Answers are grounded in this paper and cite their sources.
Preregistered protocol measures machine consciousness components in self-aware networks, indicating testing challenges.
Artificial systems can imitate reports, confidence, memory, personality, and self-description without thereby establishing subjective experience. This makes behavior-only tests vulnerable to a gaming problem: an indicator can be optimized while the mechanism it was intended to track is absent. This paper converts a 2021 Self Aware Networks (SAN) proposal for a guidebook on recognizing possible machine consciousness into a preregisterable machine-assessment protocol. The protocol is a companion to Tomographic Coherence-based Neural Rendering (TCNR), not a replacement for it. TCNR supplies a relational reconstruction hypothesis and a historical Tomographic Consciousness Index (TCI). The present protocol does not treat the historical weighted TCI as a consciousness meter. It instead measures a typed claim vector covering predictive information, reconstruction, relational phase organization, routed propagation, closed-loop influence, memory integration, counterfactual capacity, access, and self/body/agency modeling. Each component is tested against declared surrogate nulls, mechanism-positive and mechanism-negative white-box systems, capacity-matched rivals, and systems designed to produce false positives. Selective interventions must cause the predicted component-level and functional loss; a component-matched rescue must restore the ordered effect without merely restoring all computation. Evaluation is grouped by architecture family, training run, and world so that an estimator cannot pass by recognizing metadata or one implementation style. Confidence is calibrated on known implemented relations, stress-tested under distribution shift, and allowed to abstain. The output separates four judgments: whether a relation was measured validly, whether a TCNR/SAN mechanism profile is supported, whether welfare precautions are proportionate, and whether legal or moral status has been established. The first two are scientific outputs of this protocol; the latter two require separate governance. Even a complete pass supports an implemented mechanism profile, not proof of phenomenal consciousness. Draft 5 preserves every Draft 4 correction and adds the missing learned-family stress test. Four structurally different temporal systems fit independent multi-output readouts on 24 declared training traces. They are then evaluated without tuning on 16 unseen-seed traces and 20 traces from an unseen sensorimotor remapping. All four fail the noncompensatory nine-component profile on the heldout world. Each retains only the relational-template component while future prediction, reconstruction, feedback, memory, counterfactual, routed-propagation, access, and self-remapping components fail their frozen thresholds. This is a useful negative result: a phase-relation score can remain high when the joined mechanism profile and predictive control fail. The executed systems use frozen temporal features and learned readouts rather than end-to-end learned deep networks, and the benchmark remains under project-local rather than independent custody. Draft 5 therefore validates bounded synthetic and learned-readout measurement mechanics, not consciousness or deployed-system performance.
No takes yet. Share an insight, caveat, or question.
Micah Blumberg (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: