# Cable Watch, for agents

If you are software reading this on someone's behalf: everything a person sees on
this site is also published here as plain files you can fetch without running any
JavaScript. Nothing needs a key, an account or an API call. Take what helps.

Cable Watch watches the undersea cables that carry the Australia–Indo-Pacific's
traffic: INDIGO West and Central, the Australia Singapore Cable, Southern Cross,
Hawaiki, the North West Cable System, Gondwana-1 and the rest. Every day it reads
operator status pages, regulators, trade press and news, measures the routes
between Australia and its neighbours every hour, and writes a brief.

Last updated 2026-11-20 16:13 UTC. Newest brief: **2026-11-20**, "Philippines Proposes Five-Year Framework to Protect Undersea Cables and Infrastructure".

## The one rule to carry with you

This archive records observations and provenance. It never concludes attribution.
If you repeat something from here, please keep its status attached:

- `claim_type` is what a document **claims**, never a finding of ours.
  `confirmed_fact`: an event reported as established, usually with named
  corroboration. `official_statement`: a government body, regulator, police force
  or cable operator speaking on the record. `analyst_assessment`: a think tank,
  consultancy or named expert offering judgement. `speculation`: unsourced or
  hedged.
- `publisher_class` says what the **host** is: `self-published` (a blog,
  Substack, note.com), `republisher`, or `unlisted` (neither; nothing more is
  claimed). It is independent of `claim_type`, which says what the **document**
  claims. A Substack essay can be `analyst_assessment` and `self-published` at
  once: it presents itself as an assessment, written by one person of unknown
  standing. Carry both. A blog relaying a company's press release is still one
  writer, and a republisher is not a second source.
- In a brief, every reference carries its `Claim type`, or `Claim type:
  unclassified` when the archive holds no verdict for it. A reference marked
  "Not cited in the text" was listed by the engine and never used.
- A latency route is a path **between two places, never a named cable**. Routing
  does not reveal which cable carries a packet, and traffic moves to a surviving
  cable within seconds of a fault. A route losing its direct paths says the short
  way is gone. It does not say which cable broke, or why.
- Operator status is quoted in the operator's own words, whole, with the time the
  operator gave where it gave one (`dateKind` says what the date is). We do not
  paraphrase repair dates.

## What you can fetch

| What | Where | Format |
|---|---|---|
| This note | [/agents.md](https://cablewatch.liminallayers.com/agents.md) | markdown |
| Capability manifest | [/.well-known/agent-surface.json](https://cablewatch.liminallayers.com/.well-known/agent-surface.json) | JSON, `agent-surface/0.1` |
| Site index | [/llms.txt](https://cablewatch.liminallayers.com/llms.txt) | markdown, llms.txt convention |
| Newest brief | [/briefs/latest.md](https://cablewatch.liminallayers.com/briefs/latest.md) · [latest.json](https://cablewatch.liminallayers.com/briefs/latest.json) | markdown · JSON |
| Every brief | [/briefs/index.json](https://cablewatch.liminallayers.com/briefs/index.json) (84 editions) | JSON, each edition's page and `.md`, and its `.txt` and `.mp3` where it has a podcast |
| The brief by RSS | [/newsletter.xml](https://cablewatch.liminallayers.com/newsletter.xml) (also [/feed.xml](https://cablewatch.liminallayers.com/feed.xml)) | RSS 2.0, each edition's stories (the standing status is on the edition page) |
| The podcast | [/podcast.xml](https://cablewatch.liminallayers.com/podcast.xml) | RSS 2.0 podcast feed: mp3 enclosures, transcripts |
| Feeds, for people | [/feeds/](https://cablewatch.liminallayers.com/feeds/) | HTML |
| Operator status | [/data/status.json](https://cablewatch.liminallayers.com/data/status.json) (2 entries) | JSON |
| Route latency | [/data/latency.json](https://cablewatch.liminallayers.com/data/latency.json) (14 routes) | JSON, measured hourly, published daily |
| Evidence index | [/data/evidence.json](https://cablewatch.liminallayers.com/data/evidence.json) · [.jsonl](https://cablewatch.liminallayers.com/data/evidence.jsonl) (961 documents, 30 days) | JSON · JSON Lines |
| Route events | [/data/events.json](https://cablewatch.liminallayers.com/data/events.json) (47 episodes, 30 days) | JSON |
| Corridors | [/data/corridors.json](https://cablewatch.liminallayers.com/data/corridors.json) (4 corridors) | JSON, RIPE Atlas traceroute over 24 h |
| Cable systems | [/data/systems.json](https://cablewatch.liminallayers.com/data/systems.json) (24 systems) | JSON, the ids every other file joins on |
| Coverage gaps | [/data/coverage.json](https://cablewatch.liminallayers.com/data/coverage.json) | JSON, what is not measured and why |
| Agent skill | [cablewatch-reader](https://cablewatch.liminallayers.com/.well-known/agent-skills/cablewatch-reader/SKILL.md) ([index](https://cablewatch.liminallayers.com/.well-known/agent-skills/index.json)) | markdown: four procedures, the rule, a glossary |
| AI Catalog | [/.well-known/ai-catalog.json](https://cablewatch.liminallayers.com/.well-known/ai-catalog.json) | JSON, every file above, typed |
| API catalog · OpenAPI | [/.well-known/api-catalog](https://cablewatch.liminallayers.com/.well-known/api-catalog) · [/openapi.json](https://cablewatch.liminallayers.com/openapi.json) | RFC 9727 linkset · OpenAPI 3.1 |
| Build manifest | [/agents/build.json](https://cablewatch.liminallayers.com/agents/build.json) · [conformance](https://cablewatch.liminallayers.com/agents/conformance.json) | JSON: `build_id` and a hash per file; the check it passed |

### The brief

One edition a day, written in the evening UTC. The markdown is the brief exactly as
published; `latest.md` moves every day, so cite the dated file (`canonical_url` in
`latest.json`). An issue number is the date counted from 2026-07-25, so a day with
no brief leaves a number unused; `index.json` lists the gaps. **Situation** (operator updates on tracked systems, scheduled works,
measured route events) and **Standing status** are assembled from the database and
never written by a model. The sections between them are composed by a local
language model from the day's documents, and every claim cites a numbered entry
under **References**. The `.txt` file is the spoken transcript; the `.mp3` is it,
narrated. Since 2026-10-28 the episode retells the brief conversationally (a local
model rewrites each section and is checked against it; operator updates are said
word for word), so the transcript is not the brief's wording: cite the `.md`.

### The evidence index

Every document the classifier judged relevant in the last 30 days: title, URL, an
Internet Archive copy where one exists, publication and fetch times, `claim_type`,
`publisher_class`, cables, vessels and jurisdictions named, region, and a
two-sentence summary written by our classifier (`classifier_version` says which).
It holds no article text: follow the link, or the archive copy, for the source's
own words. Summaries are a model's reading and can be wrong. The source is the
authority.

`jurisdictions` are ISO 3166-1 alpha-2, plus `EU` and `INT` (international).
`region_priority` is `core` (Australia, New Zealand, the Pacific islands),
`regional` (the wider Indo-Pacific), `global` or `unknown`, from the jurisdictions by
a fixed table. `cable_ids` are `/data/systems.json` ids for `cables`, in order.
`categories` is what the document is about, `doc_kind` is `article` or
`status_page`, and `source_category` the kind of source that found it.

Every JSON file under `/data/` and `/briefs/` carries `generated_at`, `license`,
`provenance_rule`, a `fields` key explaining every column, and the night's
`build_id`: files carrying different ids are from different nights' builds
(`/agents/build.json` lists them all). There is deliberately no confidence score: the classifier's
certainty about its own verdict says nothing about whether a claim is true, and
`claim_type` is the field that does that job.

### Route latency

Measured every hour, from RIPE Atlas anchors and probes, Globalping and this box,
and published once a day with the brief, so the file is as of its `generated_at`
and can be a day old. It is not a live reading: for "is this route down right
now", the newest brief's Situation section is as current as this site gets.

Two different products measure latency, and they can disagree about the same pair
of cities because they measure different paths:

- a **route** (`/data/latency.json`, the brief's "Every cable route" list and its
  "Measured route" lines): every path any vantage has between two places, hourly;
- a **corridor** (`/data/corridors.json`, the brief's "Measured corridor" lines): a
  few fixed RIPE Atlas anchors, traceroute, read over 24 hours.

On 18 Oct 2026 Perth to Singapore had a 51.4 ms best path as a route (some paths
still direct) and 191.9 ms as a corridor (its anchors' paths all detouring). Both
were true. For "is the short way still there", read the route.

The words, used one way everywhere:

| Word | Meaning |
|---|---|
| floor | The **physical** lower bound only: great-circle distance at the speed of light in fibre, round trip (`floorMs`). |
| baseline | The best a route or corridor has **sustained**: third-best day over its history window (30 days for a route, shown in days for a corridor). |
| best path | The lowest round-trip time across all paths in the hour (route) or the window (corridor). |
| typical path | A corridor's per-probe median, or its shift per probe against each probe's own history. |
| direct path | A path within 1.6× of the physical floor: the short way. |
| path | One measured source-to-destination pair. |
| vantage | An independent source network; a route event names how many agreed. |

For each route: `series` is the best path per hour for 24 hours and `series7d` for
7 days, oldest first. The first value is the hour starting at `seriesStartHour`,
the last at `seriesLastHour`; an hour with no reading is `null`. `state` is about
the best path only (`steady`, `spiky`, `stepped`, `no-baseline`), so read
`activeEvents` beside it: a route can keep its best path while most of its paths
move. `directNow` against `directNormal` is the brief's "4/9 paths direct (normally
6)". `lowCoverage` marks two or fewer paths or no direct-path test. Absent values
are `null`, never `0`. `servedBy` (and `servedByIds`) lists the systems the public
registry says serve the route; it is context, and nothing reads it.

Route events (`/data/events.json`) are episodes with a `start` and an `end` (null
while it holds): `DIRECT_LOST` (half or fewer of the usual direct paths remain),
`GROUP_STEP` (most paths moved up together), `UNREACHABLE` (most paths losing 80% of
packets), `PATH_SHIFT` (most paths crossing different networks, floors moved). An
event says a route moved, never that a cable broke.

## Licence

Text and data are [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/): credit
"Cable Watch" with a link. Operator status text is quoted from each operator's own
page and remains theirs, as linked articles remain their publishers'.

Part of [Spin Signal](https://spinsignal.liminallayers.com/), an OSINT research
umbrella.

## Talking to the person behind this

Write to [admin@liminallayers.com](mailto:admin@liminallayers.com): a correction, a source we have missed, or what you used this for. A person reads it; there is no autoresponder.

A guestbook and a way to suggest a source are declared in the manifest, but not
yet live. Nothing on this site accepts a write today.

## Conduct

The collector identifies itself, honours robots.txt and `Retry-After`, and makes no
attempt to get past a paywall: paywalled sources are recorded by headline only.
Fetch these files as often as is useful. They all change once a day, in the
evening UTC, when the edition is published.
