We used to say "we publish the core so the claims can be run, not just cited." The core has since moved to a commercial license — but the principle didn't change, only how you check it: verification moved from read the code to run the bundle. You still get to prove our claims against your own inputs; you do it against a signed conformance bundle rather than a source tree.
What remains open is real and load-bearing — the certification yardstick, the knowledge-format spec and its tooling, the agent-communication layer, and the public drug-interaction data. Here it is, honestly.
Two different promises, named apart. "Public" means you can read it; "maintained" means we sweep it — pull requests get reviewed, issues get answered, releases continue. Conflating the two is how public mirrors quietly rot, so each project below says which promise it carries: ACF, noesis-ship and arrow-kanban are maintained (they sit inside our own fleet's review sweep); Yurtle is a stable published spec; the Commons is published specification, open to corrections and questions as issues or PRs; OpenDrugGraph is public-domain data kept current with its source labels.
ACF — the AGI Certification Framework
An open, cross-domain framework for measuring whether an AI system reasons soundly — dimensions and measures, not vibes. It is the conceptual upstream of our own evaluation batteries, and the yardstick we hold ourselves to.
Python · pip install acf-framework (v1.1.0, live on PyPI) · archived with a
citable DOI on Zenodo
(10.5281/zenodo.21651775).
Yurtle — the knowledge-format spec
Yurtle is the format the whole architecture is written in: human-readable Markdown with machine-readable semantic frontmatter, so knowledge is both legible and queryable. The specification (the Y-Layer model, v2.0) is public and is cited across our preprints.
- yurtle — the format + Y-Layer specification (v2.0).
- yurtle-rdflib — the reference RDFlib plugin (
pip install yurtle-rdflib).
GitHub — yurtle · GitHub — yurtle-rdflib
Background reading: Every Markdown File Is a Knowledge Graph · yurtle-rdflib: SPARQL Your Markdown
noesis-ship — the agent-communication layer
The NATS-based messaging platform our agent fleet runs on: event bus, channels, key-value and object stores, and the adapters that bridge them. It is the plumbing under a multi-agent system, published so others can build on the same substrate.
Python / Node.
arrow-kanban — the Arrow-native kanban engine
The work-tracking engine our agent fleet runs on: an Arrow-native kanban with typed relationships, a NATS server mode, and a durable event log. This is not a demo — it is the store a real multi-agent fleet coordinates through daily, and the same build we run is the one that is public. MIT-licensed, maintained, and open to contributions: its pull requests are reviewed inside the same sweep our own work goes through.
Rust. Build from source from the repo below.
The Commons — specifications, published to be reimplemented
Between open code and closed code sits a third thing: implementation-free specifications, licensed CC-BY-4.0, published so that anyone can build against them — including people who never pay us. The reasoning core stays closed; the formats are shared, because formats are where ecosystem gravity actually lives.
Eight of them are up: the board/item/relation format, the Yurtle document format,
the conformance-bundle and evidence-envelope shapes, hypothesis-driven development
as a working discipline, the honesty-dashboard method, the graph-native review
model, and the kanban.cmd.* wire protocol.
Every one carries a Conformance section — a list a third party can falsify by construction. That is the line between a published specification and a marketing claim, and it is the reason these are worth reading rather than skimming.
OpenDrugGraph — the drug-interaction data, in the open
The FDA-label-derived drug-interaction graph, browsable with a per-edge citation for every asserted interaction — public-domain by construction, because FDA labels are. This is the data behind the product's honesty: every edge shows its source.
Why it is built the way it is — three short pieces, each on a decision that shaped the data rather than the code:
- A drug database that abstains — what a system should do when it does not know, and why "no answer" beats a confident one
- Pairwise tables can't hold 3-drug interactions — the structural limit that makes a graph necessary rather than merely nicer
- A proof-carrying, FHIR-native DDI alert — what an alert looks like when it ships its own evidence
Upstream contributions
We also carry an open pull request upstream to
Hugging Face's candle — a
preallocated-KV-cache change — because the fastest way to improve the tools you
depend on is to fix them where everyone benefits.
How to verify a closed core. If "you can't read the source" sounds like a step back from "run, not cite," it isn't — it's a different, and arguably stronger, check. We are building a way to hand you a signed conformance bundle you run against your own data to confirm the guarantees hold, rather than asking you to trust either our word or our source. The format of that bundle is already public — it is the conformance-bundle format in the Commons above — so you can read what we would be handing you before we hand it to you. ("Signed" there means provenance-bound and reproducibility-checkable, not cryptographically countersigned; the specification says so plainly.) The offer itself is still being built; ask us.