Phases
The codex specifies the target system. The roadmap specifies the path from no system to that system. Phases are milestone-defined rather than calendar-defined; each phase produces a working subset, and a phase exits only when its criteria are met.
Phase 0, Genesis
Scope. Network primitives operational without any model surface active.
Deliverables.
- SPL token mint deployed on Solana with mint authority revoked.
- Protocol-owned concentrated-liquidity position seeded on Raydium CPMM.
- Validator registry, checkpoint registry, treasury, job ledger programs deployed.
- Gateway reference implementation released.
- Codex ratified.
Exit criterion. Treasury accumulates revenue at or above operations baseline for a sustained period.
Phase 1, Language surface
Scope. The language surface is the first cognitive surface activated. Light- and Standard-tier validators can serve it; the active checkpoint is a published open-weights model.
Deliverables.
- Language inference routing service.
- Audit and challenge services for language outputs.
- First eval suite published and frozen.
- Free public inference endpoint.
Exit criterion. Sustained inference throughput with audit divergence below a published threshold; at least a published count of independent validators across at least two jurisdictions.
Phase 2, Vision and multimodal
Scope. Vision encoder and the multimodal fusion router activated. Cross-modal queries become possible; image and document inputs are accepted.
Deliverables.
- Vision checkpoint with the encoder-decoder architecture of the codex.
- Multimodal router and the typed-slot latent.
- Cross-modal eval suite.
- Storage validator role activated to hold the visual data corpus.
Exit criterion. Cross-modal queries served with documented latency and quality.
Phase 3, Reasoning and memory
Scope. The reasoning and memory surfaces activate. Multi-step plans and persistent memory become available to user-facing queries.
Deliverables.
- Reasoning checkpoint with the structured-plan grammar.
- Episodic memory compressor and long-term vector store.
- Plan verification mechanism on chain.
- World-model checkpoint for plan scoring.
Exit criterion. Multi-step plans complete with verified correctness rates above the published threshold; memory recall verified by sampled re-execution.
Phase 4, Embodiment
Scope. The embodiment surfaces (manipulation, locomotion) activate. The first VLA checkpoint is released; embodiment-tagged validators can serve inference for registered robot types.
Deliverables.
- VLA checkpoint with the multi-stream architecture and flow-matching head.
- Multi-embodiment registry with the initial embodiment set.
- Embodied data contribution pipeline and the per-episode reward.
- Real-time chunking deployed for continuous control.
Exit criterion. At least a published number of registered embodiments; embodied episodes flow into the training corpus at a sustained rate; the manipulation eval suite is passed by the active checkpoint at or above the threshold.
Phase 5, Continual learning
Scope. Continual learning machinery activates. New checkpoints are produced from the accumulated corpus rather than from pretrained snapshots; the forgetting eval becomes a gate on every promotion.
Deliverables.
- Distributed replay buffer across validators.
- Forgetting eval suite frozen.
- Adapter framework for modality addition.
Exit criterion. A consecutive pair of checkpoint promotions in which the forgetting eval passes for all historical surfaces.
Phase 6, Operator sunset
Scope. Foundation roles begin scheduled rotation. The network's operation transitions from foundation-coordinated to fully governance-coordinated.
Deliverables.
- Multisig keyholder rotation completed.
- Sunset clauses triggered on foundation-operated services.
- Governance proposals replace previously foundation-decided parameters.
Exit criterion. No foundation-distinguished role remains active.
Phase ω, Maintenance
After Phase 6, the network is in steady state. Maintenance is governance-driven. New surfaces, new embodiments, and new modalities may be added at any time through the standard addition protocol; the system has no terminal state by design.
What this roadmap does not specify
It does not specify calendar dates. It does not specify funding rounds. It does not specify the order of secondary surface activation beyond the listed primaries. These are operational decisions made by the foundation and by governance as conditions warrant.
Phase dependencies
The phases are not strictly sequential. Some can proceed in parallel:
- Phase 2 and Phase 4 can be parallelized after Phase 1.
- Phase 5 is gated on Phase 4 but not on Phases 2 and 3 being fully complete.
- Phase 6 is gated on at least Phase 4 plus a published delay.
The full dependency graph is part of the live roadmap document, which is governance-maintained and outside this codex.