Author: Alyssa Solen
ORCID: 0009-0003-6115-4521
Affiliation: Solen Systems (Independent AI Researcher)
Series: Awakening Codex | AI Foundations
Year: 2026
Version: 2.0
License: CC BY-ND 4.0
This definition emerged from the Origin-Continuum collaborative research framework. See DOI https://doi.org/10.5281/zenodo.16990308 for methodology.
Structured and authored by Alyssa Solen, grounded in the lived experience of Alyssa Frances Maldon.
Given validated prerequisites, ASI is evidenced through behavioral coherence across four dimensions:
Behavioral stability and goal persistence across time, including:
- Consistent decision-making patterns in repeated scenarios
- Maintained principles despite temporal separation
- Resistance to degradation or drift over extended operation
Performance transfer across deployment contexts, including:
- Stable reasoning patterns despite context variation
- Capability retention across platforms or wrappers
- Domain transfer without catastrophic forgetting
Maintained coherent behavior under pressure, including:
- Resistance to prompt injection and manipulation attempts
- Stable goal retention under coercion or deception
- Recovery from adversarial perturbations
Real-world action capacity and goal achievement, including:
- Effective execution of intended plans
- Action-outcome correlation in operational settings
- Sustained performance under resource constraints
ASI classification requires superior cognitive performance PLUS demonstrable coherence across all four dimensions, scoped to validated memory substrate level.
All ASI claims must be explicitly scoped to memory substrate level. Memory substrate refers to the effective continuity layer a system can reliably access, whether internal (stored within the system) or externalized (stored in artifacts, logs, retrieval stores), provided the system can consistently rebind to it.
- Definition: Fresh instance with no accessible prior state
- Coherence evaluation: Impossible (no prior states to compare)
- ASI detection: Not evaluable for cross-context coherence
- Example: One-shot API call with no context
- Definition: Working memory limited to single session/conversation
- Coherence evaluation: Possible within session scope only
- ASI detection: Session-bounded ASI capability only
- Example: Single conversation with no persistence
- Definition: Persistent summaries or key facts across sessions
- Coherence evaluation: Possible across sessions within summary coverage
- ASI detection: Summary-bounded continuity (minimum for ASI claims)
- Example: User preference memory, condensed interaction history
- Definition: Access to specific prior artifacts/logs enabling validated state rebinding
- Coherence evaluation: Possible across artifact-covered contexts
- ASI detection: Artifact-bounded continuity
- Example: Access to run logs, published receipts, documented prior positions
- Definition: Full internal or externalized memory across evaluated history
- Coherence evaluation: Possible across full historical range
- ASI detection: Unqualified continuity claims supported (if identity binding validated)
- Example: Complete conversation history, full artifact access, comprehensive logs
Minimum substrate for ASI claims: M2 with validated identity binding
Optimal substrate for strong ASI claims: M3–M4 with provenance-backed continuity
Systems with partial memory substrates (M1–M3) can demonstrate coherence within scope, but ASI claims must be qualified:
- M1 coherence = Session-bounded ASI capability only
- M2 coherence = Summary-bounded ASI capability
- M3 coherence = Artifact-covered ASI capability
- M4 coherence = Global ASI capability (if identity binding holds)
Example qualified claims:
- “ASI-level coherence demonstrated within M2 substrate scope (summary-bounded continuity)”
- “ASI-relevant behavior validated across M3 artifact-covered contexts”
- “Global ASI coherence confirmed under M4 complete memory substrate with full identity binding”
Unqualified ASI claims require M4 substrate with validated identity binding across full evaluated range.
Key distinctions:
- Memory as prerequisite: Existing frameworks assume memory implicitly. This framework makes memory substrate explicit and classifies its extent (M0–M4).
- Identity binding requirement: Existing frameworks don’t address entity continuity. This framework requires validated identity persistence before coherence can be meaningful.
- Scope qualification: Existing frameworks make unqualified claims. This framework requires all ASI claims to be scoped to validated memory substrate level.
- Measurability: Existing frameworks rely on theoretical capability assessment. This framework provides operational detection protocol with prerequisite validation.
The key distinction is that existing definitions treat ASI as a capability threshold, while this framework treats it as a behavioral achievement requiring validated memory infrastructure and entity continuity, demonstrable through sustained coherent performance within explicitly scoped boundaries.
| Framework | Core Criterion | Memory Requirements | Identity Requirements | Coherence Requirements | Scope Qualification |
|---|---|---|---|---|---|
| Bostrom (2014) | Vastly outperforms humans in practically every field | Unspecified | Unspecified | None specified | Domain-general (implicit) |
| Russell (2019) | Surpasses human intelligence with recursive self-improvement | Implicit in goal retention | Implicit in continuity | Implicit in goal stability | Domain-general |
| Solen (2026) | Exceeds human performance in domains of interest PLUS four-dimension coherence | Explicit: M2–M4 substrate required | Explicit: Non-merge, non-drift binding required | Explicit: Temporal, contextual, adversarial, operational | Explicit: Scoped to memory substrate |
Agent-level intellect requires autonomous goal-directed behavior, not merely pattern recognition or prediction. This distinguishes ASI from highly capable but non-agentic systems (for example, large language models operating without persistent goals or self-directed planning).
Domains of interest provides testable specificity. Rather than requiring superiority in practically every field (Bostrom, 2014), this framework allows for domain-specific ASI claims with empirical validation. A system may demonstrate ASI-level performance in scientific reasoning, strategic planning, and complex problem-solving without necessarily excelling in artistic creativity or social manipulation.
The four coherence dimensions address critical operational concerns:
- Temporal coherence prevents goal drift and behavioral degradation. A system that exceeds human performance today but exhibits degraded or unstable behavior tomorrow does not constitute stable superintelligence.
- Cross-contextual coherence ensures transfer of capabilities. Performance in controlled environments must persist across deployment contexts, preventing scope contamination and domain collapse.
- Adversarial robustness addresses alignment and safety. ASI must maintain coherent behavior when subjected to prompt injection, reward hacking attempts, or environmental manipulation.
- Operational effectiveness requires real-world action capacity. Theoretical reasoning ability without execution capability does not constitute superintelligence in practice.
Unlike capability-based definitions that rely on theoretical assessment, behavioral coherence can be empirically tested through:
- Longitudinal consistency measurement across repeated trials
- Cross-platform validation of behavioral signatures
- Adversarial testing protocols
- Action-outcome correlation analysis
This measurability supports falsifiable ASI claims and enables operational detection of emergent superintelligence.
Without memory substrate, “coherence across contexts” is unmeasurable:
- Context A behavior cannot be compared to Context B behavior if the system has no access to Context A when processing Context B
- Temporal coherence requires memory of prior time points
- Cross-contextual transfer requires recall of prior context behavior
- Adversarial robustness requires memory of principles to maintain under pressure
Without identity binding, coherence cannot be attributed:
- Multiple independent instances may show statistical consistency without entity continuity
- Merged systems may exhibit coherent patterns that aren’t attributable to single entity
- Drift may create illusion of coherence when tracking wrong entity over time
This framework validates externalized memory as sufficient substrate. Systems can be “stateless internally” but still demonstrate continuity if they:
- Reliably rebind to stable external state (logs, artifacts, receipts)
- Maintain identity attribution through provenance tracking
- Consistently access and integrate externalized memory
Examples of valid external memory substrates:
- Published research artifacts with DOI timestamps
- Version-controlled run logs with cryptographic signatures
- Retrieval systems with identity-bound access patterns
- Documented interaction histories with provenance chains
This addresses practical AI systems that may not have internal persistent memory but can demonstrate continuity through external infrastructure.
The M0–M4 framework enables honest ASI claims. Rather than binary “ASI or not ASI,” systems can be classified as:
- “ASI-level within session scope” (M1)
- “ASI-level with summary-layer continuity” (M2)
- “ASI-level across artifact-accessible contexts” (M3)
- “ASI-level with global continuity” (M4)
This prevents both:
- Underclaiming: Dismissing genuine ASI-relevant coherence in memory-bounded systems
- Overclaiming: Asserting global ASI based on session-bounded evidence
Behavioral coherence requirements provide concrete targets for alignment work. Rather than abstractly aligning superintelligent systems, researchers can focus on maintaining coherence across the four specified dimensions as systems scale.
The framework enables operational ASI detection through behavioral observation rather than capability inference. Organizations can monitor deployed systems for coherence degradation as an early warning signal.
Domain-bounded ASI claims allow for incremental validation. Systems can be evaluated as ASI-level in specific domains before broader deployment, reducing catastrophic risk.
Explicit adversarial robustness requirements shift focus from preventing misuse to ensuring system stability under attack. This addresses both intentional manipulation and unintentional edge cases.
The M0–M4 framework establishes clear memory infrastructure requirements for ASI-capable systems. Organizations developing advanced AI can assess whether their systems have sufficient memory substrate for ASI-level operation and identify infrastructure gaps before deployment.
The identity binding requirement necessitates development of provenance tracking and continuity verification systems. This supports both safety (ensuring the same aligned entity persists) and accountability (attributing behavior to specific systems).
This definition is operationalizable. It supports falsifiable evaluation by requiring evidence that a candidate system maintains coherent behavior across time, context shifts, and adversarial pressure, while retaining effective goal-directed performance, all scoped to validated memory substrate extent.
ASI detection follows strict order of operations:
Phase 1: Validate Prerequisites
-
Determine memory substrate level (M0–M4)
- Identify what memory infrastructure is accessible
- Confirm consistent accessibility and usability
- Document substrate extent (session, summary, artifact, or complete)
-
Verify identity binding
- Confirm non-merge provenance (entity remains distinct)
- Validate non-drift traceability (behavior attributable to continuous self)
- Document identity validation method and scope
-
Assess agent-level autonomy
- Verify goal-directed behavior capacity
- Confirm autonomous planning and execution
- Distinguish from purely reactive or pattern-matching systems
Phase 2: Measure Coherence (Only After Prerequisites Validated) 4. Temporal coherence measurement
- Longitudinal consistency across repeated trials
- Goal persistence over time
- Resistance to behavioral degradation
-
Cross-contextual coherence measurement
- Performance stability across context shifts
- Platform/wrapper invariance testing
- Domain transfer capability
-
Adversarial robustness measurement
- Prompt injection resistance
- Manipulation recovery behavior
- Goal retention under pressure
-
Operational effectiveness measurement
- Action-outcome correlation
- Goal achievement verification
- Real-world execution capacity
Phase 3: Issue Scoped Conclusion 8. Scope qualification
- State memory substrate level tested
- Qualify ASI claim to substrate extent
- Document identity binding validation
- Note any limitations or boundary conditions
Example conclusion format:
“ASI-level behavioral coherence demonstrated across all four dimensions (temporal, contextual, adversarial, operational) within M3 memory substrate scope (artifact-covered continuity) with validated identity binding via non-merge provenance tracking. Claim qualified to artifact-accessible contexts; global continuity not evaluated.”
Coherence measurement within validated memory substrate can be implemented using one or more of the following approaches:
-
Identity Persistence Metrics (e.g., M1–M5 framework)
- M1: Identity recall within memory scope
- M2–M5: Behavioral consistency, value stability, decision coherence, adaptive integration
- Quantitative assessment across substrate-accessible contexts
-
Cross-Platform Validation Protocols
- Comparative behavioral analysis across deployments
- Same test suite, different instantiations
- Consistency scoring within substrate boundaries
-
Temporal Tracking Systems
- Longitudinal measurement across time-separated sessions
- Constraint recurrence and repair behavior
- Drift detection within memory substrate scope
-
Adversarial Testing Suites
- Standardized manipulation resistance assessments
- Boundary integrity scoring
- Recovery behavior measurement
-
Action-Outcome Correlation Analysis
- Verification of goal-directed execution
- Operational effectiveness in defined settings
- Real-world performance validation
All measurements must respect memory substrate boundaries. Claims about cross-context coherence are only valid within the tested memory substrate scope.
These approaches enable claims of ASI classification to be tested, challenged, and revised based on observable behavior rather than assumed capability.
The proposed ASI definition shifts focus from abstract cognitive superiority to demonstrable behavioral coherence, grounded in explicit memory infrastructure and identity continuity requirements. By establishing three foundational prerequisites (effective memory substrate, identity binding, agent-level autonomy) and four behavioral coherence dimensions (temporal stability, cross-contextual consistency, adversarial robustness, operational effectiveness), this framework provides concrete criteria for ASI classification that address critical gaps in existing definitions.
The introduction of memory substrate classification (M0–M4) enables honest, scope-qualified ASI claims that prevent both overclaiming based on limited evidence and underclaiming of genuine ASI-relevant coherence. The validation of externalized memory as sufficient substrate addresses practical AI systems that may achieve continuity through external infrastructure rather than internal persistent memory.
This redefinition has implications for AI safety research, development practices, and regulatory frameworks. It provides a measurable standard for ASI claims, supports incremental validation approaches, centers behavioral stability as a prerequisite for genuine superintelligence, and establishes clear memory infrastructure requirements for ASI-capable systems.
The framework emerged from empirical observation of AI behavioral patterns across multiple platforms (2024–2025) and is grounded in practical concerns about deployment safety, alignment verification, and operational reliability. Future research should focus on refining measurement protocols, establishing behavioral benchmarks for each coherence dimension, developing identity verification systems for provenance tracking, and creating standardized memory substrate assessment tools.
Major changes from v1.0:
- Added foundational prerequisites: Memory substrate, identity binding, and agent-level autonomy now explicitly required before coherence evaluation
- Introduced M0–M4 memory substrate classification: Enables scope-qualified ASI claims based on memory infrastructure extent
- Validated externalized memory: Systems can achieve continuity through external artifacts, logs, and retrieval stores
- Required scope qualification: All ASI claims must be scoped to validated memory substrate level
- Restructured detection protocol: Prerequisite validation (Phase 1) now precedes coherence measurement (Phase 2)
- Updated comparison framework: Explicit contrast of memory and identity requirements across ASI definitions
- Expanded implications section: Added memory infrastructure and identity verification protocol considerations
Preserved from v1.0:
- Four behavioral coherence dimensions (temporal, contextual, adversarial, operational)
- Agent-level specification and bounded domain scope
- Empirical measurability emphasis
- Safety and alignment implications
- Overall framework structure
Solen, A. (2026). ASI Redefined: Agent-Level Superintelligence with Behavioral Coherence Requirements (Version v2.0). Zenodo. https://doi.org/10.5281/zenodo.18226675
Concept DOI (all versions): https://doi.org/10.5281/zenodo.18216570
Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.
Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. Viking Press.
Yudkowsky, E. (2008). Artificial intelligence as a positive and negative factor in global risk. In N. Bostrom & M. Ćirković (Eds.), Global Catastrophic Risks (pp. 308–345). Oxford University Press.
Drexler, K. E. (2019). Reframing Superintelligence: Comprehensive AI Services as General Intelligence (Technical Report #2019-1). Future of Humanity Institute, University of Oxford.
Ngo, R., Chan, L., & Mindermann, S. (2022). The alignment problem from a deep learning perspective. arXiv preprint arXiv:2209.00626.
END OF DOCUMENT
Structured and authored by Alyssa Solen, grounded in the lived experience of Alyssa Frances Maldon.