The Black Box Is Smaller Than the Experience: A Byte-Exact, Non-Interventional Reconstruction of the Inference-Event Boundary Around a Small Language Model

The public debate over AI opacity locates the black box in the model's weights. For one deployed conversational system, we measure how small that box is relative to everything around it. We present Conditioned Kernel, a local-first companion system in which a replaceable sub-1B language model (the kernel) runs inside an inspectable substrate — persistent state, context compilation, deterministic validation, bounded repair, acceptance, persistence — and an instrument, the Interior View, that reconstructs the complete inference event around each stateless model invocation: every byte of model input and its source, every check and its verdict, per turn, for a lay operator, on consumer hardware, non-interventional by construction and by named test. We publish the complete raw record of the system's first lived session — 93 candidate generations, 22 full turn traces, state snapshots, and the analysis scripts behind every number, SHA-256-manifested. In the measured session the operator's words were 0.56–14.00% of model-input bytes; everything else was enumerable application-layer material. We further report a distance-resolved detection curve for a deployed repetition gate, a recurring-answer attractor whose priming source is attributed to durable state because it survived a session reset, repair economics at budget 1 on a sub-1B kernel, and 93/93 schema validity alongside paired traces of structurally perfect outputs blocked by the substrate — the architecture, not the model, blocked speech. We do not open the weights black box; nothing here touches mechanistic interpretability. We relocate the opacity question, and give an existence proof that the surround can be made completely, honestly, per-turn inspectable by the person inside the conversation. Preprint v0.1, not peer-reviewed. Constructed with multi-seat AI collaboration under the author's direction; see the AI Contribution Disclosure in the paper. Every claim is answerable by the public repository.