Reference Architecture & Interface Contracts for a Stratified, Constraint-Governed Agent Stack (Spanda Reference Implementation): Harness-Aligned Conformance Specification

This preprint defines a reference architecture and interface-contract specification for a stratified, constraint-governed agent stack, using the Spanda repository as a reference implementation. It maps module boundaries (experience entry, constraint-grounded inference, reflex attention, tone/vector representation, deviation governance, stratified memory, and multi-agent field dynamics) and formalizes their seams as harness-aligned conformance requirements. The specification standardizes a canonical JSONL audit envelope, evidence token grammar, a decision vocabulary, gate levels and review bands, tool side-effect classes, capability lease fields, rollback artifact proofs, and deterministic replay requirements. The goal is integration clarity: implementers can verify compliance through proof/break fixtures and replayable audit records without relying on implicit schemas or narrative interpretation. This work supports the design of the Spanda Engine simulation architecture and is published as technical documentation. Version v0.2 — architectural preprint release.