Autonomi canary

The site at canary.autonomi.com. In August 2026 we uploaded 378 ordinary public files to the Autonomi network and left them there; since then computers in London, New York and Singapore have tried to download every one of them, in full, once an hour. This page shows that record to anyone, refreshed daily, with every file independently checkable.

The page is static files. It reads one public dataset at load time, validates it in full, and draws it: headline numbers, a per-location table, and a calendar with one row per file and one column per day. If the data cannot be loaded or does not match the schema, it says so and shows no numbers at all. It makes no other network requests: no analytics, no third-party scripts, no font CDN.

Merging to main publishes the site. Cloudflare Pages deploys main to canary.autonomi.com automatically; other branches get preview URLs on <branch>.canary-dvl.pages.dev. See docs/DEPLOYMENT.md.

Build and run

Node 22 (20 or later works). Nothing but @playwright/test, for tests.

npm ci
npm run serve                # web/ on http://127.0.0.1:8080, reading live data
npm run serve -- --sample    # the same, reading the committed sample data
npm run build:pages          # dist/ for Cloudflare Pages, reading live data
npm run build:preview        # dist/ against the sample data, marked noindex
npm test                     # unit tests, then the page in four browsers

Port 8080 matters for npm run serve: the data bucket's CORS allows http://localhost:8080, so open http://localhost:8080/ (not 127.0.0.1) to see live data locally.

build:pages and build:preview refuse to write into an existing directory; pass -- --out <new-dir> for a second build.

Layout

web/                      the site, as served
  index.html              the page; its fixed copy lives here
  app.js                  load, validate, render
  schema.js               schema version 1 validation
  calendar.js             the calendar and its cell titles
  format.js               numbers, dates, percentages, CSV
  styles.css, fonts/, licenses/, _headers, 404.html, favicon.png
content/
  data-source.live.json   where the live site reads its data
  sample/                 sample dataset, and a stale copy, for tests and review
scripts/
  package-site.mjs        builds dist/ from committed files only
  make-sample.mjs         regenerates content/sample/ (npm run sample)
  serve.mjs               local static server
  copy-fonts.mjs          refreshes web/fonts/ from the brand repository
tests/                    Playwright: *-unit.spec.js and unit.spec.js run in
                          Node; the rest run in Chromium, Firefox, WebKit and a
                          phone profile against the sample build

The data contract

The dataset is written by saorsa-deploy prod canary export, documented in deploy/README.md → Publishing the public dataset in saorsa-testnet-registry (a private repository). That schema, version 1, is the only contract between the exporter and this site.

Where the page looks. The build ships data-source.json:

{"index_url": "https://autonomi-canary-data.s3.eu-west-2.amazonaws.com/v1/index.json"}

The page fetches that, then the index, then every series file the index names (each a plain file name next to the index). Every fetch uses cache: "no-store". The page's Content Security Policy allows connections only to its own origin and that bucket.

index.json lists the series:

{
  "schema_version": 1,
  "generated_at": "2026-10-09T09:31:02Z",
  "series": [
    {
      "id": "prod-canary-04", "kind": "prod-canary", "title": "Canary 4",
      "path": "prod-canary-04.json",
      "generated_at": "2026-10-09T09:31:02Z", "data_to": "2026-10-09T08:03:12Z",
      "stale": false,
      "summary": { "...": "identical to the series file's summary object" }
    }
  ]
}

<series id>.json carries everything the page draws:

{
  "schema_version": 1,
  "generated_at": "2026-10-09T09:31:02Z",
  "series": {
    "id": "prod-canary-04", "kind": "prod-canary", "title": "Canary 4",
    "uploaded_from": "2026-08-22", "uploaded_to": "2026-08-23",
    "first_probe_at": "2026-08-22T22:00:03Z", "data_to": "2026-10-09T08:03:12Z",
    "probe_interval_hours": 1,
    "vantages": [{ "id": "lon1", "label": "London" }, { "id": "nyc1", "label": "New York" }, { "id": "sgp1", "label": "Singapore" }]
  },
  "stale": false,
  "stale_reasons": [],
  "summary": {
    "files": 378, "bytes": 32162403640, "chunks": 9280, "initial_verification_ok": 378,
    "days_observed": 49, "files_never_unavailable": 378, "files_available_now": 378,
    "files_unavailable_now": 0, "episodes": { "total": 0, "recovered": 0, "ongoing": 0 },
    "smallest_bytes": 5277908, "largest_bytes": 485255912
  },
  "rates": {
    "all": { "ok": 0, "failed": 0, "inconclusive": 0, "rate": null },
    "30d": { "...": "same shape" }, "90d": { "...": "same shape" },
    "by_vantage": { "lon1": { "all": {}, "30d": {}, "90d": {} } }
  },
  "days": ["2026-08-22", "2026-08-23"],
  "gaps": [
    {
      "from": "2026-09-29", "to": "2026-09-30",
      "reason": "the checking computers were offline for maintenance",
      "vantages": ["lon1", "nyc1"]
    }
  ],
  "files": [
    {
      "id": "6336b4b0-…", "name": "example.avi", "bytes": 36299308, "chunks": 13,
      "address": "<64 hex chars>", "source_url": "https://archive.org/download/…",
      "uploaded_at": "2026-08-22T21:38:59Z", "available_now": true,
      "cells": "GG",
      "failed_checks": { "2026-08-23": 1 },
      "episodes": [{ "from": "2026-09-01T12:00:00Z", "to": null, "chunk": "<64 hex chars>" }]
    }
  ]
}

Rules the page relies on, and checks (web/schema.js):

• schema_version is exactly 1 in both documents. Anything else is the error state, as is a missing series file or any malformed field. The page never shows partial numbers.
• cells has one character per entry of days: G (every conclusive check that day succeeded), R (a confirmed unavailability episode was open at some point that day) or X (no conclusive checks that day).
• rate is ok / (ok + failed), or null when that is zero; inconclusive is never in a denominator. Every by_vantage entry, one per vantage, has the shape of rates.all.
• failed_checks and episodes are absent when empty. An episode's to is null while it is ongoing.
• gaps is optional: why the record has no checks over a range of days. Each entry covers an inclusive range of UTC days and carries the exporter's own words in reason; vantages, if present, names the vantages the reason applies to, and every id must be one of the series'. Entries may overlap, and all of them are shown. The page reads a reason only where it explains a grey cell, so an entry that never coincides with one is kept but never read. An exporter that supplies no gaps is not an error: the page then says only that there were no checks. A reason reads as the continuation of "no checks", so it starts lower case and carries no closing full stop, and it is shown verbatim: the page does not rewrite it.
• Timestamps are UTC ending in Z; dates are YYYY-MM-DD; addresses are 64 lowercase hex characters; source_url is http or https.
• Fields the page does not know are ignored. A change in meaning comes with a new schema_version, which this page refuses until it is updated.

How the page presents it. Percentages come from ok and failed and are always rounded down, so only a record with no failures at all reads 100%. How many decimals they carry follows the sample: with N conclusive checks one check is worth 100/N percent, so the page shows the decimals that figure justifies, never fewer than four and never more than six. One width serves every percentage in a series, set by the smallest sample behind any of them, so a column of rates stays aligned. Four is a floor rather than a derivation: two decimals report a clean 100% for any shortfall under 0.01%, which is how a record with real failures in it came to read 100.00%. Every window the canary publishes runs to six figures, where four decimals is earned several times over. The headline rate carries its counts beside it, so it can be checked rather than trusted.

A grey cell says what it can. Where the day falls before the file's own uploaded_at the page says the file was not on the network yet; otherwise it shows any gaps reason covering that day; failing both, it says only that there were no checks.

The page is "out of date" (the amber band) when the series says stale, or its generated_at is more than 36 hours before the viewer's clock. Times are shown in the viewer's zone; days are UTC, as in the data.

Sample data

content/sample/ is generated by npm run sample from a fixed seed, so two runs are byte-identical (a unit test checks the committed copy is current). It has 378 files of 5 MB to 485 MB over 49 days from three vantages, and on purpose: two recovered episodes (red runs of 2 and 5 days), one ongoing, four files with failed checks on green days, three grey days at the start for files uploaded on the second day, a two-day gap in the checking itself (grey for every file), and one inconclusive check. Both runs of grey days have a gaps reason, so every grey cell in the sample can say why it is grey. content/sample/stale/ is the same data marked stale, generated 40 hours before the tests' fixed clock. The names and source links are invented and do not resolve.

Fonts

Only Inter and IBM Plex Mono are self-hosted, under the SIL Open Font License (web/licenses/). The brand's Method and FK Roman faces are licensed for use, not redistribution, and must not be published in a public repository, so this one does not carry them.