The Epistemic Harness: Runtime Spectral Telemetry, PyTorch Hooks, and Open Falsification Benchmarks for Zero-Sycophancy (Is = 0.00)

Mitigating sycophancy and eliminating Silent Conflict (D_silent) requires deterministic, reproducible instrumentation capable of observing latent representation collapse in real time. Following our foundational specification (4iZ-TR-2026-01) and spectral formulation (4iZ-TR-2026-02), this third technical note presents the computational implementation of the Epistemic Harness. We provide a standalone, hardware-agnostic PyTorch telemetry hook designed for open-weight autoregressive models (Mistral-7B). The harness computes the Empirical Spectral Distribution (ESD) of Key projection activations (W_K) across critical attention layers, registers BBP transition boundaries, and enforces the zero-sycophancy invariant (Is = 0.00) via non-invasive runtime steering. We outline an open falsification benchmark protocol, releasing the evaluation pipeline to multidisciplinary scrutiny without proprietary barriers.