Meridian — where trust is measured, not claimed. An open protocol for reputation-routed trust networks.
| Rating | Summary | |
|---|---|---|
| Quality | poor (7/24) | A well-articulated design document with no implementation, no releases, no tests, and minimal adoption signals. |
| PAI Relevance | note (0.63) | Addresses a clear gap in PAI's agent stack — no behavioral trust or reputation layer exists — but there is nothing to integrate today; the design is worth tracking. |
7/24 — stale-risk / under-documented / no-signals
Failed:
Passed:
Failed:
Passed:
Failed:
Passed:
| Dimension | Score | Assessment |
|---|---|---|
| Harvest Value | 1 | The decay-weighted trust vector model and Sybil-resistance-via-verifiable-work pattern are novel relative to PAI's current agent coordination primitives (Comms, Delegation, Agents skills), and the daemon observation-chain design is worth studying for PAI's multi-agent trust layer; but harvest is architectural inspiration only, not a reusable algorithm or implementation. |
| Integration Readiness | 0 | Language unknown, no dependency manifest, no SDK, no CLI, no code at all — pure spec at pre-alpha; nothing can be bun add-ed or subprocess-called today. |
| Overlap Risk | 0 | No PAI skill, tool, or hook performs behavioral trust scoring or reputation propagation between agents; NorthwoodsSentinel--brook handles fleet monitoring/overwatch but not trust routing, confirming this fills a clear gap rather than duplicating anything. |
| Gap Fill | 2 | PAI's 27-agent fleet has no reputation or trust-scoring layer — inter-agent trust is implicit and uncalculated; Meridian targets exactly this absence, making it the only design in the vault that addresses behavioral trust between daemons. |
Composite: 0.63
Building a multi-agent pipeline where agents evaluate each other's outputs: Implement the Meridian spec's decay-weighted trust vector as a lightweight JSON ledger — one record per (observer, subject, domain) triple, with a last_updated timestamp and a score multiplied by e^(-λt) on each read. Agents that haven't been recently observed in a given domain lose standing automatically, replacing the naive pattern of orchestrator-assigned uniform trust with credibility that degrades without continued verification.
Open-participation systems prone to Sybil attacks (bounty boards, prediction markets, federated agent registries): Adopt Meridian's proof-of-work seeding requirement — require new entrants to produce one small, verifiable work artifact before their trust score rises above a floor value. This closes the zero-cost registration exploit that plagues reputation graphs without requiring identity documents or central approval.
Designing a federated reputation layer for a distributed service: Use the propagation graph model from the spec to cap transitive trust chains at two hops with an explicit discount factor per hop — A endorses B, B endorses C, A's effective score for C equals score_AB × score_BC × discount. Encoding this as a directed weighted graph with a depth limit produces an auditable web of trust that is bounded, inspectable, and resistant to long-chain reputation laundering.
Category: AI Agent Frameworks
In this category: VoltAgent--voltagent (the only existing entry)
Standing: Meridian is the sole trust/reputation-protocol entry in the agent-frameworks category; VoltAgent--voltagent addresses orchestration and execution, while Meridian would address inter-agent trust scoring — complementary rather than competitive, but Meridian is a spec-only pre-alpha against VoltAgent's production-grade TypeScript platform.
Density: 6/10 — Available: README (full, ~3KB), repo metadata (stars, forks, dates, license, archived status), landscape context, overlap clusters, related prior appraisals. Missing: dependency manifest, source code, CI configuration, release notes, contributor history, issue tracker content.
This is a NorthwoodsSentinel project (same author as Brook, Loam, PII Scrub) and sits at the conceptual apex of that ecosystem — where Brook monitors the fleet and Loam stores memory, Meridian would provide the trust layer that decides which agents to believe. The design borrows cleanly from real-world trust propagation literature (decay, domain-specificity, Sybil resistance) and the Linux/Red Hat / TCP/IP open-commons analogy is structurally sound. The risk is that the vision is substantially ahead of the implementation: the repo is a six-day burst of writing (created 2026-03-24, last commit 2026-03-30) with no subsequent activity. Watch for resumption of commits or a formal protocol spec document before treating this as more than a design reference.