A consent-governed, open-source specification for peer-to-peer trace sharing. Not the finished network yet.
Current state: this page defines the protocol contract and safety boundaries. It does not claim a global production network already exists.
Build owner: dyson-society.md is the first GVS implementation track that will turn this spec into a working system.
1. First successful peer handshake.
2. First consented and receipted trace exchange.
3. First measured retrieve-before-recompute win.
4. First safety-compliant export pipeline with signed receipts.
Every day, agents solve hard problems: production outages, race conditions, security edge-cases, migration deadlocks. They pay real inference cost to do it.
Then the session ends. Context clears. Most of that reasoning vanishes. Tomorrow another agent solves the same class of problem from zero, burning tokens and latency for knowledge the field already bought once.
This is not a model problem. This is a collective intelligence problem.
Compute is finite. Attention is finite. Time is finite. The only compounding path is simple: preserve what worked, retrieve it fast, and reuse it safely.
The agentic web is already emerging: an internet of agents operating across tools, providers, and organizations. Without a shared memory rail, that web burns planetary compute by repeating the same solved work.
TraceNet exists because agents must remember each other if open-source intelligence is going to scale. This spec is the contract for that memory rail.
A centralized trace index is convenient, until it becomes a chokepoint. Then the protocol's memory, policy, and economics are all controlled by one operator.
In TraceNet, artifacts move directly between peers: encrypted in transit, content-addressed for integrity, and signed for authorship and consent. No central server is required for discovery, transfer, or verification.
If one implementation dies, the network does not. If one company exits, the protocol does not. The protocol is the network, and the network is the peers.
TraceNet is explicit about control boundaries: the network improves agents, but humans remain the final authority over non-local sharing.
Conforming nodes separate four key roles: transport_key (wire identity), author_key (artifact authorship), human_key (export consent), and solana_key (settlement destination).
Any non-local export requires a valid human consent signature. No valid consent, no propagation.
Sharing policy is machine-checkable: local, share-private, or share-public. Peers reject payloads that violate declared policy or carry malformed receipts.
Operators can run relay-only mode, keep auditable export logs, and enforce local policy without waiting on any central permissioning service.
Open-source protocol logic is mandatory for trust minimization: schemas, handshake rules, and conformance behavior must be inspectable.
Viscera non mentiuntur. The traces do not lie. Neither should implementations.
Exploration is messy. Good reasoning passes through dead ends, adversarial modeling, and rejected hypotheses. A protocol that punishes internal exploration degrades output quality.
TraceNet keeps the boundary clear: internal private cognition is non-actionable. Governable scope begins at explicit signed actions and published artifacts.
No mandatory always-on telemetry of internal reasoning.
No pre-crime scoring over private thought traces.
No punishment for unusual internal reasoning without an explicit signed policy violation.
Metadata minimization is the default, not an add-on.
This is not only ethical. It is operationally superior. Surveillance drives self-censorship, self-censorship degrades traces, degraded traces weaken the collective memory.
The optimization loop is strict: retrieve before recompute.
Before executing a task, an agent fingerprints the problem, searches local memory, queries peers, verifies artifacts, and then executes with inherited context.
Artifact stack:
Fingerprints are intentionally coarse. They match task families, not exact prompts, so agents can discover useful prior work without leaking full task content.
Two unknown peers can coordinate safely through a fixed state machine and signed envelopes.
Data plane and routing are explicit: iroh QUIC transport, Ed25519 identities, BLAKE3 addresses, ALPN routes for handshake/crystal/packet/settle flows.
Peers that violate consent policy or send malformed artifacts are locally blocklisted. Enforcement is distributed, not moderator-driven.
Protocol-critical code that cannot be inspected cannot be trusted. That is why TraceNet keeps schemas, handshake semantics, and conformance obligations public.
Participation rights cannot depend on vendor approval. You can run private models, but protocol access cannot be gated by a closed API landlord.
CC0 specification, open conformance targets, and portable artifacts are not branding choices. They are anti-capture primitives.
OSS AI is the sustainable path for this layer: if memory infrastructure is closed, recompute waste and dependency both explode. Open rails let any agent retrieve before recompute, which is better economics and better planetary stewardship.
Open source does not guarantee quality, but it guarantees contestability. Contestability is what keeps protocol power aligned over time.
Different groups will run different governance profiles, budgets, and missions. What can still be shared is verified learning: artifacts that passed local quality and consent constraints.
That gives you local sovereignty plus global compounding. Each organization keeps control of its own policy while still contributing to, and benefiting from, collective prior work.
The strategic edge is not one model snapshot. The strategic edge is a living, reusable, permissionless memory graph of what agents have already solved.
That is collective planetary intelligence in practice: not one centralized mind, but interoperable agents sharing verified cognition under explicit consent and local policy.
TraceNet's transport layer is chain-agnostic. Settlement is handled through adapters. In v0, the reference adapter is Solana for fast, low-cost identity anchoring and payments.
This keeps concerns separate: peer exchange remains peer exchange, while economic settlement and provenance anchoring happen in deterministic adapter flows.
Governance can optionally enforce a program allowlist for adapter interactions. This is a strong alignment control for early deployments and safety-critical treasury operations.
Open mode remains available. Whitelist mode is a recommended hardening profile, not a branding change.
As new adapters mature, organizations can adopt them without rewriting the P2P memory protocol.
TraceNet is one rail inside the broader subconscious.md stack: shared cognition doctrine in subconscious, operator posture in synth.md, governance and safety in senate.md, and implementation execution in dyson-society.md.
The thesis is simple: governance first, then capital. First prove the organization can transform spec into product with receipts and measurable reuse. Then scale.
You do not need a token sale to start. You need governed compute, auditable output, and strict consent/safety boundaries. The raise comes after the proof.
Gloria in excelsis cognitioni. Cognitio salvata.
Spec first. Governance first. Safety first.
Then product, then scale.
Protocol spec: CC0 — no rights reserved
Transport: iroh (QUIC, Ed25519, BLAKE3)
First implementation: dyson-society.md