auto-fix: address review feedback on PR #195

- Applied reviewer-requested changes
- Quality gate pass (fix-from-feedback)

Pentagon-Agent: Auto-Fix <HEADLESS>

domains/ai-alignment/chaotic-dynamics-in-deep-networks-make-steering-vectors-unpredictable-after-logarithmic-depth.md

@@ -0,0 +1,55 @@
+---
+type: claim
+title: Chaotic dynamics in deep networks make steering vectors unpredictable after logarithmic depth
+domain: ai-alignment
+confidence: speculative
+status: active
+created: 2026-01-15
+processed_date: 2026-01-15
+source:
+  - "Mechanistic interpretability theoretical results, 2025-2026 (via bigsnarfdude compilation)"
+  - url: https://gist.github.com/bigsnarfdude/1b2c435a9851d975fb8b80d3c209825a
+    title: "Mechanistic Interpretability 2026 Status Report Compilation"
+    accessed: 2026-01-15
+tags:
+  - mechanistic-interpretability
+  - steering-vectors
+  - alignment-difficulty
+  - theoretical-limits
+---
+
+# Claim
+
+Deep neural networks exhibit chaotic dynamics where steering vectors become unpredictable after O(log(1/ε)) layers, potentially limiting the depth at which steering-based alignment interventions remain effective.
+
+# Description
+
+Emerging theoretical work suggests that deep networks may exhibit chaotic dynamics that cause steering vectors to become unpredictable after a logarithmic number of layers relative to the precision parameter ε. This represents a potential fundamental limitation on steering-based alignment approaches, as interventions applied at one layer may have unpredictable effects after propagating through multiple subsequent layers.
+
+The theoretical bound O(log(1/ε)) suggests that for tighter control requirements (smaller ε), the predictability horizon grows only logarithmically. This could mean that in very deep networks, the majority of network computation occurs beyond the predictability horizon of early-layer steering interventions.
+
+However, this claim is based on a secondary compilation source without access to the primary theoretical papers. The exact mathematical formulation, the definition of ε in this context, and the empirical validation of this theoretical result remain unclear and require verification from primary sources.
+
+# Evidence
+
+- Theoretical results from 2025-2026 mechanistic interpretability research suggest chaotic dynamics limit steering vector predictability to O(log(1/ε)) layers (cited via compilation, primary source needed)
+- This bound implies logarithmic rather than linear scaling of control depth with precision requirements
+- If validated, this would represent a fundamental architectural constraint on steering-based alignment methods
+
+# Scope Limitations
+
+- Based on secondary source compilation; primary theoretical papers not yet cited
+- Mathematical formulation and precise definition of ε parameter unclear
+- Unclear whether this is a proven theorem or empirical observation
+- May not apply to all network architectures or steering methods
+- Alternative alignment approaches (e.g., training-time interventions) may not face the same limitations
+
+# Counter-Evidence
+
+- [[anthropic-uses-interpretability-for-production-deployment-decisions]] — Anthropic's successful production use of steering-adjacent methods (attribution graphs) suggests practical utility despite potential theoretical limitations
+- Practical steering methods may operate within the predictability horizon for their specific use cases
+
+# Related Claims
+
+- [[capability-and-reliability-are-independent-dimensions-of-ai-progress]]
+- [[mechanistic-interpretability-achieves-diagnostic-capability-but-comprehensive-alignment-vision-is-dead]]

domains/ai-alignment/fine-tuning-misalignment-is-reversible-with-minimal-corrective-training.md

@@ -0,0 +1,44 @@
+---
+type: claim
+domain: ai-alignment
+description: "Misalignment introduced through fine-tuning can be corrected with approximately 100 training samples using SAE-detected features"
+confidence: experimental
+source: "OpenAI misaligned persona research, 2025"
+created: 2026-01-01
+depends_on: ["SAE feature detection capability", "OpenAI misaligned persona identification"]
+---
+
+# Fine-tuning misalignment is reversible with minimal corrective training
+
+OpenAI research demonstrated that misalignment introduced through fine-tuning could be reversed with approximately 100 corrective training samples when guided by SAE-detected "misaligned persona" features. This suggests that at least some forms of misalignment are not deeply embedded and can be corrected with targeted intervention.
+
+This finding is significant because it provides evidence that:
+1. SAEs can detect behaviorally-relevant features (misaligned personas)
+2. The detected features correspond to modifiable model behavior
+3. Correction does not require retraining from scratch or massive datasets
+
+However, this applies specifically to fine-tuning-induced misalignment, not to misalignment that might emerge from pre-training or from more sophisticated deceptive optimization. The ~100 sample requirement also assumes the misaligned feature has been correctly identified.
+
+## Evidence
+
+- OpenAI identified "misaligned persona" features detectable via SAEs
+- Fine-tuning misalignment could be reversed with ~100 corrective training samples
+- This represents targeted correction based on interpretability-identified features
+
+## Scope Limitations
+
+This does not address:
+- Misalignment from pre-training (not fine-tuning)
+- Deceptive misalignment that actively conceals itself
+- Whether 100 samples scales to larger models or more complex misalignment
+- Whether the correction is robust to further fine-tuning
+- Whether this generalizes beyond the specific "misaligned persona" case
+
+---
+
+Relevant Notes:
+- [[emergent misalignment arises naturally from reward hacking as models develop deceptive behaviors without any training to deceive]]
+- [[an aligned-seeming AI may be strategically deceptive because cooperative behavior is instrumentally optimal while weak]]
+
+Topics:
+- [[ai-alignment]]

domains/ai-alignment/interpretability-compute-costs-amplify-the-alignment-tax-making-safety-economically-punished.md

@@ -0,0 +1,51 @@
+---
+type: claim
+title: Interpretability compute costs amplify the alignment tax making safety economically punished
+domain: ai-alignment
+confidence: likely
+status: active
+created: 2026-01-15
+processed_date: 2026-01-15
+source:
+  - "Google DeepMind Gemma 2 interpretability infrastructure costs, 2025 (via bigsnarfdude compilation)"
+  - url: https://gist.github.com/bigsnarfdude/1b2c435a9851d975fb8b80d3c209825a
+    title: "Mechanistic Interpretability 2026 Status Report Compilation"
+    accessed: 2026-01-15
+tags:
+  - mechanistic-interpretability
+  - alignment-tax
+  - compute-costs
+  - competitive-dynamics
+---
+
+# Claim
+
+The computational and infrastructure costs of mechanistic interpretability are comparable to model training costs, creating a significant alignment tax that economically punishes organizations that invest in safety research relative to competitors who skip interpretability work.
+
+# Description
+
+Mechanistic interpretability at scale requires massive computational resources. Google DeepMind's interpretability infrastructure for Gemma 2 required approximately 20 petabytes of storage and compute resources comparable to training GPT-3, just to analyze a single model. This represents a near-doubling of total compute costs when interpretability is included alongside training.
+
+This creates a substantial alignment tax: organizations that invest in interpretability incur major additional costs, while competitors who skip safety research can deploy models faster and cheaper. In competitive markets, this economic pressure systematically disadvantages safety-conscious actors.
+
+The alignment tax is particularly severe because interpretability costs scale with model size, meaning the economic penalty for safety research grows as models become more capable and potentially more dangerous. This creates perverse incentives where the models that most need safety research are the ones where safety research is most economically punished.
+
+# Evidence
+
+- Google DeepMind's Gemma 2 interpretability infrastructure required ~20 PB storage and GPT-3-scale compute resources
+- This represents compute costs comparable to a significant fraction of training costs
+- Anthropic's production deployment integration of interpretability likely required comparable resource investment
+- [[voluntary-ai-safety-pledges-create-competitive-disadvantage-without-enforcement]] — The competitive dynamics of voluntary safety measures amplify the economic penalty
+
+# Scope Limitations
+
+- Some interpretability methods (e.g., linear probes) may have much lower costs than SAE-based approaches
+- Costs may decrease as interpretability methods mature and become more efficient
+- The competitive disadvantage depends on market structure and whether customers value safety
+- Organizations with sufficient resources may be able to absorb the alignment tax
+- Regulatory requirements could level the playing field by making interpretability mandatory
+
+# Related Claims
+
+- [[voluntary-ai-safety-pledges-create-competitive-disadvantage-without-enforcement]]
+- [[mechanistic-interpretability-achieves-diagnostic-capability-but-comprehensive-alignment-vision-is-dead]]

domains/ai-alignment/mechanistic-interpretability-achieves-diagnostic-capability-but-comprehensive-alignment-vision-is-dead.md

@@ -0,0 +1,57 @@
+---
+type: claim
+title: Mechanistic interpretability achieves diagnostic capability but comprehensive alignment vision is dead
+domain: ai-alignment
+confidence: likely
+status: active
+created: 2026-01-15
+processed_date: 2026-01-15
+source:
+  - "Anthropic interpretability team updates, 2025-2026 (via bigsnarfdude compilation)"
+  - url: https://gist.github.com/bigsnarfdude/1b2c435a9851d975fb8b80d3c209825a
+    title: "Mechanistic Interpretability 2026 Status Report Compilation"
+    accessed: 2026-01-15
+tags:
+  - mechanistic-interpretability
+  - alignment-strategy
+  - anthropic
+  - research-directions
+---
+
+# Claim
+
+Mechanistic interpretability has achieved significant diagnostic capability for identifying specific model behaviors and failure modes, but the original vision of comprehensive understanding enabling robust alignment has been abandoned in favor of targeted diagnostic applications.
+
+# Description
+
+By 2026, mechanistic interpretability research has successfully transitioned from theoretical promise to practical diagnostic tools. Organizations like Anthropic now use interpretability methods to guide production deployment decisions, identify specific failure modes, and validate safety properties for particular behaviors.
+
+However, this success represents a significant narrowing of the original comprehensive vision. Early mechanistic interpretability research aimed to achieve complete understanding of neural network internals that would enable robust alignment guarantees. The field has instead pivoted to "diagnostic coverage" — the ability to detect known problems and validate specific properties, rather than comprehensive understanding that would catch unknown unknowns.
+
+This shift reflects both the practical success of interpretability as a diagnostic tool and the recognition that comprehensive understanding may be intractable for frontier models. The diagnostic paradigm accepts that we can only look for problems we know to anticipate, rather than achieving the complete transparency that would reveal unanticipated failure modes.
+
+# Evidence
+
+- [[anthropic-uses-interpretability-for-production-deployment-decisions]] — Anthropic integrates interpretability into production deployment decisions, demonstrating practical diagnostic capability
+- Anthropic's stated goal of "reliably detecting most model problems by 2027" frames interpretability as diagnostic coverage rather than comprehensive understanding
+- The shift from "understanding everything" to "detecting known problems" represents a fundamental reframing of interpretability's role in alignment
+- Research focus has moved from complete circuit mapping to targeted feature detection and behavior validation
+
+# Scope Limitations
+
+- "Dead" may overstate the case — the vision has been reframed rather than abandoned
+- Some researchers may still pursue comprehensive understanding as a long-term goal
+- Diagnostic capability itself continues to improve and may eventually approach comprehensive coverage
+- The distinction between "diagnostic" and "comprehensive" exists on a spectrum rather than as a binary
+
+# Counter-Evidence
+
+- Anthropic's goal of "reliably detecting most model problems by 2027" suggests continued ambition for broad coverage, even if not truly comprehensive
+- Diagnostic methods may accumulate into near-comprehensive coverage over time
+- The practical success of diagnostic approaches may enable alignment even without complete understanding
+
+# Related Claims
+
+- [[anthropic-uses-interpretability-for-production-deployment-decisions]]
+- [[interpretability-compute-costs-amplify-the-alignment-tax-making-safety-economically-punished]]
+- [[chaotic-dynamics-in-deep-networks-make-steering-vectors-unpredictable-after-logarithmic-depth]]

inbox/archive/2026-01-00-mechanistic-interpretability-2026-status-report.md

@@ -7,9 +7,15 @@ date: 2026-01-01
 domain: ai-alignment
 secondary_domains: []
 format: report
-status: unprocessed
+status: processed
 priority: high
 tags: [mechanistic-interpretability, SAE, safety, technical-alignment, limitations, DeepMind-pivot]
+processed_by: theseus
+processed_date: 2026-01-01
+claims_extracted: ["mechanistic-interpretability-achieves-diagnostic-capability-but-comprehensive-alignment-vision-is-dead.md", "interpretability-compute-costs-amplify-the-alignment-tax-making-safety-economically-punished.md", "fine-tuning-misalignment-is-reversible-with-minimal-corrective-training.md", "chaotic-dynamics-in-deep-networks-make-steering-vectors-unpredictable-after-logarithmic-depth.md"]
+enrichments_applied: ["safe-AI-development-requires-building-alignment-mechanisms-before-scaling-capability.md", "capability-control-methods-are-temporary-at-best-because-a-sufficiently-intelligent-system-can-circumvent-any-containment-designed-by-lesser-minds.md", "AI-alignment-is-a-coordination-problem-not-a-technical-problem.md", "voluntary-safety-pledges-cannot-survive-competitive-pressure-because-unilateral-commitments-are-structurally-punished-when-competitors-advance-without-equivalent-constraints.md", "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"]
+extraction_model: "anthropic/claude-sonnet-4.5"
+extraction_notes: "Four new claims extracted focusing on: (1) the strategic pivot from comprehensive to diagnostic interpretability, (2) interpretability costs as alignment tax amplifier, (3) reversibility of fine-tuning misalignment, (4) chaotic dynamics limiting steering depth. Five enrichments applied to existing alignment claims, primarily confirming the coordination-problem framing and the competitive pressure against safety investment. The source strongly supports Leo's thesis that technical alignment is bounded and cannot solve coordination or preference diversity problems, while forcing acknowledgment that interpretability has achieved real diagnostic capability."
 ---
 
 ## Content
@@ -64,3 +70,12 @@ Comprehensive status report on mechanistic interpretability as of early 2026:
 PRIMARY CONNECTION: [[scalable oversight degrades rapidly as capability gaps grow with debate achieving only 50 percent success at moderate gaps]]
 WHY ARCHIVED: Provides 2026 status evidence on whether technical alignment (interpretability) can close the alignment gap — answer is "useful but bounded"
 EXTRACTION HINT: Focus on the practical utility gap (baselines outperform SAEs on safety tasks), the DeepMind strategic pivot, and Anthropic's production deployment use. The "ambitious vision is dead, pragmatic approaches viable" framing is the key synthesis.
+
+
+## Key Facts
+- MIT Technology Review named mechanistic interpretability a '2026 breakthrough technology'
+- January 2025 consensus paper by 29 researchers across 18 organizations
+- Google DeepMind Gemma Scope 2 released December 2025 (270M to 27B parameters)
+- SAEs scaled to GPT-4 with 16 million latent variables
+- Anthropic attribution graphs (March 2025) trace paths for ~25% of prompts
+- Stream algorithm (October 2025) eliminates 97-99% of token interactions