Sardis Guard for Distributed Inference

Parent: distributed-inference

This package is a distributed inference runtime. It may describe topology, transport, cache replay, standing-wave compression, and measured mesh behavior. It should not graduate those claims into "distributed intelligence" unless the live carrier actually bears that label.

The formal boundary is Gnosis.SardisDistributedIntelligenceGuard in open-source/gnosis-math/Gnosis/SardisDistributedIntelligenceGuard.lean. It applies Gnosis.SardisGateAdmissibility to this package's naming surface:

• distributedInferenceCarrier is admitted because compiled targets, measured runs, and closure certificates cover the advertised obligations.
• sardisDistributedIntelligenceCarrier is rejected because the stronger name claims more than the live carrier bears.
• verbalOnlyDistributedIntelligenceCarrier is rejected because standing-wave or intelligence language without compiled, measured, or certified carrier evidence is a dead-name risk.

Naming Rule

Use distributed inference for the current Rust/WASM/runtime/mesh system. Use distributed intelligence only for a bounded claim that names its live carrier: a passing build, a measured run, a closure certificate, and the exact obligation the label asserts.

Do not use a resonance phrase, a topology phrase, or a true-but-task-empty phrase as admission evidence. Gnosis.Bridges.VeridicalSwayBridge already separates that case: veridical sway may help a reader settle the story, but it has zero task-information gain. That is not a distributed-intelligence carrier.

Runtime Hook

The Rust runtime mirrors this boundary in src/sardis_guard.rs. The guard classifies cue shape as usefulEvidence, veridicalSway, vacuousTruth, or phantomBullshit, then admits only task-bearing runtime artifacts with live carrier evidence. protocol69-gen-agent attaches the guard report to each successful legAck as sardisGuard and fails closed before acking an empty or task-empty generated artifact.

Admission Checklist

Before adding or preserving an intelligence-level claim, require all of:

1. A current command that builds or tests the surface.
2. A persisted measurement or certificate when the claim is empirical.
3. A theorem-lineage reference when the claim cites formal grounding.
4. A named fallback or refusal when the gate fails.
5. Wording that states the bounded relationship, not an emphatic identity.

Refusal Pattern

Prefer this shape:

This is a distributed-inference carrier with standing-wave and mesh evidence. It is not yet admitted as distributed intelligence; that label requires additional live carrier obligations named in SardisDistributedIntelligenceGuard.

That refusal keeps the work alive without giving it a dead name.