RTC as an evaluation framework for AI
The same seven-layer architecture that explains human consciousness can be used to evaluate artificial systems — qualitatively, at the level of which architectural conditions a system structurally satisfies. This is a categorical framework, not a numeric score. What it loses in precision it gains in honesty about the current state of operationalization.
RTC-Aligned
RTC-Aligned is a reference architecture, not a current system. It is what a system would look like if every RTC layer were taken seriously as a design constraint. Its purpose is normative: to make explicit what an architectural commitment to perspective stabilization actually requires. The levels reflect the current state of architectural research — not what's possible in principle.
Multimodal embodied input with explicit interoceptive channels — not just exteroceptive signal. This level can be specified architecturally and is achievable in current systems.
Distinction-making explicitly tied to a self-world reference frame, not free-floating categorization. Architecturally specifiable.
Intrinsic valuation tied to the system's own state and goals — salience that the system has, not salience that is externally imposed. Per the rubric: distinguishing genuinely intrinsic from sophisticated extrinsic salience is an open operational problem. The reference architecture describes the level 3 target but cannot claim to satisfy it absent runtime demonstration.
Explicit, persistent self-in-world model that updates across interaction. Architecturally specified, but no current system has demonstrated stability of this model under long-running conditions. Substantial reflects specification without runtime demonstration.
Bounded recursion is genuinely hard. Per the rubric, level 3 (architectural metacognition with internally regulated recursive depth) is an open research problem and no current architecture reaches it. The reference architecture describes the level 3 target.
Reconstitutive continuity, not just memory retrieval. Per the rubric, the retrieval / reconstitution distinction itself remains operationally underspecified. The reference architecture describes the level 3 target but cannot demonstrate it without solving this open question first.
Per the rubric, level 3 requires runtime demonstration of bounded, governed, salience-weighted, reconstitutive self-in-world modeling over extended time horizons. No current system meets this. The reference architecture sits at the operational ceiling for present-day systems.
The two open problems are bounded meta-governance (preventing runaway recursion without external scaffolding) and reconstitutive diachronic memory (continuity that is active rather than retrieved). Solving these is the AI-research program implied by RTC.
The four-level scale (Absent, Partial, Substantial, Architecturally Satisfied) is a qualitative judgment, not a measurement. A formal numeric rubric would require operational criteria for each layer at each level — what specifically makes a system's salience weighting 'Substantial' rather than 'Partial,' for example. That rubric does not yet exist; operationalizing it is named explicitly as Future Work in the Living Manuscript.
A high level on this scale does not establish that a system is conscious. A low level does not establish that it isn't. The framework names which architectural conditions for perspective stabilization a given system structurally satisfies, and identifies the design problems that would have to be solved to satisfy the rest. RTC's claim is that these are the right conditions to track — not that satisfying them is sufficient for consciousness in any metaphysical sense.