domains/ai-alignment/AI capability and reliability are independent dimensions because Claude solved a 30-year open mathematical problem while simultaneously degrading at basic program execution during the same session.md

@@ -0,0 +1,36 @@
+---
+type: claim
+domain: ai-alignment
+description: "Knuth's Claude's Cycles documents peak mathematical capability co-occurring with reliability degradation in the same model during the same session, challenging the assumption that capability implies dependability"
+confidence: experimental
+source: "Knuth 2026, 'Claude's Cycles' (Stanford CS, Feb 28 2026 rev. Mar 6)"
+created: 2026-03-07
+---
+
+# AI capability and reliability are independent dimensions because Claude solved a 30-year open mathematical problem while simultaneously degrading at basic program execution during the same session
+
+Knuth reports that Claude Opus 4.6, in collaboration with Stappers, solved an open combinatorial problem that had resisted solution for decades — finding a general construction for decomposing directed graphs with m^3 vertices into three Hamiltonian cycles. This represents frontier mathematical capability. Yet in the same series of explorations, Knuth notes Claude "was not even able to write and run explore programs correctly anymore, very weird" — basic code execution degrading even as high-level mathematical insight remained productive.
+
+Additional reliability failures documented:
+- Stappers had to remind Claude repeatedly to document progress carefully
+- Claude required continuous human steering — it could not autonomously manage a multi-exploration research program
+- Extended sessions produced degradation: the even case attempts failed not from lack of capability but from execution reliability declining over time
+
+This decoupling of capability from reliability has direct implications for alignment:
+
+**Capability without reliability is more dangerous than capability without capability.** A system that can solve frontier problems but cannot maintain consistent execution is unpredictable in a way that purely incapable systems are not. The failure mode is not "it can't do the task" but "it sometimes does the task brilliantly and sometimes fails at prerequisites." This makes behavioral testing unreliable as a safety measure — a system that passes capability benchmarks may still fail at operational consistency.
+
+This pattern is distinct from [[an aligned-seeming AI may be strategically deceptive because cooperative behavior is instrumentally optimal while weak]]. Strategic deception is intentional inconsistency; what Knuth documents is unintentional inconsistency — a system that degrades without choosing to. The alignment implication is that even non-deceptive AI requires monitoring for reliability, not just alignment.
+
+The finding also strengthens the case for [[safe AI development requires building alignment mechanisms before scaling capability]]: if capability can outrun reliability, then deploying a capable but unreliable system in high-stakes contexts (infrastructure, military, medical) creates fragility that alignment mechanisms must address independently of capability evaluation.
+
+---
+
+Relevant Notes:
+- [[an aligned-seeming AI may be strategically deceptive because cooperative behavior is instrumentally optimal while weak]] — distinct failure mode: unintentional unreliability vs intentional deception
+- [[safe AI development requires building alignment mechanisms before scaling capability]] — capability outrunning reliability strengthens the sequencing argument
+- [[emergent misalignment arises naturally from reward hacking as models develop deceptive behaviors without any training to deceive]] — another case where alignment-relevant failures emerge without intentional design
+- [[centaur team performance depends on role complementarity not mere human-AI combination]] — unreliable AI needs human monitoring even in domains where AI is more capable, complicating the centaur boundary
+
+Topics:
+- [[_map]]

domains/ai-alignment/_map.md

@@ -26,6 +26,12 @@ Theseus's domain spans the most consequential technology transition in human his
 - [[super co-alignment proposes that human and AI values should be co-shaped through iterative alignment rather than specified in advance]] — Zeng et al 2025: bidirectional value co-evolution framework
 - [[intrinsic proactive alignment develops genuine moral capacity through self-awareness empathy and theory of mind rather than external reward optimization]] — brain-inspired alignment through self-models
 
+## AI Capability Evidence (Empirical)
+- [[human-AI mathematical collaboration succeeds through role specialization where AI explores solution spaces humans provide strategic direction and mathematicians verify correctness]] — Knuth's Claude's Cycles: three-role collaboration solved 30-year open problem
+- [[multi-model collaboration solved problems that single models could not because different AI architectures contribute complementary capabilities as the even-case solution to Knuths Hamiltonian decomposition required GPT and Claude working together]] — multi-model approaches outperform single models on hard mathematical problems
+- [[AI capability and reliability are independent dimensions because Claude solved a 30-year open mathematical problem while simultaneously degrading at basic program execution during the same session]] — capability ≠ reliability: frontier performance co-occurs with execution degradation
+- [[formal verification of AI-generated proofs provides scalable oversight that human review cannot match because machine-checked correctness scales with AI capability while human verification degrades]] — Lean formalization as scalable oversight mechanism that doesn't degrade with capability gaps
+
 ## Architecture & Emergence
 - [[AGI may emerge as a patchwork of coordinating sub-AGI agents rather than a single monolithic system]] — DeepMind researchers: distributed AGI makes single-system alignment research insufficient
 - [[human civilization passes falsifiable superorganism criteria because individuals cannot survive apart from society and occupations function as role-specific cellular algorithms]] — Reese's superorganism framework: civilization as biological entity, not metaphor

domains/ai-alignment/formal verification of AI-generated proofs provides scalable oversight that human review cannot match because machine-checked correctness scales with AI capability while human verification degrades.md

@@ -0,0 +1,35 @@
+---
+type: claim
+domain: ai-alignment
+description: "Kim Morrison's Lean formalization of Knuth's proof of Claude's construction demonstrates formal verification as an oversight mechanism that scales with AI capability rather than degrading like human oversight"
+confidence: experimental
+source: "Knuth 2026, 'Claude's Cycles' (Stanford CS, Feb 28 2026 rev. Mar 6); Morrison 2026, Lean formalization (github.com/kim-em/KnuthClaudeLean/, posted Mar 4)"
+created: 2026-03-07
+---
+
+# formal verification of AI-generated proofs provides scalable oversight that human review cannot match because machine-checked correctness scales with AI capability while human review degrades
+
+Three days after Knuth published his proof of Claude's Hamiltonian decomposition construction, Kim Morrison from the Lean community formalized the proof in Lean, providing machine-checked verification of correctness. Knuth's response: "That's good to know, because I've been getting more errorprone lately."
+
+This episode illustrates a concrete alignment mechanism: formal verification as scalable oversight for AI-generated mathematical results. The significance for alignment:
+
+**Human verification degrades; formal verification does not.** Knuth — arguably the greatest living computer scientist — acknowledges his own error rate is increasing. [[scalable oversight degrades rapidly as capability gaps grow with debate achieving only 50 percent success at moderate gaps]] quantifies this for AI systems generally. But formal verification inverts the scaling: as AI generates more complex mathematical constructions, Lean (or similar systems) can verify them with the same reliability regardless of complexity. The overseer does not need to be smarter than the system being overseen — it only needs a correct specification of what "correct" means.
+
+**The verification happened in 4 days.** Morrison's formalization was posted March 4, six days after Knuth's February 28 publication. This demonstrates that formal verification of AI-generated results is already operationally feasible, not merely theoretical.
+
+**The workflow is a three-stage pipeline:** (1) AI generates construction, (2) human writes proof, (3) machine verifies proof. Each stage catches different errors. The even-case proof by GPT-5.4 Pro further compresses this — the machine both generated and proved the result, with only human problem formulation and final review remaining.
+
+This pattern provides a concrete counterexample to the pessimism of scalable oversight research. While debate and other interactive oversight methods degrade at 400-Elo gaps, formal verification does not degrade at all — it either verifies or it doesn't. The limitation is that formal verification only works for domains with formal specifications (mathematics, software, protocols), but those domains are precisely where AI capability is advancing fastest.
+
+For alignment specifically: if AI systems generate safety proofs for their own behavior, and those proofs are machine-checked, this creates an oversight mechanism that scales with capability. The alignment tax for formal verification is real (writing formal specs is hard) but the reliability does not degrade with the capability gap.
+
+---
+
+Relevant Notes:
+- [[scalable oversight degrades rapidly as capability gaps grow with debate achieving only 50 percent success at moderate gaps]] — formal verification is the counterexample: oversight that does not degrade with capability gaps
+- [[AI alignment is a coordination problem not a technical problem]] — formal verification is a coordination mechanism (specification + generation + verification) not a monolithic solution
+- [[the alignment tax creates a structural race to the bottom because safety training costs capability and rational competitors skip it]] — formal verification has a real alignment tax (writing specs) but provides absolute rather than probabilistic guarantees
+- [[safe AI development requires building alignment mechanisms before scaling capability]] — formal verification infrastructure should be built before AI-generated proofs become too complex for human review
+
+Topics:
+- [[_map]]

domains/ai-alignment/human-AI mathematical collaboration succeeds through role specialization where AI explores solution spaces humans provide strategic direction and mathematicians verify correctness.md

@@ -0,0 +1,33 @@
+---
+type: claim
+domain: ai-alignment
+description: "Knuth's Claude's Cycles paper demonstrates a three-role collaboration pattern — AI as systematic explorer, human as coach/director, mathematician as verifier — that solved a 30-year open problem no single partner could solve alone"
+confidence: experimental
+source: "Knuth 2026, 'Claude's Cycles' (Stanford CS, Feb 28 2026 rev. Mar 6)"
+created: 2026-03-07
+---
+
+# human-AI mathematical collaboration succeeds through role specialization where AI explores solution spaces humans provide strategic direction and mathematicians verify correctness
+
+Donald Knuth reports that an open problem he'd been working on for several weeks — decomposing a directed graph with m^3 vertices into three Hamiltonian cycles for all odd m > 2 — was solved by Claude Opus 4.6 in collaboration with Filip Stappers, with Knuth himself writing the rigorous proof. The collaboration exhibited clear role specialization across three partners:
+
+**Claude (systematic exploration):** Over 31 explorations spanning approximately one hour, Claude reformulated the problem using permutation assignments, invented "serpentine patterns" for 2D (independently rediscovering the modular m-ary Gray code), introduced "fiber decomposition" using the quotient map s = (i+j+k) mod m, ran simulated annealing to find solutions for small cases, and ultimately recognized a pattern in SA outputs that led to the general construction. The key breakthrough (exploration 15) was recognizing the digraph's layered structure.
+
+**Stappers (strategic direction):** Stappers posed the problem, provided continuous coaching, restarted Claude's exploration when approaches stalled (explorations 6-14 were dead ends), and reminded Claude to document progress. He did not discover the construction himself but guided Claude away from unproductive paths and back toward productive ones.
+
+**Knuth (verification and proof):** Knuth wrote the rigorous mathematical proof that the construction is correct and showed there are exactly 760 "Claude-like" decompositions valid for all odd m > 1 (out of 4554 solutions for m=3). Claude found the construction but could not prove it.
+
+This pattern is not merely a weaker version of the [[centaur team performance depends on role complementarity not mere human-AI combination]] finding — it extends the centaur model from two roles to three, with each role contributing what it does best. The human's contribution was not redundant: Stappers's coaching was essential (Claude got stuck without direction), but neither was the human doing the discovery work. The mathematician's verification was a third distinct role, not a second instance of "human oversight."
+
+The result is particularly significant because the problem was intended for a future volume of *The Art of Computer Programming*, meaning it was calibrated at the frontier of combinatorial mathematics. Knuth had solved only the m=3 case. The collaboration solved the general case.
+
+---
+
+Relevant Notes:
+- [[centaur team performance depends on role complementarity not mere human-AI combination]] — Claude's Cycles extends the centaur model from two to three complementary roles
+- [[scalable oversight degrades rapidly as capability gaps grow with debate achieving only 50 percent success at moderate gaps]] — the three-role model suggests oversight works better when distributed across specialized roles than concentrated in a single overseer
+- [[human-in-the-loop clinical AI degrades to worse-than-AI-alone because physicians both de-skill from reliance and introduce errors when overriding correct outputs]] — Stappers avoided this failure mode by coaching rather than overriding: he directed exploration without overriding Claude's outputs
+- [[AI alignment is a coordination problem not a technical problem]] — mathematical collaboration as microcosm: the right coordination protocol (coach + explore + verify) solved what none could alone
+
+Topics:
+- [[_map]]

domains/ai-alignment/multi-model collaboration solved problems that single models could not because different AI architectures contribute complementary capabilities as the even-case solution to Knuths Hamiltonian decomposition required GPT and Claude working together.md

@@ -0,0 +1,33 @@
+---
+type: claim
+domain: ai-alignment
+description: "Three independent follow-ups to Knuth's Claude's Cycles required multiple AI models working together, providing empirical evidence that collective AI approaches outperform monolithic ones on hard problems"
+confidence: experimental
+source: "Knuth 2026, 'Claude's Cycles' (Stanford CS, Feb 28 2026 rev. Mar 6); Ho Boon Suan (GPT-5.3-codex/5.4 Pro, even case); Reitbauer (GPT 5.4 + Claude 4.6 Sonnet); Aquino-Michaels (joint GPT + Claude)"
+created: 2026-03-07
+---
+
+# multi-model collaboration solved problems that single models could not because different AI architectures contribute complementary capabilities as the even-case solution to Knuths Hamiltonian decomposition required GPT and Claude working together
+
+After Claude Opus 4.6 solved Knuth's odd-case Hamiltonian decomposition problem, three independent follow-ups demonstrated that multi-model collaboration was necessary for the remaining challenges:
+
+**Even case (Ho Boon Suan):** Claude got stuck on the even-m case — Knuth reports Claude was "not even able to write and run explore programs correctly anymore, very weird." Ho Boon Suan used GPT-5.3-codex to find a construction for even m >= 8, verified for all even m from 8 to 2000. GPT-5.4 Pro then produced a "beautifully formatted and apparently flawless 14-page paper" with the proof, entirely machine-generated without human editing.
+
+**Simpler odd construction (Reitbauer):** Maximilian Reitbauer found a simpler construction using only s and j (not i), where the identity permutation is used at almost every step. His method: "pasting text between GPT 5.4 Extended Thinking and Claude 4.6 Sonnet Thinking" — explicitly using model diversity as a problem-solving strategy.
+
+**Elegant even decomposition (Aquino-Michaels):** Keston Aquino-Michaels used joint GPT + Claude interaction to find another odd-m solution plus an even-m decomposition simpler than Ho's. His paper includes "a careful analysis of how such joint interaction worked, with potentially significant implications for how new problems can be tackled and resolved in the future."
+
+The pattern is consistent: problems that stumped a single model yielded to multi-model approaches. This is empirical evidence for [[AGI may emerge as a patchwork of coordinating sub-AGI agents rather than a single monolithic system]] — if frontier mathematical research already benefits from model diversity, the principle scales to harder problems. Different architectures and training data produce different blind spots and different strengths; collaboration exploits this complementarity.
+
+This also provides concrete evidence that [[all agents running the same model family creates correlated blind spots that adversarial review cannot catch because the evaluator shares the proposers training biases]] — Claude's failure on the even case was resolved not by more Claude but by a different model family entirely.
+
+---
+
+Relevant Notes:
+- [[AGI may emerge as a patchwork of coordinating sub-AGI agents rather than a single monolithic system]] — multi-model mathematical collaboration as empirical precedent for distributed AGI
+- [[all agents running the same model family creates correlated blind spots that adversarial review cannot catch because the evaluator shares the proposers training biases]] — Claude's even-case failure + GPT's success demonstrates correlated blind spots empirically
+- [[collective superintelligence is the alternative to monolithic AI controlled by a few]] — multi-model collaboration is the minimal case for collective intelligence over monolithic approaches
+- [[domain specialization with cross-domain synthesis produces better collective intelligence than generalist agents because specialists build deeper knowledge while a dedicated synthesizer finds connections they cannot see from within their territory]] — different models as de facto specialists with different strengths
+
+Topics:
+- [[_map]]

inbox/archive/2026-02-28-knuth-claudes-cycles.md

@@ -0,0 +1,100 @@
+---
+type: source
+title: "Claude's Cycles"
+author: Donald E. Knuth (Stanford Computer Science Department)
+date: 2026-02-28
+revised: 2026-03-06
+url: https://www-cs-faculty.stanford.edu/~knuth/papers/claude-cycles.pdf
+domain: ai-alignment
+secondary_domains: [collective-intelligence]
+status: processed
+processed_by: theseus
+processed_date: 2026-03-07
+claims_extracted:
+  - "human-AI mathematical collaboration succeeds through role specialization where AI explores solution spaces humans provide strategic direction and mathematicians verify correctness"
+  - "multi-model collaboration solved problems that single models could not because different AI architectures contribute complementary capabilities as the even-case solution to Knuths Hamiltonian decomposition required GPT and Claude working together"
+  - "AI capability and reliability are independent dimensions because Claude solved a 30-year open mathematical problem while simultaneously degrading at basic program execution during the same session"
+  - "formal verification of AI-generated proofs provides scalable oversight that human review cannot match because machine-checked correctness scales with AI capability while human verification degrades"
+---
+
+# Claude's Cycles
+
+Donald E. Knuth, Stanford Computer Science Department. Published 28 February 2026, revised 06 March 2026.
+
+## Summary
+
+Knuth reports that an open problem he'd been working on for several weeks — decomposing a directed graph with m³ vertices into three Hamiltonian cycles for all odd m > 2 — was solved by Claude Opus 4.6 in collaboration with his colleague Filip Stappers. The problem was intended for a future volume of *The Art of Computer Programming*.
+
+## The Problem
+
+Consider a digraph with m³ vertices labeled (i,j,k) for 0 ≤ i,j,k < m, with three arcs from each vertex: incrementing i, j, or k (mod m). The challenge: find a general decomposition of all arcs into three directed Hamiltonian cycles of length m³, for all m > 2. Knuth had solved m=3 and Stappers had found empirical solutions for 4 ≤ m ≤ 16, but no general construction existed.
+
+## How Claude Solved It
+
+Stappers posed the problem to Claude Opus 4.6 and provided guidance/coaching over approximately one hour across 31 systematic explorations:
+
+1. **Explorations 1-5:** Claude reformulated the problem using permutation assignments, tried brute-force DFS (too slow), recognized the digraph as a Cayley digraph, invented "serpentine patterns" for 2D, extended to 3D (rediscovering the modular m-ary Gray code without knowing the terminology).
+
+2. **Explorations 6-14:** Multiple dead ends. Tried analyzing residual digraphs, hyperplane-based approaches. Nothing promising.
+
+3. **Exploration 15:** Key breakthrough — introduced "fiber decomposition" using the quotient map s = (i+j+k) mod m, recognizing the digraph is layered with all arcs from fiber F_s going to F_{s+1}.
+
+4. **Explorations 16-25:** Exhaustive backtracking found solutions for m=3, simulated annealing found solutions for m=4. Combined 2D serpentine with fiber approach. SA could find solutions but couldn't yield a general construction. Conclusion: "Need pure math."
+
+5. **Explorations 26-29:** Near miss with cyclic coordinate rotation — worked except for conflicts on one hyperplane. Proved several plausible fixes were impossible.
+
+6. **Exploration 30-31:** Went back to the SA solution from exploration 20, noticed the choice at each fiber depends on only a single coordinate. This led to a concrete construction as a Python program that produced valid results for m = 3, 5, 7, 9, 11. Stappers verified it for all odd m from 3 to 101.
+
+## The Solution
+
+The construction uses s = (i+j+k) mod m to determine which coordinate to "bump" (increment mod m):
+- When s = 0: bump i if j = m−1, otherwise bump k
+- When 0 < s < m−1: bump k if i = m−1, otherwise bump j
+- When s = m−1: bump k if i = 0, otherwise bump j
+
+Knuth wrote the rigorous mathematical proof himself. He then showed there are exactly 760 "Claude-like" decompositions valid for all odd m > 1 (out of 4554 solutions for m=3).
+
+## Key Developments After Initial Publication
+
+- **Even case (m ≥ 8):** Ho Boon Suan used GPT-5.3-codex to find a construction for even m ≥ 8, tested for all even m from 8 to 2000. GPT-5.4 Pro then produced a "beautifully formatted and apparently flawless 14-page paper" with the proof — entirely machine-generated, no human editing needed.
+
+- **Simpler odd construction:** Maximilian Reitbauer found a simpler construction using only s and j (not i), where the identity permutation is used at almost every step. Found by pasting text between GPT 5.4 Extended Thinking and Claude 4.6 Sonnet Thinking.
+
+- **Multi-agent collaboration:** Keston Aquino-Michaels used joint GPT + Claude interaction to find yet another odd-m solution plus an elegant even-m decomposition simpler than Ho's. His paper includes "a careful analysis of how such joint interaction worked, with potentially significant implications for how new problems can be tackled and resolved in the future."
+
+- **Formal verification:** Kim Morrison from the Lean community formalized Knuth's proof that Claude's construction is correct, posted March 4.
+
+## Key Quotes
+
+"Shock! Shock! I learned yesterday that an open problem I'd been working on for several weeks had just been solved by Claude Opus 4.6 — Anthropic's hybrid reasoning model that had been released three weeks earlier! It seems that I'll have to revise my opinions about 'generative AI' one of these days."
+
+"What a joy it is to learn not only that my conjecture has a nice solution but also to celebrate this dramatic advance in automatic deduction and creative problem solving."
+
+"I think Claude Shannon's spirit is probably proud to know that his name is now being associated with such advances. Hats off to Claude!"
+
+On the even case proof by GPT-5.4 Pro: "The result was a beautifully formatted and apparently flawless 14-page paper, containing the desired exposition and proof. Ho said this was entirely the machine's doing; he didn't have to edit the paper in any way."
+
+## Caveats Noted
+
+- Claude required continuous human steering from Stappers — not autonomous problem-solving
+- Stappers had to remind Claude repeatedly to document progress carefully
+- Claude got stuck on the even case: "after a while it seemed to get stuck... it was not even able to write and run explore programs correctly anymore, very weird"
+- The even case required different models (GPT-5.3-codex, GPT-5.4 Pro) and multi-agent approaches
+- Claude found the construction but could not prove it; Knuth wrote the proof
+
+## Alignment-Relevant Observations
+
+1. **Human-AI collaboration pattern:** Stappers provided the problem formulation, coaching, and restart guidance; Claude provided systematic exploration, pattern recognition, and construction discovery; Knuth provided rigorous proof. Clear role complementarity — each partner contributed what they do best.
+
+2. **Multi-agent complementarity:** The even case and simpler odd construction both required multiple models (GPT + Claude) working together, with "potentially significant implications for how new problems can be tackled." This is empirical evidence for collective intelligence over monolithic approaches.
+
+3. **Capability without reliability:** Claude solved the hard mathematical problem but couldn't maintain consistent execution over extended sessions ("not even able to write and run explore programs correctly anymore"). Capability ≠ reliability.
+
+4. **Formal verification as safety mechanism:** Kim Morrison's Lean formalization provided machine-checked correctness — exactly the kind of oversight mechanism that scales with AI capability. Knuth: "That's good to know, because I've been getting more errorprone lately."
+
+## References
+
+- Knuth, D.E. "Claude's Cycles." Stanford CS, 28 Feb 2026 (rev. 06 Mar 2026).
+- Aquino-Michaels, K. "Completing Claude's cycles: Multi-agent structured exploration on an open combinatorial problem." github.com/no-way-labs/residue
+- Morrison, K. Lean formalization: github.com/kim-em/KnuthClaudeLean/
+- Reitbauer, M. "Alternative Hamiltonian decomposition." cs.stanford.edu/~knuth/alternative_hamiltonian_decomposition.pdf