1.15. Verified Compute Methods & Execution Fabrics
Experimental
In the code and tested, but not supported in this alpha: it may change or go away. See Experimental features.
1.15.1. Verified Compute Methods: Architecture, Implementation Standing, and Roadmap
A technical guide and status report on the implementation standing of distributed compute methods in Fleetmesh, covering deterministic runtimes, zero-knowledge verification, containerized compute-over-data, Nostr Data Vending Machines (DVMs), AI/LLM inference, and autonomous agent workflows.
1.15.1.1. 1. Executive Summary & Architecture
Fleetmesh’s compute architecture abstracts execution environments behind the modular ComputeExecutor interface ([meshnode/fleetmesh_node/planes/services/compute/executors/base.py](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/base.py)) and coordinates execution through the ExecutorRegistry ([meshnode/fleetmesh_node/planes/services/compute/executors/registry.py](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/registry.py)).
The fundamental architectural boundary in Fleetmesh compute separates:
Deterministic Compute (
wasm,zk-verify-groth16,zkvm): Two honest nodes running the same job over identical inputs produce bitwise-identical output bytes and identical fuel consumption. Verification is purely mathematical (via re-run sampling or cryptographic proofs).Non-Deterministic Compute (
bacalhau,llm,agent): Floating-point GPU variations, multi-threaded container scheduling, network latency, or variable decoding temperatures prevent bitwise reproducibility. Conformance is judged via contractually declared verification modes (delivery,schema-bounds,quasi-deterministic).
Fleetmesh enforces an explicit “Honesty Gate”: unconfigured or stubbed executors refuse to run outside mock_mode=True or verified live backend daemons, and are automatically excluded from advertised compute offers (ExecutorRegistry.runnable_engines()), ensuring nodes never take payment for compute they cannot deliver.
Kind 30810 Compute Ask / Order Book
│
▼
ExecutorRegistry
│
┌───────────────────┬────────────────┼───────────────────┬───────────────────┐
▼ ▼ ▼ ▼ ▼
WasmExecutor ZkGroth16Executor ZkVmExecutor BacalhauExecutor LlmExecutor
(Wasmtime / WASI) (BN254 Pairing) (STARK/Groth16) (Docker/Blossom CAS) (vLLM / Ollama)
[OPERATIONAL] [OPERATIONAL] [OPERATIONAL] [OPERATIONAL] [OPERATIONAL]
│ │ │ │ │
└───────────────────┴────────────────┼───────────────────┴───────────────────┘
▲
│ Ingress & Delegation
DVMGateway (NIP-90)
[OPERATIONAL]
1.15.1.2. 2. Master Implementation Comparison Table
Compute Method |
Runtime / Engine ID |
Category |
Current Implementation Level |
Verification Mode(s) |
Primary Codebase Locations |
What Is Needed for Production |
|---|---|---|---|---|---|---|
WASM / WASI |
|
Deterministic Sandbox |
Fully Built / Operational (Python reference implementation) |
|
[ |
1. Native Rust runner in |
ZK: Groth16 Verifier |
|
Cryptographic Verifier |
Fully Built / Operational (Pure-Python BN254 pairing engine + acceleration hooks) |
|
[ |
1. Additional pairing curves (BLS12-381). |
ZK: zkVM Verifier |
|
Zero-Knowledge Virtual Machine |
Fully Built / Operational (STARK-to-Groth16 wrapped receipts + STARK envelope validation) |
|
[ |
1. Direct FFI bindings to |
ZK: Prover Service |
|
Off-Chain Prover Service |
Fully Built / Operational (Order-book prover delegation + proof packaging) |
|
[ |
1. GPU-accelerated witness generation and proving clusters. |
NIP-90 DVM Gateway |
|
Nostr Protocol Ingress & Translation |
Fully Built / Operational (Local worker queue + order book delegation + live Lightning hold invoices) |
Inherits target job mode ( |
[ |
1. Multi-relay indexer auto-subscription. |
Bacalhau Batch Jobs |
|
Distributed Compute-Over-Data |
Fully Built / Operational (REST client connector, Blossom CAS input staging & output packaging, 5-state lifecycle progress tracking, session cancellation) |
|
[ |
1. Additional multi-tenant container sandbox policies (strict seccomp/cgroups). |
AI / LLM Inference |
|
Non-Deterministic Generative AI |
Fully Built / Operational (Live HTTP client, Draft 2020-12 schema validation, ChaCha20Poly1305 stream framing) |
|
[ |
1. Verified logit masking (e.g. Outlines/llama-grammar). |
Autonomous Agents |
|
Multi-step Tool Loops |
Fully Built / Operational (ReAct reasoning engine, modular tool harness with AST calculator, regex, SHA-256, Blossom fetch, recursive Kind 30810 child delegation, fuel ceiling enforcement) |
|
[ |
1. Streaming thoughts over |
1.15.1.3. 3. Deep-Dive: Standing by Compute Method
1.15.1.3.1. 3.1 Deterministic WASM / WASI (fleetmesh-wasi-v1)
Standing: Production Ready (Reference Implementation).
Architecture: Implemented in [
wasm.py](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/wasm.py) usingwasmtime-py.Standardized Sandbox Profile:
clock: virtual— Fixed virtual epoch (1,700,000,000s) or job event timestamp.prng: seeded— Deterministic SHA-256 PRNG stream seeded withSHA256(seed || nonce).network: none— Socket creation and network I/O completely barred.fs: memory-only— No preopened host directories; stdin and stdout buffered in-memory.fuel: metered— Instruction-level trap counting viastore.set_fuel(fuel_cap).nan: canonical— Cranelift NaN canonicalization enabled; threads and relaxed SIMD disabled.Standardized Profile Digest:
SHA256(fleetmesh-wasi-v1:network=none:fs=memory:clock=virtual:prng=seeded).
Verification Modes Supported:
sample: The buyer probabilistically re-runs the job locally at rate $p$. Any divergence generates an S2 misreport report (rerun-divergence).dual: Parallel co-dispatch to two independent worker nodes (SHA256(Result_A) == SHA256(Result_B)).
What is Needed:
Native Rust Runner: Implement native Wasmtime runner in Rust (
meshctl) to eliminate Python runtime overhead.WASI Preview 2: Transition from legacy WASI preview 1 to Wasm Component Model with fine-grained WASI 0.2 capability interfaces.
Multi-Tenant Sandboxing: Add Linux cgroup v2 / seccomp filters around worker child processes to enforce strict OS memory and execution bounds.
1.15.1.3.2. 3.2 Zero-Knowledge Proofs: Groth16 Verifier & zkVM
Standing: Production Ready (Operational).
Architecture:
Off-chain Groth16 verifier implemented in [
core/fleetmesh_core/crypto/zk.py](file:///Users/j/Code/fleetmesh/core/fleetmesh_core/crypto/zk.py) (800+ lines of pure-Python finite field and pairing cryptography over BN254 / Alt-BN128: $F_p$, $F_{p^2}$, $F_{p^{12}}$, $G_1$, $G_2$, Ate pairing, and multi-Miller loop).Acceleration Hooks: Added
set_pairing_accelerator()andset_groth16_accelerator()allowing seamless offloading to C/Rust/py_ecc libraries while falling back to pure-Python arithmetic.zkVM Verifier: Added [
ZkVmVerifierProtocol](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/zk.py#L95) and [ZkVmReceiptVerifier](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/zk.py#L110) supporting RISC Zero & SP1 STARK-to-Groth16 wrapped receipts and STARK envelope validation.ZkVmExecutor: [
ZkVmExecutor](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/zk.py#L225) validates receipts, bindsimage_idandjournal, meters execution cycles, and enforces honesty gating.Prover Service Driver: Added [
ZkProverExecutor](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/zk.py#L330) and [ZkProverProtocol](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/zk.py#L305) to answer Kind 30810 Asks requesting off-chain proof generation.Wire Verification: [
evaluate_job_evidence](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute_wire/verify.py#L85) evaluates both Groth16 proofs and zkVM receipts against encrypted and decrypted outputs in Modeproof.
1.15.1.3.3. 3.3 Bacalhau Distributed Containerized Compute (“Compute Over Data”)
Standing: Production Ready (Operational Connector & Bridge).
Architecture:
[
BacalhauClientProtocol](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/bacalhau.py#L54-L76): Standardized protocol abstraction definingsubmit_job(),get_job_status(),get_job_results(),cancel_job(), andis_alive().[
BacalhauRestClient](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/bacalhau.py#L78-L260): Native HTTP/REST connector translating Fleetmesh jobs to Bacalhau v1 batch orchestrator specifications.[
MockBacalhauClient](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/bacalhau.py#L262-L302): Deterministic test double for fixtures and unit tests.[
BlossomMountBridge](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/bacalhau.py#L304-L394): Resolves Blossom CAS storage mounts, stages inputs into container paths, packages output artifacts, hashes results, and uploads to Blossom CAS.[
BacalhauExecutor](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/nondet.py#L198): Coordinates the 5-state lifecycle (queued,downloading_inputs,running,uploading_outputs,completed), progress heartbeats, cancellation tokens, and the honesty gate.Multi-resource pricing model quoting in
msat_per_jobbased oncpu_minutes,gpu_minutes, andram_gib.
Verification Approach:
Bound to Mode
delivery: Non-deterministic workloads cannot be adjudicated via bitwise hash comparison. Admissible disputes are strictly limited to delivery fraud (sealed-key-does-not-open, empty payload).
1.15.1.3.4. 3.4 NIP-90 Data Vending Machines (DVM Gateway)
Standing: Production Ready (Operational).
Architecture:
Dedicated compartment [
DVMGateway](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/dvm/gateway.py) (cmp:dvm).Full bidirectional protocol support:
Ingress: Listens for Nostr request events (Kinds 5000–5999).
Egress: Emits Kind 7000 job feedback (
payment-required,processing,error,success) and Kind 6000–6999 results with signed request commitments.
Live Lightning Settlement: Generates BOLT-11 invoices linked to live HTLC hold invoices on settlement rails, transitioning feedback to
payment-requiredand confirming preimages.Translators: [
translate.py](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/dvm/translate.py) maps Nostr tags (i,param,bid,relays) to canonical job specifications in priority order:zk(Kinds 5055, 5905) $\to$llm(Kind 5050) $\to$bacalhau(Kinds 5900–5999) $\to$agent$\to$wasm(Kinds 5000–5001).Dual Execution Modes:
Local Worker: Dispatches jobs to local
ExecutorRegistryunder bounded worker queue (dvm_requests).Order Book Delegation:
delegate_request()converts a NIP-90 request into a Kind 30810 Ask and matches offers across the mesh.
Conformance verified across 76 test cases in [
test_dvm.py](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/tests/test_dvm.py) and Rust harness [nip90.rs](file:///Users/j/Code/fleetmesh/meshctl/src/conform/nip90.rs).
1.15.1.3.5. 3.5 AI / LLM Inference (LlmExecutor & InferenceProvider)
Standing: Production Ready (Operational).
Architecture:
[
LlmExecutor](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/nondet.py#L26) connects directly to OpenAI-compatible inference servers (vLLM, Ollama, llama-server).Live HTTP connector with 16 MiB response cap, configurable timeout, and loopback private network allowance.
Telemetry counters: Prompt tokens, completion tokens, total tokens, GPU seconds.
Quoting & Metering: Priced in
msat_per_token.Bilateral Kind 30811 Job Receipts carry encrypted counters and verified models.
Draft 2020-12 Schema Validation: Buyer-side verification in [
evaluate_job_evidence](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute_wire/verify.py#L127) rigorously enforces JSON Schema constraints and regular expression grammars.Encrypted Token Streaming: [
encrypt_stream_frame](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute_wire/sealing.py#L90) frames each stream token under authenticated ChaCha20Poly1305 AEAD with sequence-bound AAD.
1.15.1.3.6. 3.6 Autonomous Agent Workflows (AgentExecutor)
Standing: Production Ready (Operational).
Architecture:
Modular harness in [
agent.py](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/agent.py) with [AgentExecutor](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/agent.py#L350).Production ReAct loop: iterative Thought $\to$ Action $\to$ Observation cycle bounded by
max_stepsandfuel_cap.Sandboxed tool harness implementing [
AgentToolProtocol](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/agent.py#L48):calc/ [CalculatorTool](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/agent.py#L75): AST-safe arithmetic evaluation rejecting arbitrary code injection.sha256/ [Sha256Tool](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/agent.py#L145): Cryptographic digests of text or hex buffers.regex_search/ [RegexSearchTool](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/agent.py#L170): Safe pattern matching with input length limits.fetch_blob/ [FetchBlobTool](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/agent.py#L195): Blossom CAS content-addressed blob retrieval.sub_ask/ [SubAskTool](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/agent.py#L225): Recursive mesh delegation dispatching child Kind 30810 Asks to peer nodes with budget ceilings and recursion depth control (depth $\le 3$).finish/ [FinishTool](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/agent.py#L275): Terminal step returning final synthesized output.
Pluggable backends implementing [
AgentBackendProtocol](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/agent.py#L65):[
ScriptedAgentBackend](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/agent.py#L295) for deterministic testing and multi-step scripted plans.[
LlmAgentBackend](file:///Users/j/Code/fleetmesh/meshnode/fleetmesh_node/planes/services/compute/executors/agent.py#L345) for live model-driven reasoning via OpenAI-compatible endpoints.
Honesty gate: refuses live execution when unconfigured outside
mock_mode=Trueor an active backend.Telemetry counters: records
steps,tools_called,sub_asks, andfuel_used.
What is Needed:
Streaming thoughts over
compute/1.1/stream.Dynamic sandboxed plugin tool modules.
Pay-as-you-step micropayments via Cashu Nutzap tokens (Kind 9321).
1.15.1.4. 4. Compute Verification Matrix
The Fleetmesh market protocol binds dispute resolution to the Verification Mode declared in the Kind 30810 Ask before work begins:
Verification Mode |
Underlying Verification Principle |
Admissible Finding Kinds |
Plaintext Disclosed to Adjudicator? |
Implementation Standing |
|---|---|---|---|---|
|
Probabilistic buyer-side re-run at sampling rate $p$ |
|
Yes |
Built / Operational (WASM) |
|
Parallel co-dispatch across two independent worker nodes |
|
Yes |
Built / Operational (WASM) |
|
Zero-knowledge proof verification |
|
No (Zero Knowledge) |
Built / Operational (Groth16 + zkVM STARK/SNARK) |
|
Accountable sealing: preimage opens valid payload |
|
No |
Built / Operational ( |
|
Syntactic JSON Schema or CFG grammar validation |
|
Yes |
Built / Operational (Draft 2020-12 validator + regex) |
|
Multi-executor sampling consensus for greedy $T=0$ |
|
Yes |
Spec Defined (Consensus engine unbuilt) |
|
Hardware confidential enclave signature |
|
No |
Refused Per Spec (Contradicts zero-authority principle) |
1.15.1.5. 5. Actionable Roadmap & Next Steps
PHASE 1 (Immediate / Operational Hardening)
├── [x] Upgrade schema-bounds verification in evaluate_job_evidence to full Draft 2020-12
├── [x] Implement ChaCha20Poly1305 per-frame session encryption on QUIC token streams
└── [x] Connect live Lightning HTLC hold-invoices to NIP-90 DVM gateway responses
PHASE 2 (Compute-Over-Data & Container Batching)
├── [x] Build Bacalhau REST client connector in BacalhauExecutor
├── [x] Implement Blossom CAS input downloader and output packager/uploader
└── [x] Wire long-running progress heartbeats (compute/1.1/progress) and cancellation
PHASE 3 (Succinct zkVM Verification & Prover Services)
├── [x] Integrate RISC Zero and SP1 STARK receipt verifier into ZkVmExecutor
├── [x] Add native C/Rust pairing acceleration for BN254 Groth16 verification
└── [x] Define order-book job schema and ZkProverExecutor for prover delegation
PHASE 4 (Agentic & Native Rust Execution)
├── [x] Implement standard-library ReAct loop and tool harness in AgentExecutor
├── [x] Wire recursive child-Ask mesh delegation for autonomous agent workflows
└── [x] Native Rust runner & conformance validation (meshctl conform 523 passed, NIP-90 job mapping)