07 โ Trust-graph formation (the dream-agent cycle) โ notes
What this runtime is. An agentprivacy-harness runtime that models how the DTG trust graph forms, run as
a dream-agent cycle (Mage proposes โฅ Swordsman proves, across the Gap). It is the bridge runtime: it takes
the personhood anchor from runtime 01 and shows the graph growing on top of it, with every object named in
both the agentprivacy framework and the DTG Credentials Core Spec. Reference model, not a circuit โ
like runtime 01 it demonstrates the relations a ZK trust-graph would prove, zero-dep and offline.
Run: node test.mjs โ 11/11 (see below). Reuses: ../01-uniqueness-nullifier (enrol + nullifier).
The spec: how the graph forms, one collision at a time
The spellweb grammar states it in five words โ collision โ edge โ propagation โ "the trust graph grows one
collision at a time." This runtime makes each arrow mechanical and names it on both sides:
| Step | Framework (agentprivacy) | Cred-spec (DTG) | In this runtime |
|---|---|---|---|
| Node | personhood-anchored member; three-layer identity's principal + data layers | member of a VTC with an M-DID; personhood via the PHC rule | joinCommunity() โ member = nullifier(secret, community); one node per human per community |
| Collision | an encounter producing a matching compression (proverb/spell match) | the out-of-band event a VRC attests | encounter() โ a shared value only the two parties can derive |
| Edge | a VRC (relationship layer): bilateral commitment, "proof of bilateral comprehension without central authority" | a VRC between two fresh R-DIDs; the pairwise construction | Mage.propose() โ candidate carrying each side's rdid() + consent |
| Consent gate | the shared graph grows only on mutual consent (SPELLWEB spec ยง6B) | holder MUST present unaltered; bilateral issuance | Swordsman rejects unilateral-no-mutual-consent |
| Propagation | trust reaches transitively; value lives in the Gap (max betweenness) | community-anchored ZKP: two members of one VTC prove relationship existence | TrustGraph.connected() (BFS reachability) |
The two cred-spec ZK constructions appear exactly here: pairwise (VRC-based) is the single edge; the
community-anchored (VMC-based) is propagation within a shared C-DID. See ../CRED-SPEC-COHERENCE.md.
The dream-agent cycle (why it is a harness runtime, not just a graph)
The graph does not grow by fiat โ it grows through the harness's dual-agent fold, which is the anti-Sybil
discipline made structural:
- Mage (proposer,
bnot) โ proposes the smallest edge: "these two members share a matching compression;
here is the candidate VRC." Conceal/reduce: it asserts, minimally. - Swordsman (prover,
neg) โ proves before signing: both endpoints personhood-anchored (PHC), no self-edge
(self-Sybil), mutual consent, fresh per-counterparty R-DIDs, and โ across the Gap โ it recomputes the
VRC commitment from the candidate's own public parts rather than trusting the Mage's claim (Fiat-Shamir
reseed, non-collusion). Signs only a validated edge; names and rejects every mirage. - The Gap (
โ) โ the proposal and the proof are held apart.neg(bnot(x)) = succ(x): the graph advances
by exactly one edge only when a reduction (the Mage's minimal proposal) is signed by a proof (the
Swordsman) that did not collude with it. A bare proposal never touches the graph (property G5a).
This is the same algebra as the shareable dual-agent harness; here the "artifact being made cheaper" is the
trust graph itself, and the "held-out gate" is edge validity.
Properties verified by test.mjs (11/11)
- G1 bilateral consent (unilateral rejected; mutual forms the edge).
- G2 personhood anchor required (a non-member endpoint is rejected โ PHC/VMC gate).
- G3 no self-edge (self-Sybil forbidden).
- G4 R-DID uniqueness โ one node's R-DID differs per counterparty (unlinkable, cred-spec R-DID rule).
- G5 the fold advances only on a signature (a bare Mage proposal does not grow the graph).
- G6 one VRC per pair (duplicate edge rejected, idempotent).
- G7 the Gap catches a forged VRC commitment (prover recomputes, never trusts the proposer).
- G8 community-anchored propagation (trust reaches transitively; an edgeless member stays unreachable).
Feedback to the strawman / cred-spec (candidate notes)
- Trust-graph formation is where our two ZK constructions get their operational definition. The cred-spec
names pairwise and community-anchored ZKP but does not say when an edge may be minted. This runtime's
Swordsman checks (personhood + mutual consent + fresh R-DIDs + non-collusion recompute) are a candidate
edge-admissibility rule for the "registry-ZK interaction" the spec defers โ worth raising as a profile. - The consent gate is a normative requirement, not a UX detail. G1 shows the shared graph must not grow on
unilateral action. Suggest the spec state bilateral-consent-to-edge explicitly (it currently implies it via
two-sided VRC issuance). - A VWC slots straight in. A third party witnessing an encounter is a VWC (taskContext-bound); adding a
witness channel to the cycle is the natural next extension (a Swordsman that also accepts a witness's
signed attestation of the collision). Not built here; flagged.
Next steps for this runtime
- Add a witness (VWC) channel: an optional third-party attestation of the encounter, taskContext-bound.
- Add VPC/P-DID personas (runtime 04) so a node can present different personas per edge.
- Port the edge-admissibility statement to the circom gadget once 01/03 land (Poseidon nullifier + Merkle
membership prove the endpoints without revealing the members).