Bicameral AI Architecture: Fusing Long-Context Holographic Strategy Models with Local Deterministic Execution Agents: chapter answers
Three popular beliefs promise fewer errors from large language models: use a bigger and stronger model; let two models vote and release only what they agree on; let the model reflect on and correct itself. This book refutes their shared premise, that errors are independent. After checking the literature it states its thesis as follows: the gain from combining two heterogeneous models comes from the lack of correlation between their errors, so heterogeneity is the crux and duplication is not; a long-context strategy model deliberates, while a local deterministic execution engine uses gates, te…
DOI pending · monograph page goes live after publication · 5DL-B-410 · 中文版
- Chapter 1 · Correlated Errors: Where the Second Model's Gain Comes From
Chapter 1 uses a Gaussian copula to show that, at illustrative error rates, the second model's gain is 0.0880 at a latent correlation of 0 and only 0.0243 at 0.9, and that a weaker but different reviewer leaks only 0.49… - Chapter 2 · Diversity Measures and the Ambiguity Decomposition: How Different Is Different Enough
Chapter 2 shows that binary correlations on a ledger fall far below the latent correlation, and proves by enumeration that diversity and voting accuracy are not monotonic: majority accuracy can range from 0.4 to 0.9. - Chapter 3 · The Three-Stage Pipeline: Turning Errors into Intercepted Errors
Chapter 3 shows that a common-factor correlation of only 0.1 makes a three-stage pipeline leak 1.87 times the independent product, and 8.65 times at 0.5. - Chapter 4 · The Intent Translation Protocol: Slots, Ambiguity Entropy and Upper Bounds
Chapter 4 counts ambiguity with a five-slot intent packet and its entropy, and rewrites "zero ambiguity" as an audit bound: 299 clean packets in a row are needed before the ambiguity rate can be stated below 0.010. - Chapter 5 · State-Sync Packets: Hash Chains, Version Vectors and Idempotency Keys
Chapter 5 builds state-sync packets with hash chains, version vectors and idempotency keys; when 10,000 instructions are delivered 849 extra times, the keys keep side effects at exactly 10,000. - Chapter 6 · Two-Way Correction: Gain, Noise and the Limits of Self-Correction
Chapter 6 gives the stable interval of the correction gain and shows how it narrows when receipts arrive a round late; self-reflection without an outside signal pulls accuracy towards 0.6250, receipt-driven revision tow… - Chapter 7 · Eastern and Western Classical Sources: Temple Reckoning, Breadth with Precision, and Friction
Block A is the arithmetic behind the remark that the simplest thing is difficult. A per-step success rate of 0.98 sounds reassuring. After twenty steps the whole order succeeds with probability only 0.6676, and after fo… - Chapter 8 · The Empirical Loop: Field Tools and Checker Specifications
Chapter 8 gives the bicameral ledger, the distilled Bicameral Error-Orthogonality Gain and State-Sync Packet Checker, a checker specification table grouped by kit, and a pre-registered test: lowering the leak from 0.030…
Illustrative model calculation, not a live quote