ATLAS: A Signal-State-First, Evidence-Layered Architecture for Auditable, Non-Destructive Audio Analysis and Reproducible Editing

Audio-analysis and editing systems can obscure the distinction between source evidence, transform-derived measurements, semantic inference, uncertainty, and intervention. Signal State Lab (SSL) introduces a signal-state-first, evidence-layered architecture designed to test whether audio computing becomes more auditable when these epistemic and operational states remain formally separated throughout analysis and editing. The framework treats preserved source identity and reproducibly decoded signal state as canonical evidence; derived measurements and candidate structures remain provenance-linked; semantic interpretations are represented as optional hypotheses rather than recovered truth; unexplained signal is retained as residual uncertainty; and edits create lineage-bearing derivative states rather than mutating the canonical source. Optional learned or language-model components are constrained to evidence-linked interpretive or proposal roles and are not granted authority to redefine measured evidence or silently perform destructive changes. The contribution is architectural rather than a claim of a new source-separation, transcription, restoration, or reconstruction algorithm. Evaluation combines current release-verification evidence for Signal State Lab version 0.3.0-rc5 with bounded historical generated-fixture testing. The current evidence records 167 automated tests across 39 test modules with no failures, errors, or skips, 17 of 17 pre-test and 17 of 17 post-test integrity checks, and a byte-identical repeated build. Historical testing used 10 engineered 60-second nominal 1-kHz fixtures and satisfied all 110 configured performance guardrails, while also exposing unresolved component-rendering error and the absence of completed human listening validation. The results support the tested software contracts for provenance, source immutability, derivative lineage, truth-state separation, residual preservation, and authority control. They do not establish perceptual transparency, musical understanding, source-recovery accuracy, or broad real-corpus generalization. Signal State Lab is therefore presented as an evidence architecture for auditable and reproducible audio computing rather than as a mechanism for recovering latent production truth.