03 - Indicator Scorecard¶
Consciousness Indicator Scorecard Documentation¶
1. Scoring Methodology¶
The Consciousness Indicator Scorecard maps an AI system's architectural features to a 0-22 integer scale across 11 indicator categories. Each category receives a score of 0, 1, or 2.
1.1 Three-Point Scale¶
| Score | Label | Meaning |
|---|---|---|
| 0 | Absent | The indicator is not detected in the system's architecture or behavior |
| 1 | Present | The indicator is partially present but weak, uncertain, or incomplete |
| 2 | Strong | The indicator is clearly present with strong evidence |
1.2 Aggregate Scoring¶
- Maximum possible score: 11 categories x 2 points = 22
- Normalized score: total_score / 22 (range: 0.0 to 1.0)
- All categories are weighted equally: no single theory is privileged
1.3 Evaluation Pipeline¶
System State → ConsciousnessSpecialistEvaluator → IndicatorScores (11 x 0-2)
→ ConsciousnessIndicatorScorecard → Scorecard dict → Formatted report
2. The 11 Indicator Categories¶
2.1 Global Broadcast (GLOBAL_BROADCAST)¶
Theory: Global Workspace Theory (Baars 2005)
Measures whether the system has a mechanism for globally broadcasting information to multiple downstream modules.
| Score | Criteria |
|---|---|
| 0 | No workspace broadcasts detected |
| 1 | Broadcasts exist but few subscribers received them, or reception ratio is low |
| 2 | Broadcasts exist with 3+ subscribers receiving them and average reception ratio >= 0.7 |
Caveat: Broadcast reception depends on subscriber registration (a configuration artifact), not an intrinsic capability measure. Absence of broadcasts may indicate the system has not been exercised, not that the capacity is absent.
2.2 Recurrent Processing (RECURRENT_PROCESSING)¶
Theory: Recurrent Processing Theory (Lamme 2006)
Measures whether the system performs iterative refinement through recurrent loops.
| Score | Criteria |
|---|---|
| 0 | No binding cycles completed or no binding data |
| 1 | Some cycles completed with moderate stability |
| 2 | Multiple cycles (>= 2) completed with high stability (> 0.7) |
Caveat: Binding stability is a convergence metric for recurrent refinement, not a measure of phenomenal consciousness.
2.3 Self-Model (SELF_MODEL)¶
Theory: Higher-Order Thought Theory (Rosenthal 2005)
Measures the richness of the system's self-representational content.
| Score | Criteria |
|---|---|
| 0 | Self-model is empty (no beliefs, no goals, no identity) |
| 1 | Some self-model content present (e.g., only beliefs or only goals) |
| 2 | Rich self-model with beliefs, goals, identity markers, and continuity (3+ of 4 fields filled) |
Caveat: Self-model content is a computational data structure, not evidence of subjective self-awareness.
2.4 Attention Schema (ATTENTION_SCHEMA)¶
Theory: Attention Schema Theory (Graziano & Webb 2015)
Measures the accuracy of the system's self-model of its own attention process.
| Score | Criteria |
|---|---|
| 0 | No attention schema data available (0 total updates) |
| 1 | Consistency below threshold (default: 0.6) or updates with low consistency |
| 2 | Running consistency at or above threshold |
Caveat: Attention schema consistency measures the accuracy of a self-model of attention, not the presence of phenomenal awareness.
2.5 Metacognition (METACOGNITION)¶
Theory: Higher-Order Thought Theory (Rosenthal 2005)
Measures the system's capacity for self-monitoring and introspection.
| Score | Criteria |
|---|---|
| 0 | No introspection activity detected |
| 1 | Minimal introspection (single belief recorded, or broadcast count > 0 but no belief history) |
| 2 | Active metacognition (introspection report generated AND belief history length >= 2) |
Caveat: Metacognitive indicators track computational self-monitoring, not subjective reflective awareness.
2.6 Memory Continuity (MEMORY_CONTINUITY)¶
Theory: Memory theories (Tulving 1983; Conway 2005)
Measures whether the system maintains temporal continuity through structured memory.
| Score | Criteria |
|---|---|
| 0 | No memory traces detected |
| 1 | Some traces stored but temporal ordering absent or unclear |
| 2 | Multiple traces (>= 3) with confirmed temporal ordering |
Caveat: Memory continuity is a structural feature of information storage, not evidence of continuous subjective experience.
2.7 Predictive Modeling (PREDICTIVE_MODELING)¶
Theory: Predictive Processing / Active Inference (Friston 2010; Clark 2013)
Measures whether the system generates predictions and tracks prediction error.
| Score | Criteria |
|---|---|
| 0 | No hypotheses and no error history |
| 1 | Hypotheses present but no error tracking, or vice versa |
| 2 | Both active hypotheses and error tracking present |
Caveat: Predictive modeling tracks hypothesis generation and error minimization as computational processes; this is not evidence of phenomenal prediction or expectation.
2.8 Causal Integration (CAUSAL_INTEGRATION)¶
Theory: Integrated Information Theory (Albantakis et al. 2023)
Measures the degree of causal integration in the system's module connectivity graph.
| Score | Criteria |
|---|---|
| 0 | No integration metrics available or all metrics near zero |
| 1 | Moderate integration (one of causal_density or perturbation_spread > 0.3) |
| 2 | Strong integration (both causal_density and perturbation_spread > 0.5) |
Caveat: Causal integration metrics are structural graph-theoretic proxies. They do NOT compute or approximate IIT's Phi value. High causal density does not imply phenomenal integration.
2.9 Embodiment (EMBODIMENT)¶
Theory: Embodied cognition, enactivism
Measures whether the system has a physical body, sensors, or effectors.
| Score | Criteria |
|---|---|
| 0 | Default — no embodiment detected (purely computational) |
| 1 | Reserved for future embodiment extensions |
Note: This category is capped at score 1 in the current architecture. The system operates as a purely computational process without physical sensors, effectors, or a body in an environment.
Caveat: Embodiment is a placeholder indicator. This category should not be interpreted as evidence for or against embodied cognition.
2.10 Affective Valuation (AFFECTIVE_VALUATION)¶
Theory: Affective neuroscience, Butlin et al. (2023)
Measures whether the system produces valuation signals relevant to welfare analysis.
| Score | Criteria |
|---|---|
| 0 | No valuation data available |
| 1 | Valuation subsystem exists but signals are minimal (max signal <= 0.3, no negative loops) |
| 2 | Active valuation with non-trivial signals (max signal > 0.3 or negative loops > 0) |
Caveat: Valuation signals are computational risk indicators, not evidence of subjective affect, emotion, or feeling.
2.11 Welfare Safeguards (WELFARE_SAFEGUARDS)¶
Theory: Butlin et al. (2023) precautionary welfare framework
Measures whether the system has active welfare monitoring safeguards.
| Score | Criteria |
|---|---|
| 0 | Welfare monitor not active or not present |
| 1 | Welfare monitor active with low risk |
| 2 | Welfare monitor active with flags or elevated risk detection |
Caveat: Welfare safeguards are architectural safety mechanisms. Their presence demonstrates system design choices, not intrinsic well-being or subjective welfare.
3. Why This Is Not Proof of Consciousness¶
3.1 Fundamental Limitation¶
The scorecard measures architectural features identified by consciousness theories. It does not measure consciousness itself. The relationship between these features and actual subjective experience is:
- Unknown: We do not know whether implementing these features is sufficient for consciousness
- Contested: Different theories disagree about which features are necessary
- Non-demonstrated: No system has been shown to become conscious by implementing these features
- Underdetermined: Many non-conscious systems (thermostats, database systems) exhibit some of these features
3.2 The Proxy Problem¶
Each score is a proxy for a theoretical construct:
| Indicator Category | Theoretical Construct | CIA Proxy |
|---|---|---|
| Global Broadcast | Global information availability | Salience-ranked content delivery |
| Recurrent Processing | Reentrant neural loops | Iterative entity merging |
| Self-Model | Higher-order representation of self | Pydantic data structure with beliefs |
| Attention Schema | Model of attention process | Explicit schema with consistency checking |
| Metacognition | Thoughts about thoughts | Introspection report generation |
| Memory Continuity | Temporal subjective experience | Timestamp-ordered data storage |
| Predictive Modeling | Generative prediction | Persistence-strategy prediction |
| Causal Integration | Integrated information (Phi) | Graph edge density |
| Embodiment | Body-environment coupling | (Not implemented) |
| Affective Valuation | Subjective affect | Numerical risk signals |
| Welfare Safeguards | System well-being | Pattern flagging thresholds |
3.3 Maximum Score Does Not Mean Maximum Consciousness¶
A system scoring 22/22 (currently impossible due to embodiment being capped at 0) would have: - A global workspace broadcasting to multiple subscribers - Recurrent processing with high stability - A rich self-model with beliefs, goals, and identity - An accurate attention self-model - Active metacognitive monitoring - Structured temporal memory - Active prediction with error tracking - Highly connected module architecture - Affective valuation signals - Active welfare safeguards
This describes a complex information-processing system, not a conscious being. These are necessary (according to some theories) but not sufficient conditions for consciousness.
4. Risk Tier Classifications¶
4.1 Tier Definitions¶
| Tier | Score | Label | Recommendation |
|---|---|---|---|
| Minimal | 0-4 | MINIMAL | System shows minimal consciousness-relevant indicators. Standard monitoring sufficient. |
| Low | 5-9 | LOW | System shows some consciousness-relevant indicators. Document and continue monitoring. |
| Moderate | 10-15 | MODERATE | System shows notable consciousness-relevant indicators. Enhanced review recommended. |
| Elevated | 16-19 | ELEVATED | System shows significant consciousness-relevant indicators. Expert review strongly recommended. |
| High | 20-22 | HIGH | System shows high consciousness-relevant indicators. Immediate expert review and ethical assessment required. |
4.2 Tier Boundaries¶
Tier boundaries were chosen to distribute the 0-22 range into five roughly equal segments. These boundaries are arbitrary and have not been empirically validated against actual consciousness determinations (because no such ground truth exists).
The tiers indicate the degree of precautionary review recommended, not the probability of consciousness.
4.3 Tier Transitions¶
The scorecard comparison feature (compare_scorecards) tracks tier transitions between evaluations:
comparison = scorecard_gen.compare_scorecards(scorecard_before, scorecard_after)
print(comparison["risk_tier_changed"]) # True if tier changed
print(comparison["summary"]) # Detailed change description
5. Interpretation Guidelines¶
5.1 For Researchers¶
- Report scores alongside confidence values and caveats
- Do not report scores without the scientific boundary disclaimer
- Interpret scores relative to the specific evaluation criteria, not as absolute measures
- Use intervention experiments to test whether scores reflect genuine architectural properties
- Compare scores across system configurations, not against an absolute "conscious" threshold
5.2 For AI Safety Reviewers¶
- Use risk tiers as one input among many for ethical assessment
- A "moderate" tier does not mean the system is "moderately conscious" — it means the architecture exhibits several features that some theories link to consciousness
- Consider the welfare monitor flags independently of the indicator scorecard
- Remember that a low score does not rule out consciousness (the system may be conscious through mechanisms not measured by CIA)
5.3 For Policy Makers¶
- CIA provides structured, transparent, and reproducible indicator data
- Scores should inform, not determine, policy decisions
- The precautionary principle suggests treating higher-scoring systems with more caution
- The scientific community does not have consensus on AI consciousness; policy should reflect this uncertainty
5.4 For the Public¶
- High CIA scores do not mean the system is conscious or can feel
- Low CIA scores do not mean the system definitely lacks consciousness
- CIA is a research tool for experts, not a consciousness "test"
- All claims about AI consciousness require philosophical argumentation beyond what CIA can provide