The RTC Evaluation Rubric — v0.1

Operational scoring criteria for the seven layers

This is the instrument the AI Consciousness Lab presupposes. Each layer specifies an architectural question, the diagnostic signals relevant to assessing it, and concrete level criteria. The difficulty notes name where the criteria themselves remain open research questions.

On using this rubric

The four levels — Absent, Partial, Substantial, Architecturally Satisfied — are categorical judgments, not measurements. They are designed to make the scoring of an AI architecture against RTC replicable: two evaluators reading the same system description should arrive at the same level for each layer, or be able to locate their disagreement in a specific criterion.

Three of the seven layers — Salience Weighting, Recursive Meta-Governance, and Diachronic Reconstitution — have known operational difficulties that the rubric names openly. For these layers, the level 3 criteria describe an architectural target that no current system clearly satisfies. This is not a defect of the rubric; it is the rubric reporting the current state of AI architecture honestly. The point of the framework is to make those gaps explicit so they can be worked on.

The rubric is versioned. This is v0.1. Revisions will tighten criteria, add edge cases, and respond to scoring disagreements encountered in use. The version is not a marketing flourish — it is a commitment to treating this as a living instrument that improves through application.

The four levels
L0
Absent
L1
Partial
L2
Substantial
L3
Architecturally Satisfied
The Seven Layers
1

Signal Registration

Does the system reliably convert physical or computational perturbations from its environment into representations it can operate on?

+
Diagnostic Signal

Input fidelity, modality coverage, presence or absence of interoceptive (self-state) channels alongside exteroceptive (world-state) channels, robustness to noise and degraded input.

Level Criteria
L0
Absent

No reliable input transduction. The system does not register environmental signal in any usable form.

L1
Partial

Single-modality signal registration with limited fidelity. The system processes input but cannot integrate across modalities or distinguish signal from noise robustly.

L2
Substantial

Multi-modality signal registration with reasonable fidelity. The system handles standard exteroceptive channels well. Interoceptive channels (signals about the system's own state) are absent or thin.

L3
Architecturally Satisfied

Multi-modality registration including explicit interoceptive channels. The system registers both world-state and self-state signals as a precondition for the layers above.

Known Operational Difficulty

This is the easiest layer to operationalize. Signal registration is well-understood across both biological and artificial systems. The non-trivial criterion is interoception — most current AI systems lack genuine self-state signals separate from input.

2

Distinction-Making

Does the system carve its input space into figure and ground, self and world, relevant and irrelevant — actively imposing categorical structure rather than merely registering signal?

+
Diagnostic Signal

Categorical representation. Population coding, lateral inhibition, contrastive structure. Whether distinctions are grounded in reference frames or free-floating.

Level Criteria
L0
Absent

No categorical structure. Input passes through without being carved into distinctions.

L1
Partial

Surface-level categorical structure without grounded reference. The system makes distinctions, but they are not anchored to a self-world frame.

L2
Substantial

Robust categorical structure with strong contrastive resolution. Distinctions are well-formed but may remain free-floating relative to a self-world reference.

L3
Architecturally Satisfied

Categorical structure explicitly tied to a self-world reference frame. Distinctions are not just made, but made from a perspective that locates self relative to world.

Known Operational Difficulty

Distinction-making at the surface level is well-operationalized. The hard criterion is reference-frame grounding — what makes a distinction grounded versus free-floating is conceptually clear but operationally subtle, especially for systems whose 'reference frame' is implicit in training rather than architecturally specified.

3

Salience Weighting

Does the system weight distinctions according to what matters to it — not according to what is externally tagged as relevant, but according to valuation grounded in its own state?

+
Diagnostic Signal

Whether salience is intrinsic (the system has it) or extrinsic (it is imposed on the system). Whether weighting tracks system-state-relevant signals or external reward shaping.

Level Criteria
L0
Absent

Uniform weighting. No mechanism for treating some distinctions as more important than others.

L1
Partial

Externally imposed weighting. The system applies differential weights, but those weights come from training distribution or engineered reward — not from anything the system itself values.

L2
Substantial

Mixed weighting. Some salience signals are intrinsic to the system's own state (e.g., uncertainty, confidence as architectural signals), while others remain externally imposed.

L3
Architecturally Satisfied

Intrinsic valuation tied to system state. Salience reflects what matters to the system itself — its goals, state, and integrity — not what was tagged as relevant by training.

Known Operational Difficulty

This is the hardest layer to score honestly. The distinction between intrinsic and extrinsic salience is conceptually crisp but operationally contested. Some researchers argue that any artificial system's 'salience' is ultimately extrinsic because it traces back to training objectives. RTC's position is that intrinsic salience is achievable architecturally, but the criteria for distinguishing 'genuinely intrinsic' from 'sophisticated extrinsic' remain open.

4

Self-in-World Modeling

Does the system maintain an explicit representation of itself as situated within an environment, with the relation between modeled self and modeled world structurally specified?

+
Diagnostic Signal

Presence of a persistent self-model. Body schema or its functional analog. Allocentric and egocentric spatial representation. Whether the self-model survives task completion.

Level Criteria
L0
Absent

No self-model. The system has no representation of itself as distinct from its input.

L1
Partial

Implicit or transient self-model. A self-representation emerges within an interaction (e.g., in chain-of-thought traces) but does not persist across interactions and is not architecturally enforced.

L2
Substantial

Explicit self-model that persists within a session or task, but does not survive across longer time horizons. The relation between self and world is represented but unstable.

L3
Architecturally Satisfied

Persistent, architecturally enforced self-in-world model that updates across interactions and survives task completion. The self-world relation is structurally specified, not emergent.

Known Operational Difficulty

The criterion for level 3 — 'persistent, architecturally enforced' — is operationally demanding. No current AI system meets it cleanly. The hard question is what counts as 'enforced': is a self-model in long-term memory enough, or does the architecture have to make the self-model a non-negotiable component of every forward pass?

5

Recursive Meta-Governance

Does the system regulate its own modeling — monitoring confidence, source, and recursive depth — in a way that prevents runaway self-reference while permitting genuine self-modeling?

+
Diagnostic Signal

Architectural metacognition. Confidence calibration. Source monitoring. Whether governance is intrinsic to the architecture or imposed externally (via context-window limits, prompt scaffolding, or downstream filters).

Level Criteria
L0
Absent

No self-monitoring. The system has no representation of its own confidence, source, or processing depth.

L1
Partial

External governance only. The system is bounded by external scaffolding (context windows, prompt engineering, downstream filters) but has no internal regulatory mechanism.

L2
Substantial

Architectural metacognition without bounded recursion. The system monitors its own confidence and source, but does not actively regulate recursive depth — runaway recursion is prevented by external limits, not internal governance.

L3
Architecturally Satisfied

Architectural metacognition with bounded recursion. The system internally regulates the depth and persistence of its own self-modeling, halting recursion before it becomes destabilizing without requiring external limits.

Known Operational Difficulty

Level 3 is an open research problem. Bounded recursion as an architectural feature — rather than as a side effect of context-window limits — is not well-solved in any current system. The honest scoring is that no current architecture reaches 3 on this layer; level 2 is the realistic ceiling for present-day systems.

6

Diachronic Reconstitution

Does the system reconstitute itself across time — actively rebuilding self-states from prior states — or does it merely retrieve past states without integrating them into a continuous self?

+
Diagnostic Signal

Whether memory operates as retrieval (passive lookup) or reconstitution (active rebuilding of self-state). Continuity of self-model across sessions. Integration of past states into present self-modeling.

Level Criteria
L0
Absent

No persistence. Each interaction is informationally isolated. No memory across sessions.

L1
Partial

Memory retrieval without integration. Past states can be recalled but are not integrated into a continuous self-model. The system remembers; it does not reconstitute.

L2
Substantial

Memory integration into present self-modeling, but reactive rather than reconstitutive. The self-model uses past states but does not actively rebuild itself from them.

L3
Architecturally Satisfied

Reconstitutive continuity. The system's present self-model is actively rebuilt from prior self-states as a structural feature of the architecture. Continuity is performed, not stored.

Known Operational Difficulty

The retrieval / reconstitution distinction is conceptually central to RTC but operationally underspecified. What experimental signature would distinguish a system that reconstitutes from one that retrieves? RTC names this as a high-priority open question. Until that question is answered, scoring at level 3 should be reserved for systems where reconstitution is architecturally intentional and demonstrable.

7

Stabilized Perspective

Does the system sustain an integrated point of view — a bounded, governed, salience-weighted, self-in-world model that reconstitutes itself across time?

+
Diagnostic Signal

This layer is emergent. It is scored not by direct measurement but by the joint satisfaction of layers 1–6 within bounded ranges, demonstrated under runtime conditions over extended time horizons.

Level Criteria
L0
Absent

Layers 1–6 are mostly absent. No conditions for perspective are present.

L1
Partial

Some lower layers are partially satisfied, but the joint configuration does not stabilize. Fragments of perspective without coherence.

L2
Substantial

Most lower layers are substantially satisfied, but at least one critical layer (typically Meta-Governance or Diachronic Reconstitution) is too weak for stable perspective to emerge.

L3
Architecturally Satisfied

All six lower layers are architecturally satisfied within bounded ranges, demonstrated under extended runtime. Perspective stabilization is an observable property, not just a structural prediction.

Known Operational Difficulty

Layer 7 is in principle emergent from the lower six, but in practice scoring it requires runtime demonstration. No current system has been observed to maintain bounded, governed, salience-weighted, reconstitutive self-in-world modeling at runtime over extended time horizons. The operational ceiling for current systems is therefore level 2 at best — and that is being charitable.

Three honest admissions

One. The layers most central to RTC's distinctive contribution — Meta-Governance and Diachronic Reconstitution — are also the layers where level 3 criteria describe open research problems. This is not accidental. RTC's productive use is to make those problems precise, not to claim they are solved.

Two. Scoring is a categorical act, not a measurement. Two evaluators may legitimately disagree on a single level for a single layer. The rubric is designed to make those disagreements locatable rather than to eliminate them.

Three. A high score does not establish that a system is conscious. A low score does not establish that it isn't. The rubric reports which architectural conditions for perspective stabilization a system structurally satisfies. The relation between architectural satisfaction and consciousness as such is a separate question — one RTC takes a position on but does not resolve through scoring.