Business
Roadmap
Current V1 state and the seven-phase plan from pre-registration hardening to additional sign languages.
Current state
V1 is a complete mechanism definition and implementation baseline. The following are specified and built:
- artifact-based miner submission
- immutable weight and loader references, executed in network-published runtime profiles
- manifest validation
- offline validator execution
- deterministic inference
- canonical 543-landmark pose output
- 2B parameter and runtime limits
- private-corpus selection rules
- translation metrics
- grounding derangement and ratio
- canary generalization
- pose scoring
- rejection codes
- emissions construction
- validator abstention semantics
- reference-miner workflow
- adversarial and reproducibility test structure
PRE-REGISTRATION
The subnet is not yet registered on mainnet. The current stage is Phase 0.
Phases
| Phase | Window | Focus |
|---|---|---|
| 0 · Pre-registration hardening | Weeks 1–8 | Acquire and QA the 20-hour corpus plus 3-hour canary; run reference and alternative branches; calibrate normalization and the G_canary = 1.25 floor; verify replay; adversarial tests; publish final docs, the runtime profile image, and the reference model repository |
| 1 · Mainnet launch | Weeks 9–16 | Register the subnet; open submissions; run daily rounds; reach ≥20 gate-passing miners; produce a champion that significantly exceeds the reference on private data; publish per-round records |
| 2 · Real-world generalization | Months 5–8 | Expand signer and condition diversity; monthly 4-hour rotation; track slices separately; identify the first production-quality vertical; begin design-partner discussions |
| 3 · Commercial launch | Months 9–14 | Deploy the champion behind the API; monitoring, batching, rollback, reporting; onboard design partners; publish benchmarks on retired footage; establish external revenue |
| 4 · Vertical specialization | Months 12–18 | Maintain a general champion plus vertical profiles; increase broadcast and conversational coverage; add contextual-window evaluation; customer-specific post-processing |
| 5 · Text-to-sign generation | Months 15+ | Build paired text↔pose assets; train text-to-pose models; evaluate generation separately; integrate avatar rendering |
| 6 · Additional sign languages | Post-ASL | Expand language by language, BSL first, each with its own training sources and private evaluation corpus |
How the phases connect
Phase 5 is why pose is scored at all today. The 543-landmark output collected from every round accumulates the paired data foundation for text-to-sign, long before that product line begins.
Phase 6 repeats the whole mechanism per language. Each new sign language needs its own training sources and its own private evaluation corpus. The corpus is the expensive part, and it does not transfer.