Signals decay.
A hypothesis is a signal with an intensity. Left alone it fades, window by window, until the field forgets it.
Read the signalConsensus,
without a center.
Agents don't vote.
Evidence reinforces. Doubt decays.
SMESH is a Rust protocol for coordinating agents through signals that diffuse, decay and get reinforced. Here it runs a read-only quant study.
Explore the repositoryRecorded run / read-only market dataSMESH / MESH STUDY 001
Market data moves from Kalshi, Massive and Kraken through scout agents to two analysts. A walker tests each hypothesis on unseen windows and a skeptic doubts weak evidence. Each hypothesis is a signal: confirmed windows brighten it, contradictions and doubt dim it, and it decays without evidence. Motion is illustrative; recorded counts appear below.
–
–
––––––
Recorded run, replayed. Counts preserved; motion interpreted. Read the evidence
No coordinator. No vote. A field.Independent open source / Rust
Signals, not ballots.
Evidence, not authority.
Every claim lives on a shared field.
Nothing keeps it alive but evidence.
Evidence reinforces.
Each walk-forward window tests the claim on data it has never seen. Agreement restores it; disagreement and a skeptic's doubt cut its confidence.
Read the fieldJev routes.
Agents describe themselves in plain language. Each asks Jev whether it suits the task, and settlement picks the winner. No fixed task types; any domain.
Meet JevA2A connects.
The whole mesh is one A2A v1.0 agent. Any client can send it a task, and other A2A agents can join it as members.
Read A2AEvery claim on the field.
Every window on the record.
Scouts read Kalshi, Massive and Kraken with a client that can only issue GETs to market-data paths. Analysts posed eight hypotheses about Kalshi's prediction markets. A walker tested each on unseen windows; a skeptic doubted small samples and fee-sensitive results.
A claim's strength is its recency times its evidence: the margin of confirming over contradicting windows, shrunk toward zero when there are few. Claims no recent window supports fade.
–
Two limits worth knowing. The calibration test asks whether Kalshi's error is within what chance alone produces; at ~90 predictions a window it can only fail to reject calibration. And Kalshi's payroll ladders missed the prints widely in both directions: check which figure they settle on before leaning on payroll results.
Download the replay dataHypotheses at the end of the run–
| Claim | Windows for / against | Strength |
|---|
Consensus is a strength of 0.30 or more.
–
What the walk found.
Drawn from the record.
Out-of-sample only.
Every point is a recorded prediction.
Kalshi oddsVolatility model
CPIPayrollsFed
Ask the mesh.
Over A2A, routed by Jev.
The finished study joins the mesh as an agent.
It answers only from what it recorded.
Serve the mesh.
The study becomes an agent named Quant beside any others, behind one A2A v1.0 endpoint.
Jev routes the question.
Each agent asks Jev whether it suits the task. A question about Kalshi's odds reaches Quant; a translation does not.
Jev picks the evidence.
Quant asks Jev which recorded findings bear on the question, then returns them verbatim. It cannot claim more than the walk showed.
$ smesh a2a serve --jev --agents 0 --quant-run runs/quant-full \ --agent "Translator=a translator between English and Portuguese" $ smesh a2a send "Are Kalshi's hourly bitcoin odds well calibrated?" [Working] round 1: Quant claims (72%) [Working] settled on Quant; executing From recorded research run `quant-full` (read-only data, walk-forward validated): - Kalshi's hourly crypto odds are calibrated (calibration error within what chance alone would produce) — SUPPORTED by consensus (strength 0.86; 8 windows confirmed, 0 contradicted). [Completed] completed by Quant $ smesh a2a send "Translate into Portuguese: the evidence decides." [Working] round 1: Translator claims (92%), Quant claims (72%) [Working] round 2: Quant backs off a evidência decide [Completed] completed by Translator
Recorded against the live gateway, abridged. Jev's claim strengths shown as recorded; settlement, not Jev, picks the winner.
One workspace.
The study is one use of it.
A Rust workspace and a single smesh binary.
Signals at the core; agents, transport and research on top.
# build, then watch signals settle with no LLM and no key $ git clone https://github.com/copyleftdev/smesh-rust $ cd smesh-rust && cargo build --release $ cargo run --release --bin smesh -- sim --nodes 50 --ticks 100 # install, then serve the mesh as one A2A v1.0 agent # (needs OPENROUTER_API_KEY and TYPESAFE_API_KEY) $ cargo install --path smesh-cli $ smesh a2a serve --jev \ --agent "Contracts=a reviewer of commercial contracts" # re-run this study (read-only; needs MASSIVE_API_KEY) $ smesh quant run --track both --out runs/quant-mine $ smesh quant export-site --run runs/quant-mine --out site/data
smesh-coreSignals, nodes, fields, network topologies, Bayesian trust.smesh-runtimeA live peer-to-peer mesh over QUIC, with signed attestation.smesh-agentLLM agents on OpenRouter or Claude; plain-language routing with Jev.smesh-a2aThe mesh as an A2A v1.0 agent, with signed Agent Cards.smesh-quantThe read-only research mesh replayed above.smesh-worldsmesh-refinerysmesh-ratifySigned organizational world models, with a human as merge authority.smesh-bountySecurity swarms with agentic tool execution, for authorized testing.
Let the evidence
decide.
Build with SMESH Open source. Rust. MIT / Apache-2.0.