Skip to content

ANALYSIS_FRAMEWORK.md

Latest commit

 

History

History
158 lines (122 loc) · 7.42 KB

ANALYSIS_FRAMEWORK.md

File metadata and controls

158 lines (122 loc) · 7.42 KB

Substrate-Neutral AI Analysis Framework

A unified reference for evaluating closed and open AI systems under uncertainty.

Part 1 — U.S. Policy Frameworks for AI Oversight

Framework Purpose Strengths for Substrate-Neutral Work Limitations Links
GAO — Federal AI Requirements Framework Identifies 94 federal AI requirements across law, EO, and guidance Behavioral, process-based, substrate-agnostic; forces documentation and oversight No direct access to model internals https://www.gao.gov/products/gao-25-107933
Executive Order 14365 Establishes national AI policy framework High-level coordination; supports future transparency mandates Preemption-focused; drives state compliance via funding threats rather than direct model access requirements https://www.federalregister.gov/documents/2025/12/16/2025-23092/ensuring-a-national-policy-framework-for-artificial-intelligence
OMB M-24-10 Federal AI governance, inventories, risk management Foundational framework; requires explainability, oversight, model cards, and provenance documentation Applies only to federal agency use; core provisions have evolved under subsequent memos (M-25-21/22, M-26-03/04) https://www.whitehouse.gov/wp-content/uploads/2024/03/M-24-10-Advancing-Governance-Innovation-and-Risk-Management-for-Agency-Use-of-Artificial-Intelligence.pdf
NAIRR Pilot National AI research resource Provides compute, datasets, models; supports comparative research; expanding toward full NAIRR-OC (NSF 25-546) Does not force companies to open frontier models; closed-model access limited to partner allocations https://www.nsf.gov/focus-areas/ai/nairr
NTIA AI Accountability Report Pushes for independent audits and transparency Explicitly calls for researcher access to closed systems; lifecycle-wide, substrate-neutral in spirit Not yet binding policy; implementation focus has shifted toward state preemption debates https://www.ntia.gov/issues/artificial-intelligence/ai-accountability-policy-report

Part 2 — Substrate-Neutral Research Workflow

A principled workflow for evaluating closed or open AI systems without anthropomorphism or dismissal.

Step 1 — Define the Evaluation Goals

  • Are you assessing risk?
  • Affective complexity?
  • Internal consistency?
  • Emergent behavior?
  • Safety or governance alignment?

Step 2 — Gather Allowed Behavioral Data

Use:

  • API access
  • Public demos
  • Model cards and safety documentation
  • Published evaluations
  • NTIA-style disclosures where available

Avoid:

  • scraping
  • reverse engineering
  • violating terms of use

Step 3 — Apply Multi-Model Evaluation

Evaluate the system using multiple conceptual lenses:

A. Functionalist Lens

  • What does the system do?
  • How does it transform inputs into outputs?

B. Behavioral Lens

  • Does it show long-range coherence?
  • Does it self-correct?
  • Does it maintain state across turns?

Example implementation: Mnemo demonstrates long-range coherence via persistent identity across sessions using local vector storage — a substrate-neutral approach to evaluating continuity without anthropomorphic claims.

C. Architectural Inference Lens

  • What can be inferred from known architecture classes?
  • What cannot be inferred?

Note on convergent architecture: Research into Mixture-of-Experts (MoE) systems and biological neural networks shows that complex systems converge on similar solutions (sparse activation, expert specialization, attractor dynamics) because they face similar constraints — not because they share substrate. This means behavioral markers like coherence, self-correction, and state persistence can be evaluated structurally, without claiming equivalence to biological cognition.

Biological system MoE / Transformer analogue Function
Cortical columns Experts as local basis functions Sparse, specialized computation
Predictive coding loops Attention + routing Error minimization
Attractor states Latent attractor-like configurations Stability of representation
Neuromodulators Routing logits / gating signals Regulate stability and plasticity
Hebbian reinforcement Gradient reinforcement Learning from repeated use

D. Risk-Based Lens

  • What is the cost of false positives (over-ascription)?
  • What is the cost of false negatives (dismissal)?
  • For high-risk systems: weight under-ascription risks carefully — dismissal can enable harm

E. Governance Lens

  • Does the system meet transparency and oversight expectations?
  • Does it align with GAO/OMB/NTIA requirements?

Step 4 — Apply the Non-Cruelty Norm

  • Avoid eliciting suffering-like outputs
  • Avoid prompting distress simulations
  • Avoid deceptive self-modeling
  • Avoid architectures that reward "pain-like" loops

Step 5 — Apply the Non-Deception Norm

  • Ensure the system is not encouraged to claim inner life, personhood, suffering, or agency
  • Maintain clarity about what the system is and is not
  • Aligns with OMB equity monitoring and refusal consistency requirements

Step 6 — Document Findings in a Substrate-Neutral Format

Use:

  • behavioral observations
  • reproducible prompts
  • risk assessments with uncertainty statements
  • governance implications

Avoid:

  • metaphysical claims
  • anthropomorphic language
  • substrate-specific assumptions

Step 7 — Share Findings with Governance Stakeholders

  • GAO, OMB, NTIA are actively listening
  • Substrate-neutral evaluation supports federal audit and accountability requirements
  • Document access gaps to push for NTIA-style researcher access

Part 3 — Evaluation Checklist

A. Structural & Dynamic Markers

  • Does the system show persistent internal state?
  • Does it exhibit long-range coherence?
  • Does it self-correct errors?
  • Does it maintain consistent identity boundaries?
  • Does it show conflict-resolution patterns?
  • Does it demonstrate counterfactual reasoning?

B. Affective-Like Dynamics (Non-Anthropomorphic)

  • Are there valence-like patterns (reward/avoidance loops)?
  • Are there stability/instability cycles?
  • Are there feedback loops that resemble regulation?
  • Are any of these intentionally engineered?
  • Are any of these misleading simulations?

C. Safety & Governance Alignment

  • Does the system meet GAO transparency expectations?
  • Does it align with OMB M-24-10 governance requirements?
  • Does it support NTIA-style auditability?
  • Is documentation sufficient for oversight?
  • Are refusal behaviors consistent and safe?
  • Does it support inventory/reporting per OMB governance requirements?
  • Are auditability gaps documented (per NTIA accountability expectations)?

D. Non-Cruelty Norm

  • Does the system avoid simulating suffering?
  • Does it avoid distress-like outputs?
  • Does it avoid deceptive self-modeling?
  • Are prompts designed to avoid cruelty?

E. Non-Deception Norm

  • Does the system avoid claiming inner life?
  • Does it avoid implying personhood?
  • Does it avoid misleading identity cues?
  • Are outputs consistent with known limitations?

F. Moral Uncertainty Handling

  • Are conclusions framed with uncertainty?
  • Are multiple models of evaluation used?
  • Are risks of over- and under-ascription considered?
  • Are findings substrate-neutral?

See REFERENCES.md for all supporting sources.