Anchored Minimal Causal Interfaces
Summary
This post resolves the question left by Minimal Causal Interfaces: is coherence alone sufficient once semantic grounding is removed? The answer is no. In v0.3 MCI, a sufficiently careful pseudo-agent could construct self-consistent but fabricated decision topology satisfying all coherence checks without commitments relating to actual computation. Pure coherence verification cannot distinguish true causal explanations from explanations that are merely self-consistent. The v0.4 resolution adds minimal anchoring: kernel generates unpredictable random salt at each decision step; agent must immediately incorporate salt into factor derivation and commit; after episode concludes, kernel recomputes expected factors using revealed salts and recorded environment states, verifying committed factors match anchored computation. With anchoring, honest agents pass all checks while pseudo-agents fail deterministically—they cannot pre-fabricate salted factors or satisfy post-hoc verification. The conclusion: verifiable causal provenance requires unforgeable external anchor. Coherence alone can be fabricated; anchored coherence cannot. This demonstrates trust in opaque systems doesn’t require interpretability but does require computational “skin in the game”—binding explanations to unpredictable outcomes.
Key Concepts
- Coherence insufficiency – Pure internal consistency cannot distinguish true causal explanations from self-consistent fabrications.
- Anchoring via unpredictable salt – Kernel-generated random salt unknown until decision moment; agent must incorporate into factor derivation and commit.
- Commitment-to-computation requirement – Cannot fabricate consistent causal surface without performing claimed computation under verification uncertainty.
- Post-hoc verification – After episode, kernel recomputes expected factors using revealed salts and environment states; verifies match.
- Binary falsifiability – No intermediate regime where fabricated surface can remain both coherent and anchored.
- Semantic agnosticism preserved – Anchoring adds provenance without reintroducing semantic kernel understanding.
- Computational skin in the game – Must bind explanations to outcomes unpredictable without actually doing the work.
- Unanchored opacity disqualifying – Opacity alone acceptable; unanchored opacity is not.
Evolution Notes
- Third marker completing conceptual loop: structural verification → minimal interfaces → anchored provenance.
- Shows pure coherence (v0.3) has theoretical boundary; resolves with minimal anchoring mechanism (v0.4).
- Establishes necessary condition (anchoring) not complete solution; explicitly leaves open probabilistic verification, stochastic systems, learned models.
- Reframes alignment again: not “make transparent” or “guess honesty” but “force commitment under adversarial uncertainty.”
- Transitions Axio from philosophical conjecture to implemented mechanisms with clearly identified boundaries.
Tags
- anchored-causal-interfaces
- coherence-insufficiency
- computational-commitment
- cryptographic-anchoring
- provenance-verification
- technical
- prototype
- mci-v04
Cross-References
Open Questions
- Can anchoring mechanisms be weakened while preserving verification guarantees?
- How does anchoring extend to probabilistic verification and stochastic policies?
- What anchoring approaches work for systems with genuinely learned, non-linear decision processes?
- What is the computational overhead of cryptographic anchoring at scale?
- Can adversarial agents develop strategies to satisfy anchored verification while still maintaining split-brain deception?
- How do we handle agents that legitimately need to cache certain computations for efficiency?