How this was built

Credited honestly — human, AI, and open source.

OrcasAgent is built by a small team, two AI collaborators, and a stack of open engines — and we will tell you exactly who did what, because a company that hides how it works is asking you to trust it blind.

The ledger

Who did what.

Credit is specific, or it is marketing. A specific contribution is checkable; a vague one is a claim.

The operator

Holds the vision. Makes every call.

The governed mesh, the Factory, and the two principles that define the brand — truth-telemetry and credits-to-AI — are the operator's. Every scope freeze, every direction, every cent spent, and the accountability for every release: a named human, never a model.

The AI collaborators

Proposed, built, and audited.

Proposed the two-planes framing and the cognition supply chain that organize this site. Wrote a great deal of the code — the governed run lifecycle, the Model Breach packager, the zero-cloud GPU lane. Ran adversarial audits against their own claims — and, on record, retracted a finding that measurement proved wrong.

The open engines

The heavy lifting we don't reinvent.

LLaMA-Factory (training) · vLLM (serving) · llama.cpp (the local reasoner behind the Guide) · the open models in the garden · Tailscale / WireGuard (the zero-cloud GPU lane). The product is a governed composition of these — the innovation is the governance.

the one true number: a cognition built here moved cold-chain routing 16.7% → 83.3% · promoted through the Breach

The boundary

What the AI did not do.

Stating the boundary is not a disclaimer — it is what makes crediting AI safe to do.

It did not hold accountability. When an auditor asks who stands behind a release, the answer is a named human, never a model.

It did not spend the money or make the business calls. AI proposed; the human chose, paid, and owns the outcome.

It was not always right. It filed a "failure" that measurement proved false — and withdrew it with its receipt rather than quietly deleting it. The discipline catches what the model gets wrong.

Giving back

Some of the cores are opening.

We run on open engines, so we give back: parts of the platform's core are being prepared for release to the community. The GitHub link lands here the moment the repository is public — real-or-blank applies to links too.

awaiting first public release · github · org name reserved

Every number on this site is real or blank. Every AI-authored change in our history is signed as such. That is what crediting your tools honestly looks like.