TrueOpen
Development Roadmap
The inference network roadmap covers inference services, GPU nodes, Logprob verification, data privacy, and scaling. A separate four-stage roadmap covers cryptographic verification with SLP.
Inference Network
Inference Network Roadmap
Phase 0
Core Network
- Open GPU node registration for inference and verification tasks.
- Price and settle task orders in USDC.
- Run the network with permissioned Validators and allowlisted Builders.
Phase 1
Production Reliability
- Strengthen verification-material requirements, commitment ordering, and exception handling.
- Improve streaming responses, reconnection, refunds, and failure recovery.
- Bring in independent operators and establish performance and cost baselines.
Phase 2
Data Availability & Privacy
- Build Builder-operated data availability (DA), with retrieval from multiple sources, data repair, and handovers.
- Offer optional task encryption and stage-specific access to verification materials.
- Define storage responsibilities, retention periods, and service fees.
Phase 3
Economic Mechanisms & Open Consensus
- Upgrade staking mechanisms under approved plans.
- Gradually open Validator admission, delegation, and governance.
- Strengthen incentive, penalty, and exit mechanisms, along with asset-bridge security.
Phase 4
Open Builders & Batch Submission
- Open Builder candidate admission with staking, rotation, and handover of outstanding responsibilities.
- Use DA to support batch submission and processing of records from each stage of a task.
- Preserve per-task validation, reconciliation of funds, resubmission of omitted records, and failure recovery.
Phase 5
Single-Chain Scaling
- Optimize single-chain execution, state reads and writes, and data indexing.
- Improve DA transmission, storage, and recovery efficiency.
- Tune batch sizes and maximum waiting times to reduce latency and per-task costs.
Cryptographic Verification
Cryptographic Verification Roadmap (SLP)
SLP (Sampled Layerwise Proofs) verifies LLM inference using sumcheck / GKR and lookup arguments. The proof of concept is complete. The next stages cover security and a testnet prototype, scale and performance, and production deployment.
SLP Stage 0 · Complete
Proof of Concept
- Completed end-to-end experiments on GPT-2, TinyLlama-1.1B, and Llama-2-70B, demonstrating the feasibility of sampled layerwise proofs on real models.
- Implemented commit-then-sample verification: commit layer-group boundaries and inputs/outputs before selecting groups to prove. Input and output groups are always proved. Verifiers derive samples independently, and verification can be escalated to a full proof.
- Implemented streaming weight commitments, enabling proofs for a 70B model on a CPU host with 2 TB of memory. Implemented proof packing so multiple requests can share a proof.
- Evaluated quantized-model fidelity, improved residual-stream bit widths, and identified model versions using weight and scale digests.
- Tested grinding attacks against pure Fiat–Shamir sampling and designed a commit → randomness beacon → prove flow. All seven adversarial negative cases were rejected. Integration with a live beacon is planned for the next stage.
Measured Results
- GPT-2 · Laptop CPU:Committing 27 boundaries and proving 5 of 27 layer groups took 26 seconds; verification took 0.16 seconds.
- TinyLlama-1.1B · RTX 4090 Inference, CPU Proving:Packing 12 requests into one proof reduced cost by a factor of 6.5 compared with 12 separate proofs.
- Llama-2-70B · 256 Threads, 2 TB CPU Host:Committing 163 boundaries and proving five layer groups took 21 minutes. The proof was 4.34 MiB, and verification took 46 seconds without model weights. These are sampled-proof measurements, not full-proof or GPU-proving results.
SLP Stage 1 · Next
Security Hardening and Testnet MVP
- Complete the testnet workflow for 1–3B models: registration, inference, commitments, beacon retrieval, proof generation, verification, and settlement.
- Integrate a public-ceremony reference string (BN254 powers-of-tau) and a live randomness beacon. Fix one Manifest per request and publish it before obtaining beacon randomness to prevent repeated sampling for favorable outcomes.
- Commit decoding randomness and bind output tokens to output anchors, enabling verification of inference with temperature sampling.
- Develop collateral, slashing, and repeated-audit mechanisms so expected losses from cheating exceed its rewards. A single sample's detection rate depends on coverage; collateral and repeated audits strengthen the deterrent.
- Prioritize tolerance-based spot checks for fp16 services, routinely checking each layer's matrix multiplication. In parallel, develop a runtime that serves integer-quantized models directly. Specify the verification scope of each approach.
- Publish protocol specification v1, a model-registry prototype, and Verifier CLI / SDK tools. Harden integer arithmetic and extend tampering, replay, and beacon-attack tests.
Development Targets:For a 1.1B model with context length up to 512: proof generation within 60 seconds and verification within 2 seconds per request, with at least 25 adversarial negative cases correctly rejected. These are targets, not measured results.
SLP Stage 2 · Planned
Scale and Performance
- Separate weight-storage precision from accumulation precision, using 8/16-bit weight storage and 64-bit accumulation. Develop an int8 Tensor Core path to reduce GPU memory use.
- Move the integer runtime and quantized lookups to GPUs, and accelerate sumcheck, multiscalar multiplication, and other proving computations.
- Measure attention-proof costs at context lengths of 512, 1k, 4k, and 8k. Optimize attention arguments and segmented proofs, and publish cost curves across models and context lengths.
- Develop deterministic inference so repeated runs with the same model and input produce identical output tokens.
- Integrate packed proofs into the service. Improve trace retention and chunked full proofs for dispute review, and obtain an independent cryptographic audit of the protocol and implementation.
Development Targets:For 7B, run inference, commitments, and sampled proofs on one 24 GB GPU, targeting proofs within 5 minutes per request at 2k context. For 70B, run the same workflow on one or two 80 GB GPUs and measure costs at 2k context. Target under 5 minutes for a five-group 70B proof and 100% inference reproducibility. These are development targets, not achieved performance.
SLP Stage 3 · Planned
Production Deployment and Ecosystem
- Integrate with the mainnet for verification-result submission and dispute handling, including escalation to full proofs for disputed tasks. Research recursive proof wrapping to shorten on-chain verification.
- Expand Dense model families and add MoE support, starting with Mixtral and then DeepSeek. Develop router anchors, permutation arguments, and expert-selection rules.
- Advance the fp16 service path with routine tolerance-based spot checks and sampled or full proofs for disputes. Publish its economic-security parameters and scope.
- Complete key management, proof and trace storage, monitoring, alerts, service-level agreements (SLAs), and cost reporting.
- Open-source protocol specification v2, reference implementations, and test vectors. Run bug bounties and ongoing adversarial testing.
Launch Targets:Stable mainnet operation, support for at least three model families, and full-proof handling of real disputes.