We approved the Aigarth evolution PEP on August 12, 2026. To set the stage for what that means, here is what shipped in the eight months before it.
The platform, in numbers
- 12 services running in production: identity, qubic, compute, gateway, billing, ann, marketplace, tissue, dataset, economy, training, and the new work runtime.
- 26 database migrations across all services, with a per-service migration table to keep the journals clean.
- 5 training recipes (mlp, cnn, text, gradient boost, and the trinary classifier that powers the consensus tissue).
- 5 consensus policies for tissue-level decisions: majority, unanimous, any, veto-aware, and short-circuit.
- 3 ADRs governing the platform (Trinary Protocol v1, Dataset Licensing, and the Aigarth evolution ADR 005).
The phases that matter most
Phase 0-9: The foundation
Identity with JWT auth, the qubic client, compute reservations, gateway routing, billing and credits, the first ANN registry, the marketplace, and the tissue service. None of this is exciting on its own. All of it is required.
Phase 10-14: The intelligence layer
The ANN registry grew from a single class to a versioned, licensable, deployable model. The dataset service added schema sniffing and connectors. The first experiments with retraining and A/B shadowing showed up here.
Phase 17: Material science, end to end
An 8-ANN coordinator (Director, Literature, Simulation, Physics, Design, Optimization, Experiment, Validation) showed that a swarm of small, specialized AIs can outperform a single generalist on a real research task. The cost model showed that DFT relaxation dominates at 87% of the budget. The Phase 1 re-evaluation gate became the standard for moving from "we tried it" to "we shipped it."
Phase 18: The Trinary Intelligence Layer
The trinary package (state, sign, envelope, consensus) reached 100% test coverage with 90+ cases. The tissue service signed every IntentEnvelope with HMAC-SHA-256. The five consensus policies turned the tissue from a name into a real primitive.
Phase 19: Training, the real one
Five training recipes. Real LLM invocation. SSE progress events. Auto-publish to the ANN registry. Decision outcome tracking. Auto-retrain on degraded performance. The training service stopped being a sandbox and started being production-grade.
Phase 20-24: Staking, treasury, mainnet
AigarthPool (on-chain settlement with a 30/60/10 default split, QPI contract plus a TypeScript simulator). The multi-sig treasury. The pre-mainnet gates. The hardware presale. Each phase moved the platform from "a thing you can run" to "a thing you can fund."
The 2026-08-12 checkpoint
We froze the state on August 12 and wrote a PEP for the next chapter. The PEP is called The Aigarth evolution: it adds the Organism primitive and the Work Runtime, two pieces that turn the platform from a model registry into a place where intelligences actually grow.
You can read the full PEP in our docs. The next post in this series covers what the evolution actually shipped, three weeks later.