Resources
Changelog
Release notes by area, covering governance and connections through to risk, observability, integrations and developer tooling. Factual, not promotional.
Factual, not promotional
The standard is written down before the first entry rather than after fifty, because a changelog loses this gradually and nobody notices the day it happens.
Entries
0
Areas
6
Nothing has shipped publicly
An empty changelog, and the standard the entries will be held to.
There is nothing to list, so the useful thing to publish is the rule. A changelog is the one page on a site like this with a natural resistance to spin, and most products lose it gradually — a bug fix becomes 'improved reliability' and the page stops being worth reading.
Release notes
Nothing has been released publicly yet. Backfilling entries for internal work would be inventing a history, and dating them would be inventing a timeline.
What it will contain
- What changed, in the words somebody would use to describe it to a colleague.
- Behaviour changes marked as such, including ones that are technically fixes but which you might have been relying on.
- What is still missing in a partial release, named specifically rather than omitted.
- A version and a date, once there are versions and dates.
Factual, not promotional
The same change, written two ways.
This is the drift worth guarding against, and it never happens deliberately. An entry gets rewritten to sound better, then the next one matches it, and eventually the changelog is a feed of announcements that tells a developer nothing.
01
The first tells you what you can now do. The second tells you how to feel about it.
- factual
- Replay now accepts an unpublished draft version.
- promotional
- Powerful new policy simulation capabilities.
02
The first is checkable and explains a behaviour you might have been relying on. The second is what you write when you do not want to say a duplicate was possible.
- factual
- A retried evaluation with the same idempotency key returns the original decision rather than producing a second one.
- promotional
- Improved reliability and consistency.
03
The first tells somebody planning a compliance review whether to wait. The second tells them nothing at all.
- factual
- The BigQuery connector does not yet expose residency metadata. Reads work.
- promotional
- Expanded data platform support.
The only people who read a changelog are the people it is easiest to lose.
Somebody debugging a behaviour that changed, or deciding whether to upgrade. Neither is served by an announcement, and both notice immediately when the page starts producing them.
How it will be organised
By the 6 areas the product is already organised into.
Rather than by release number, which tells a reader nothing about whether an entry concerns them. Somebody who cares about policy does not want to read the connector entries to find out.
- Connect
- Bring existing AI systems, agents, models, data and tools under one layer.
- Understand
- Know what AI exists, who owns it, and how it behaves.
- Govern
- Identity, accountability, policy and data controls.
- Control
- Risk, approval and the boundaries an action runs inside.
- Observe
- Activity, traces, incidents and the evidence record.
- Improve
- Routing, evaluation, posture and continuous governance.
Two things every entry will carry
- Whether behaviour changed
- Marked explicitly, including when the change is technically a fix. Somebody may have been depending on the old behaviour, and “it was a bug” does not help them at two in the morning.
- What is still missing
- A partial release says what does not work yet. Shipping half a connector and describing it as support is how a catalogue stops being trustworthy — which is why the integrations pages name the gap on each entry.
Where to go
The roadmap is the closest thing to a changelog today.
It is assembled from the gaps this site already admits, so nothing on it is aspirational — and unlike this page, it has content.