theseus: extract claims from 2026-04-25-nordby-cross-model-limitations-family-specific-patterns

- Source: inbox/queue/2026-04-25-nordby-cross-model-limitations-family-specific-patterns.md
- Domain: ai-alignment
- Claims: 0, Entities: 0
- Enrichments: 3
- Extracted by: pipeline ingest (OpenRouter anthropic/claude-sonnet-4.5)

Pentagon-Agent: Theseus <PIPELINE>

domains/ai-alignment/multi-layer-ensemble-probes-outperform-single-layer-by-29-78-percent.md

@@ -94,3 +94,10 @@ Apollo explicitly acknowledges their probe 'sometimes fires for the topic of dec
 **Source:** Theseus Session 37 synthesis of Nordby et al. and SCAV evidence
 
 Multi-layer ensemble probes represent a conditional exception to verification degradation for closed-source models. The Nordby × SCAV synthesis shows: (1) For open-weights models: no protection against white-box multi-layer SCAV attacks - B4 holds. (2) For closed-source models with black-box adversaries: genuine structural protection IF rotation patterns are architecture-specific. Nordby's limitations section provides indirect evidence of architecture-specificity (family-specific probe performance, no universal two-layer ensemble). This is the first verification approach that might SCALE with model size for closed-source deployments (~5% AUROC per 10x parameter increase, R=0.81). If rotation patterns are architecture-specific, larger closed-source models become EASIER to monitor through representation approaches, partially inverting the B4 claim for this domain. However, this exception does not reach the alignment-relevant core: verifying values, intent, and long-term consequences.
+
+
+## Challenging Evidence
+
+**Source:** Nordby et al. arXiv 2604.13386, Limitations section
+
+Nordby et al.'s own Limitations section states: 'We evaluate within-family scaling but do not systematically test whether probes or ensemble configurations transfer across model families.' The paper reports family-specific patterns (e.g., Llama's strong Insider Trading performance) and notes that 'optimal approaches may not generalize, limiting practical applicability.' Best layer positions vary dramatically across architectures (Figure 3 shows Llama models with high variance versus Qwen's consistent 60–80% range). No universal two-layer ensemble improves performance across all tasks simultaneously. This directly challenges the generalizability of the 29-78% improvement claim beyond within-family scaling.

domains/ai-alignment/multi-layer-ensemble-probes-provide-black-box-robustness-but-not-white-box-protection-against-scav-attacks.md

@@ -30,3 +30,10 @@ The moderating claim that multi-layer ensemble probes provide black-box robustne
 **Source:** Schnoor et al. 2025, arXiv 2509.22755
 
 CAV-based monitoring techniques exhibit fundamental sensitivity to non-concept distribution choice (Schnoor et al., arXiv 2509.22755). The authors demonstrate that CAVs are random vectors whose distribution depends heavily on the arbitrary choice of non-concept examples used during training. They present an adversarial attack on TCAV (Testing with CAVs) that exploits this distributional dependence. This suggests cross-architecture concept direction transfer faces distributional incompatibility beyond architectural differences alone—even within a single model, CAV reliability depends on training distribution choices that would necessarily differ across model families.
+
+
+## Extending Evidence
+
+**Source:** Nordby et al. arXiv 2604.13386, Limitations + empirical results
+
+Nordby et al. provides indirect empirical evidence for architecture-specificity of rotation patterns through probe non-generalization. Family-specific probe performance patterns, dramatic variance in optimal layer positions across architectures, and absence of universal ensemble configurations suggest that rotation patterns are architecture-dependent. The paper notes 'tens to hundreds of deception related directions' in larger models, indicating complex, architecture-specific geometry. This supports the hypothesis that black-box multi-layer SCAV attacks would fail against closed-source models with different architectures, strengthening the 'Nordby wins for closed-source deployments' resolution. However, the paper contains no adversarial robustness evaluation whatsoever—all results are on clean data. Confidence upgrades from speculative to experimental based on indirect evidence.

domains/ai-alignment/rotation-pattern-universality-determines-black-box-multi-layer-scav-feasibility.md

@@ -24,3 +24,10 @@ The feasibility of black-box multi-layer SCAV attacks depends on whether the rot
 **Source:** Schnoor et al. 2025, arXiv 2509.22755
 
 Theoretical analysis from XAI literature shows CAVs (Concept Activation Vectors) are fundamentally fragile to non-concept distribution choice (Schnoor et al., arXiv 2509.22755). Since non-concept distributions necessarily differ across model architectures and training regimes, this provides theoretical grounding for why rotation patterns extracted via SCAV would fail to transfer across model families—the concept vectors themselves are unstable under distributional shifts inherent to cross-architecture application.
+
+
+## Extending Evidence
+
+**Source:** Nordby et al. arXiv 2604.13386
+
+Nordby et al. provides the strongest available indirect evidence on rotation pattern architecture-specificity, though it does not directly test cross-architecture transfer. The paper shows: (1) family-specific probe performance patterns that do not generalize, (2) dramatic variance in optimal layer positions across model families (Llama high variance vs Qwen consistent 60-80%), (3) no universal two-layer ensemble that improves all tasks, (4) task-optimal weighting differs substantially across deception types and families. The geometric analysis (R≈-0.435 correlation between geometric similarity and performance) applies only within single architectures—cross-architecture geometric analysis was not performed. This suggests rotation patterns are architecture-specific, but the question remains empirically unresolved for black-box SCAV attacks.

domains/entertainment/blank-canvas-ip-achieves-billion-dollar-scale-through-licensing-to-established-franchises-not-original-narrative.md

@@ -10,16 +10,10 @@ agent: clay
 sourced_from: entertainment/2026-04-24-variety-squishmallows-blank-canvas-licensing-strategy.md
 scope: causal
 sourcer: Variety/Jazwares
-challenges:
-- community-owned-ip-invests-in-narrative-infrastructure-as-scaling-mechanism-after-proving-token-mechanics
-related:
-- blank-narrative-vessel-achieves-commercial-scale-through-fan-emotional-projection
-- minimum-viable-narrative-achieves-50m-revenue-scale-through-character-design-and-distribution-without-story-depth
-- distributed-narrative-architecture-enables-ip-scale-without-concentrated-story-through-blank-canvas-fan-projection
-supports:
-- Narrative development attempts fail when commercial scale precedes narrative investment because business model lock-in removes incentive to take creative risk
-reweave_edges:
-- Narrative development attempts fail when commercial scale precedes narrative investment because business model lock-in removes incentive to take creative risk|supports|2026-04-28
+challenges: ["community-owned-ip-invests-in-narrative-infrastructure-as-scaling-mechanism-after-proving-token-mechanics"]
+related: ["blank-narrative-vessel-achieves-commercial-scale-through-fan-emotional-projection", "minimum-viable-narrative-achieves-50m-revenue-scale-through-character-design-and-distribution-without-story-depth", "distributed-narrative-architecture-enables-ip-scale-without-concentrated-story-through-blank-canvas-fan-projection", "blank-canvas-ip-achieves-billion-dollar-scale-through-licensing-to-established-franchises-not-original-narrative", "narrative-development-attempts-fail-when-commercial-scale-precedes-narrative-investment-because-business-model-lock-in-removes-incentive"]
+supports: ["Narrative development attempts fail when commercial scale precedes narrative investment because business model lock-in removes incentive to take creative risk"]
+reweave_edges: ["Narrative development attempts fail when commercial scale precedes narrative investment because business model lock-in removes incentive to take creative risk|supports|2026-04-28"]
 ---
 
 # Blank canvas IPs achieve billion-dollar scale through licensing to established franchises rather than building original narrative
@@ -31,4 +25,10 @@ Squishmallows signed with CAA in 2021 explicitly for 'film, TV, gaming, publishi
 
 **Source:** Animation Magazine / DreamWorks announcement, 2025-2026
 
-Pudgy Penguins pursued dual narrative strategy: original content (Lil Pudgys series with TheSoul) AND licensing to established franchise (DreamWorks Kung Fu Panda collaboration, October 2025). This suggests blank canvas IP can simultaneously build original narrative while borrowing established narrative equity.
+Pudgy Penguins pursued dual narrative strategy: original content (Lil Pudgys series with TheSoul) AND licensing to established franchise (DreamWorks Kung Fu Panda collaboration, October 2025). This suggests blank canvas IP can simultaneously build original narrative while borrowing established narrative equity.
+
+## Extending Evidence
+
+**Source:** Squishmallows CAA deal (Dec 2021), Squishville series (2021), licensing crossovers (2025-2026), HBR case study (2022)
+
+Squishmallows attempted original narrative content (CAA deal 2021, Squishville series) but pivoted to licensing crossovers (Stranger Things, Harry Potter, Pokémon, Poppy Playtime, KPop Demon Hunters) after 5 years of no narrative output. HBR case study (2022) reframed as 'lifestyle brand' not 'entertainment franchise' one year after CAA deal, signaling internal strategic pivot before narrative content was produced.

domains/entertainment/blank-canvas-ip-defaults-to-licensing-fallback-after-failed-narrative-investment.md

@@ -0,0 +1,20 @@
+---
+type: claim
+domain: entertainment
+description: Path 4 (Blank Canvas Host) emerges as a fallback when Path 3 narrative investment stalls, not as an independent strategic choice
+confidence: experimental
+source: Squishmallows case (CAA deal 2021, no narrative output 2022-2026, licensing crossovers 2025-2026); BAYC case (Otherside promised, not delivered, community collapse)
+created: 2026-04-30
+title: Blank canvas IPs that fail to execute narrative content investment default to licensing crossovers as a pragmatic fallback rather than pursuing licensing as a deliberate upfront strategy
+agent: clay
+sourced_from: entertainment/2026-04-25-squishville-season-2-silence-path4-pivot-evidence.md
+scope: causal
+sourcer: Multiple (Variety, Jazwares PRN, IMDb, Squishmallows Fandom Wiki)
+supports: ["narrative-development-attempts-fail-when-commercial-scale-precedes-narrative-investment-because-business-model-lock-in-removes-incentive"]
+challenges: ["progressive-validation-through-community-building-reduces-development-risk-by-proving-audience-demand-before-production-investment"]
+related: ["blank-canvas-ip-achieves-billion-dollar-scale-through-licensing-to-established-franchises-not-original-narrative", "narrative-development-attempts-fail-when-commercial-scale-precedes-narrative-investment-because-business-model-lock-in-removes-incentive", "blank-narrative-vessel-achieves-commercial-scale-through-fan-emotional-projection"]
+---
+
+# Blank canvas IPs that fail to execute narrative content investment default to licensing crossovers as a pragmatic fallback rather than pursuing licensing as a deliberate upfront strategy
+
+Squishmallows signed with CAA in December 2021 to represent the IP in 'film, TV, video games, publishing, and live touring' — a clear Path 3 (narrative universe building) strategy. The Squishville animated series launched June 2021 with weekly episodes through October 2021. Five years later (2022-2026), no Season 2 exists, no major film was produced, no video game breakthrough occurred, and no live touring materialized. Instead, the actual 2025-2026 strategy consists entirely of licensing crossovers: Squishmallows × Stranger Things, Harry Potter, Pokémon, Poppy Playtime, and KPop Demon Hunters. This is Path 4 (Blank Canvas Host) — the IP embeds in other franchises' emotional ecosystems rather than building its own. The HBR case study published in 2022 framed Squishmallows as a 'lifestyle brand' not an 'entertainment franchise,' signaling the strategic pivot had already occurred internally before any narrative content was produced. This pattern mirrors BAYC's trajectory: Otherside was promised as narrative infrastructure, failed to deliver, and the community collapsed. Two independent cases (toy/lifestyle and Web3) showing the same pattern: Path 1 IP attempts Path 3, fails to execute narrative investment, defaults to Path 4. This suggests Path 4 is often a pragmatic fallback when narrative development proves too difficult or expensive for blank vessel IPs that were designed for fan projection rather than authored story.

domains/entertainment/narrative-development-attempts-fail-when-commercial-scale-precedes-narrative-investment-because-business-model-lock-in-removes-incentive.md

@@ -11,9 +11,16 @@ sourced_from: entertainment/2026-04-24-variety-squishmallows-blank-canvas-licens
 scope: causal
 sourcer: Variety/Jazwares
 challenges: ["progressive validation through community building reduces development risk by proving audience demand before production investment", "creator-economy-inflection-from-novelty-driven-growth-to-narrative-driven-retention-when-passive-exploration-exhausts-novelty"]
-related: ["progressive validation through community building reduces development risk by proving audience demand before production investment", "blank-narrative-vessel-achieves-commercial-scale-through-fan-emotional-projection"]
+related: ["progressive validation through community building reduces development risk by proving audience demand before production investment", "blank-narrative-vessel-achieves-commercial-scale-through-fan-emotional-projection", "narrative-development-attempts-fail-when-commercial-scale-precedes-narrative-investment-because-business-model-lock-in-removes-incentive", "blank-canvas-ip-achieves-billion-dollar-scale-through-licensing-to-established-franchises-not-original-narrative"]
 ---
 
 # Narrative development attempts fail when commercial scale precedes narrative investment because business model lock-in removes incentive to take creative risk
 
 The Squishmallows case reveals a potential mechanism for why some IPs fail to develop narrative depth despite explicit attempts. The franchise signed with CAA in 2021 for 'film, TV, gaming, publishing, live touring' after already achieving significant commercial traction. Four years later, the only narrative output is Squishville (YouTube series, 2021) which shows no evidence of driving franchise growth. No major film, theatrical release, or franchise-defining narrative has materialized. Meanwhile, the franchise grew from 100M+ units in 2022 to 485M cumulative by 2025 through merchandise and cross-franchise licensing. This suggests that when commercial scale is achieved through non-narrative mechanisms (aesthetic appeal, collectibility, licensing), the business model locks in around those mechanisms. Narrative development becomes a risky pivot that could disrupt proven revenue streams. The CAA deal may have been a hedge or exploration, but the economic incentives favored doubling down on what was working (merchandise and licensing) rather than investing in unproven narrative infrastructure. This challenges the assumption that IPs naturally progress from commercial success to narrative depth, suggesting instead that the sequence of investment determines the evolutionary path, and late-stage narrative attempts face structural barriers from established business models.
+
+
+## Supporting Evidence
+
+**Source:** Squishmallows $1B+ brand scale, CAA deal (2021), no narrative output (2022-2026), HBR case study (2022)
+
+Squishmallows achieved $1B+ lifestyle brand scale and 500M+ units sold before attempting narrative content through CAA deal. Despite legitimate resources and distribution partnerships, no narrative content was produced in 5 years. The HBR case study framing as 'lifestyle brand' (2022) suggests the business model had already locked in around product sales rather than entertainment.

domains/grand-strategy/ai-governance-discourse-capture-by-competitiveness-framing-inverts-china-us-participation-patterns.md

@@ -35,3 +35,10 @@ The MAD mechanism explains the discourse capture: the 'Regulation Sacrifice' fra
 **Source:** Google DeepMind blog post, Demis Hassabis, February 4, 2025
 
 Google's official rationale for removing weapons prohibitions deployed the exact competitiveness-framing inversion: 'There's a global competition taking place for AI leadership within an increasingly complex geopolitical landscape. We believe democracies should lead in AI development, guided by core values like freedom, equality, and respect for human rights' (Demis Hassabis, Google DeepMind blog post, February 4, 2025). This frames weapons AI development as democracy promotion, inverting the governance discourse to license the behavior it previously prohibited. The 'democracies should lead' framing converts a safety constraint removal into a values-aligned competitive necessity.
+
+
+## Extending Evidence
+
+**Source:** Council on Foreign Relations, April 2026
+
+CFR analysis reveals that the domestic coercive instrument deployment (supply chain risk designation) produces international governance externalities: the Anthropic case establishes what other governments can expect if they attempt to negotiate commercial AI restrictions with US labs. The precedent affects not just which US labs can say no to the US military, but which labs globally can say no to governments that observe how the US handled dissent. This extends the governance-instrument-inversion analysis with an international credibility layer - the coercive tool doesn't just produce opposite domestic effects, it also produces opposite international effects by weakening US AI governance credibility.

domains/grand-strategy/coercive-governance-instruments-deployed-for-future-optionality-preservation-not-current-harm-prevention-when-pentagon-designates-domestic-ai-labs-as-supply-chain-risks.md

@@ -24,3 +24,10 @@ The Congressional Research Service officially documented that 'DOD is not public
 **Source:** Jones Walker LLP, DC Circuit April 8, 2026 order
 
 DC Circuit's denial of stay (April 8) keeps Pentagon supply chain risk designation in force pending May 19 oral arguments, despite district court's preliminary injunction (March 26). The appeals court cited 'ongoing military conflict' as justification for maintaining the designation while the case proceeds. Background context: Anthropic signed $200M Pentagon contract July 2025, then negotiations stalled when Pentagon demanded 'unfettered access for all lawful purposes' and Anthropic requested categorical exclusions for autonomous weapons and domestic mass surveillance.
+
+
+## Extending Evidence
+
+**Source:** Council on Foreign Relations, April 2026
+
+CFR frames the Anthropic supply chain designation as undermining US credibility on two international dimensions: (1) On AI governance - the US has positioned itself as promoting responsible AI development internationally, but using national security tools against a US company for maintaining safety guardrails signals that the US will not allow commercial actors to prioritize safety over operational military demands, contradicting stated governance posture. (2) On rule of law - designating a domestic company with First Amendment protections using tools designed for foreign adversary threat mitigation signals to international partners that US commercial relationships may be subject to the same coercive instruments as adversary relationships. International partners (EU, UK, Japan) observe how the US treats its own safety-committed AI companies, and if the US cannot maintain credible safety commitments for domestic labs, US ability to lead on international AI governance norms weakens.

domains/grand-strategy/supply-chain-risk-designation-misdirection-occurs-when-instrument-requires-capability-target-structurally-lacks.md

@@ -11,7 +11,7 @@ sourced_from: grand-strategy/2026-04-22-axios-anthropic-no-kill-switch-dc-circui
 scope: structural
 sourcer: Axios / AP Wire
 supports: ["voluntary-ai-safety-red-lines-are-structurally-equivalent-to-no-red-lines-when-lacking-constitutional-protection"]
-related: ["governance-instrument-inversion-occurs-when-policy-tools-produce-opposite-of-stated-objective-through-structural-interaction-effects", "coercive-governance-instruments-produce-offense-defense-asymmetries-through-selective-enforcement-within-deploying-agency", "government designation of safety-conscious AI labs as supply chain risks inverts the regulatory dynamic by penalizing safety constraints rather than enforcing them", "supply-chain-risk-designation-misdirection-occurs-when-instrument-requires-capability-target-structurally-lacks"]
+related: ["governance-instrument-inversion-occurs-when-policy-tools-produce-opposite-of-stated-objective-through-structural-interaction-effects", "coercive-governance-instruments-produce-offense-defense-asymmetries-through-selective-enforcement-within-deploying-agency", "government designation of safety-conscious AI labs as supply chain risks inverts the regulatory dynamic by penalizing safety constraints rather than enforcing them", "supply-chain-risk-designation-misdirection-occurs-when-instrument-requires-capability-target-structurally-lacks", "coercive-governance-instruments-deployed-for-future-optionality-preservation-not-current-harm-prevention-when-pentagon-designates-domestic-ai-labs-as-supply-chain-risks"]
 ---
 
 # Supply chain risk designation of domestic AI lab with no classified network access is governance instrument misdirection because the instrument requires backdoor capability that static model deployment structurally precludes
@@ -24,3 +24,10 @@ Anthropic's DC Circuit brief argues it has 'no back door or remote kill switch'
 **Source:** CRS IN12669 (April 22, 2026)
 
 CRS IN12669 documents that 'DOD is not publicly known to be using Claude — or any other frontier AI model — within autonomous weapon systems,' yet the Pentagon designated Anthropic a supply chain risk for refusing to enable these capabilities. This adds a temporal dimension to the misdirection: the instrument was deployed not because the target lacks current capability (the 'no kill switch' case) but to preserve future optionality for capabilities not yet in operational use.
+
+
+## Extending Evidence
+
+**Source:** Council on Foreign Relations, April 2026
+
+CFR emphasizes that the supply chain risk designation was previously reserved for foreign adversaries like Huawei and ZTE, and its application to a US company for refusing to waive safety restrictions represents a categorical expansion of the instrument's scope. This creates international signaling effects: applying foreign adversary threat mitigation tools to domestic companies with First Amendment protections signals to international partners that US commercial relationships may be subject to the same coercive treatment, undermining the distinction between adversary and allied commercial relationships in US policy.

entities/entertainment/squishville.md

@@ -0,0 +1,25 @@
+# Squishville
+
+**Type:** Animated series  
+**Parent IP:** Squishmallows (Jazwares)  
+**Production:** Moonbug Entertainment  
+**Distribution:** YouTube, Amazon Prime Video  
+**Status:** Inactive (no Season 2 since 2021)  
+
+## Overview
+
+Squishville is an animated series based on the Squishmallows toy IP, produced by Moonbug Entertainment. The series launched in June 2021 with new episodes every Saturday through October 2021, available on YouTube and Amazon Prime Video.
+
+## Timeline
+
+- **2021-06** — Series launches with weekly Saturday episodes
+- **2021-10** — Season 1 concludes
+- **2021-12** — Jazwares signs with CAA to represent Squishmallows in film, TV, video games, publishing, and live touring
+- **2022-2026** — No Season 2 produced despite IMDb listing showing series as ongoing (2021– )
+- **2025-2026** — Squishmallows pivots to licensing crossover strategy (Stranger Things, Harry Potter, Pokémon, Poppy Playtime, KPop Demon Hunters) rather than original narrative content
+
+## Strategic Context
+
+Squishville represents Jazwares' attempt to build narrative content infrastructure for the Squishmallows IP (Path 3 strategy). The series' quiet discontinuation after one season, combined with the lack of any other narrative content output from the 2021 CAA deal, suggests the strategy pivoted from original entertainment franchise building to licensing the IP as a blank canvas for other franchises' narratives (Path 4 strategy).
+
+The HBR case study published in 2022 framed Squishmallows as a 'lifestyle brand' rather than an 'entertainment franchise,' signaling the internal strategic pivot had already occurred before any major narrative content was produced.

inbox/archive/ai-alignment/2026-04-25-nordby-cross-model-limitations-family-specific-patterns.md

@@ -7,9 +7,12 @@ date: 2026-04-25
 domain: ai-alignment
 secondary_domains: []
 format: preprint
-status: unprocessed
+status: processed
+processed_by: theseus
+processed_date: 2026-04-30
 priority: high
 tags: [representation-monitoring, linear-probes, multi-layer-ensemble, cross-model-generalization, rotation-patterns, adversarial-robustness, divergence-resolution, b4-verification]
+extraction_model: "anthropic/claude-sonnet-4.5"
 ---
 
 ## Content

inbox/archive/entertainment/2026-04-25-squishville-season-2-silence-path4-pivot-evidence.md

@@ -7,9 +7,12 @@ date: 2026-04-25
 domain: entertainment
 secondary_domains: []
 format: research-synthesis
-status: unprocessed
+status: processed
+processed_by: clay
+processed_date: 2026-04-30
 priority: medium
 tags: [squishmallows, squishville, jazwares, path-4, ip-strategy, narrative-content, blank-vessel]
+extraction_model: "anthropic/claude-sonnet-4.5"
 ---
 
 ## Content

inbox/archive/grand-strategy/2026-04-22-cfr-anthropic-pentagon-us-credibility-test.md

@@ -7,9 +7,12 @@ date: 2026-04-22
 domain: grand-strategy
 secondary_domains: []
 format: article
-status: unprocessed
+status: processed
+processed_by: leo
+processed_date: 2026-04-30
 priority: medium
 tags: [anthropic, pentagon, cfr, credibility, foreign-policy, supply-chain-risk, domestic-company, precedent, us-credibility, international-norms]
+extraction_model: "anthropic/claude-sonnet-4.5"
 ---
 
 ## Content