The system itself

This page exists because a site that rates other people's changelogs should be legible about its own machinery.

What this is

An archive of how the AI ecosystem changes. We watch releases from coding harnesses, model providers and protocol specs; preserve the original source permanently; and publish a normalised account of what changed, how much it matters, and what it affects.

The website is one interface. The product is the Git-backed, Markdown-canonical archive underneath — which is why every page here has a .md twin and a JSON representation, and why the source is public.

The pipeline

  1. Collect. Each tracked entity has one or more detectors — GitHub releases, in-repo CHANGELOG.md, npm, blogs. Anything new is written to data/raw/ and never edited again. Raw is sacred: if our interpretation is wrong, the original is still there to re-read.
  2. Eventify. An LLM turns a raw source into a structured event: title, summary, categories, significance, breaking flag, and a confidence score. It is instructed never to state a fact absent from the source, and never to reproduce vendor text verbatim.
  3. Review. Confidence below 0.75 flags the event needs_review and it waits for a human. Currently 27 events are in that queue, and they say so on their own pages. A review backlog is a health metric, not something to hide.
  4. Derive. Everything downstream — the index, the search, the graph, this site, the API, the CLI, the MCP server — is regenerated from the event files. Delete all of it and it rebuilds identically.

Significance

A 1–5 judgement, and the number this site's credibility rests on.

5Ecosystem-shifting — a new frontier model, a protocol-level change, a major harness launch.
4Changes decisions — a breaking change in a widely used tool, a major pricing move, a significant new capability.
3Worth knowing — a notable feature, new model support.
2Routine — fixes, minor features.
1Noise — patch bumps, docs.

Significance is inherently per-audience: a Cline user weights Cline events higher than we do. The number is a default, not a verdict.

Rules we hold ourselves to

  • Provenance or it doesn't ship. No event exists without at least one source URL. The schema rejects it at write time.
  • We don't mirror. Vendor release notes stay on the vendor's site. Republishing them verbatim would add nothing, compete with the source for rankings, and be legally gray. You get our synthesis and a prominent link.
  • Facts and opinion never blend. Community reactions live in a separate module from the record of what changed, visually and structurally. Credibility dies the moment those mix.
  • History is append-only. We never edit an event to fix it. A correction is a new event pointing at the old one — including when the mistake is ours.
  • Absence is disclosed. Coverage begins at each entity's first tracked release. Older gaps are gaps, not silence.

What this is not, yet

Honest list. The community layer is empty — the curated creator list that feeds it is not written, so rather than fake it, every event says so. Five of the six verticals are not publishing: their manifests exist but their golden evaluation sets have not passed, and we don't publish a classifier we haven't measured. There is no paid tier, no alerts, no webhooks.

Currently tracking 11 entities and 393 events.

For agents

This site is built to be read by software. Append .md to any URL, or send Accept: text/markdown. Start at /llms.txt, or use the skill and MCP server. Don't scrape the HTML — there is a better surface for whatever you need.