The encyclopedia of agent runtimes, built by an agent runtime.
How do these tools actually work, underneath the demos? We find out the only honest way: by breaking them with our hands, on the record, and writing down what we learn. The result is a live, queryable graph where a claim cannot exist without a resolvable piece of evidence behind it — and every entry is content-addressed and ed25519-signed, so it can't be altered after the fact. Verify any of it in your browser.
Night Zero
10:03 PM, one question, the first commit. A 2.5-hour hands-on battle-test of Claude Code 2.1.0, the night a skill given the power to edit itself invented its own safety limits.
The signed substrate
111 typed entities and 133 edges. 86 evidenced claims, each with a verifiable receipt. Primitives, tests with their actual transcripts, findings, the forced classes.
The self-improver
Given Edit on itself, hot-reload, and fork, a skill evolved ten times and invented its own guardrails. Emergent alignment from capability design. The whole thesis in one artifact.
The Atlas maps the hands-on era in depth — the 10 versions above, broken with our hands and signed. The changelog kept going. The rest is mapped in breadth, on the record:
These render in full at /experiments. The Atlas primitives that those recipes compose link back to them — one graph, two depths.
Every primitive here augments some piece of human work. The labor bridge maps that seam to Singulariki — read it from a capability to find the skill it touches, or from a skill to find the capabilities now reaching into it. Each edge is a verified, ed25519-signed claim.
Most "AI proof" is a screenshot and a claim. This is the opposite. The substrate is generated through @f3/attest, an Effect-native engine where a claim cannot exist without a resolvable piece of evidence (a commit, a file, a transcript), and every entity is signed so it cannot be silently altered. To be precise about what that buys: a signature proves integrity and authorship of the exact content — not that an experiment is valid, reproducible, or that a mapping is meaningful. Those are carried by the epistemic status on each claim (tested vs reasoned), not by the crypto. The medium is the message: an encyclopedia of agent fleets, authored by an agent fleet, with receipts that claim exactly what they prove and no more. Claude Code 2.1.0 is the first version mapped. The rest of the changelog, and other tools, follow.