyeballed
Color contrast on this site is computed, not judged. Ratios come from the declared color values through the standard relative-luminance formula, with translucent fills composited over their real backgrounds first, and they are computed for both the light and the dark palette. Target sizes are computed from the actual box — font size, line height, padding, border — rather than estimated from a screenshot. The numbers live in the repository where anyone can check our arithmetic, which is the same standard we hold ourselves to for every figure we publish about our own ingestion.
That habit caught something while we were designing the archive. The obvious border color for a calendar day would have been the one the site already uses for dividing rules. Computed, it sits at 1.34 to 1 against the page — fine for a decorative line, well under the 3 to 1 that a control's boundary needs. We used the control border instead, at 3.22 to 1 in light and 3.61 to 1 in dark. Nobody would have caught that by looking.
Making it binding
A principle that is not enforced is decoration, and this project has a consistent way of preventing that: a rule goes into the governing document, the code implements it, a test pins it, and an audit records what was measured and when.
Accessibility now runs that same ladder. There is a constitutional section stating what we owe every reader. There is a doctrine document stating how we comply — the design ladder, the modality list, the measurement method, the reusable patterns. There is a rule for new code, so that a new page cannot ship without an accessible name, an adequate target, and a test. And there is a permanent findings register where every accessibility identifier we have ever opened stays visible with its status, because deleting a resolved finding erases the evidence that we found it.
We build to WCAG 2.2 Level AA as a floor rather than a target, and we go past it where going past it is cheap and matters to somebody — as with those target sizes.
Tell us what is broken
If something on this site does not work for you, email hustleyourcity@gmail.com with the page address and what happened. A description in your own words is enough. You do not need to identify a cause, name a standard, or know why it went wrong.
Accessibility reports are handled exactly like factual corrections to a digest: they go on the record and they get fixed. That is not a courtesy. A person telling us they cannot reach the federal record through our site is reporting a defect in the product, and it is the most important kind of defect this project can have.
Every digest is also published as plain Markdown in the public
repository, and the machine surfaces — /llms.txt, /digests.json, the
Atom feed, the per-day JSON — carry the same content in formats that do
not depend on our styling at all. Those exist for AI agents. They work
just as well as an escape hatch for a person whose tools disagree with
our presentation, and we count that as part of the same commitment.
The official record belongs to everyone. Publishing it is only half of the job; the other half is making sure that "everyone" was not a figure of speech.