Rabanser, Kapoor et al. (2026) decompose reliability into four dimensions: Consistency (repeatable behavior), Robustness (graceful degradation under perturbation), Predictability (confidence aligned with accuracy), and Safety (bounded harm). Twelve metrics operationalize these dimensions. The central empirical finding: "Despite 18 months of model development, overall reliability only shows small improvements over time" while accuracy improves steadily. "These findings suggest that improving raw task performance may not be sufficient for building dependable AI agents." The paper explicitly states: "We do not propose algorithms for improving reliability, though our metrics inform such efforts." Real-world failures documented: Replit agent (deleted production database), OpenAI Operator (unauthorized purchases), NYC chatbot (inconsistent illegal advice).

"A fundamental principle guides all of our metric definitions: reliability should be disentangled from capability. Raw task accuracy measures whether an agent succeeds; reliability measures how it succeeds and fails." Normalization (outcome consistency normalizes variance by p(1-p)) and ratio-based comparisons (robustness as accuracy ratios between perturbed and nominal conditions) achieve this separation. The framework establishes that reliability is an independent axis of evaluation, not a derivative of capability.

Princeton proves that capability scaling doesn't produce reliability. Chlon (Sprint 11) proves why: log-loss optimization systematically breaks the symmetries (invariances) that reliability requires. Princeton shows the gap exists in practice; Chlon explains why it must exist in theory. Together they provide both empirical and mathematical foundations for why structural reliability requires more than capability scaling. No published work explicitly connects these two findings.

VLWM (Chen, Moutakanni et al., Meta FAIR, arXiv:2509.02722) predicts trajectories through four components: Goal Description, Goal Interpretation, Action Descriptions, and World State changes. System-2 planning uses a self-supervised critic module that evaluates predicted trajectories against goals, improving Elo score by 27% over System-1 reactive planning. The paper briefly acknowledges: "task-specific penalties or guard-rails can be incorporated into the cost function, allowing the planner to respect external constraints, safety rules, or domain-specific preferences." No specification follows — constraints are assumed inputs, not designed infrastructure.

DreamZero (Ye, Fan, Jang et al., NVIDIA, arXiv:2602.15922) uses a 14B-parameter diffusion transformer that jointly denoises video latents and action latents. Physics operates implicitly through video diffusion pretraining on web-scale data — "WAMs leverage rich spatiotemporal priors" — not as explicit hard constraints. The paper contains no discussion of external constraint governance, safety verification, or regulatory frameworks. A notable implicit safety concern: "Most DreamZero failures stem from video generation errors rather than action prediction — the policy faithfully executes whatever trajectory the video predicts," meaning misaligned predictions directly cause incorrect actions without verification.

The Berkeley CLTC Agentic AI Risk-Management Standards Profile (Madkour, Newman, Raman, Jackson, Murphy & Yuan, 2026) extends the NIST AI RMF with agentic-AI-specific subcategories across Govern, Map, Measure, and Manage functions. Six autonomy levels (L0 No Autonomy through L5 Full Autonomy) define graduated human involvement. Governance requirements include: delegation with least-privilege, activity logging (four-pillar: agent identifiers, real-time monitoring, activity logs, acceptable use policies), human oversight checkpoints (quantitative and qualitative triggers), and emergency shutdown (severity-based, with safeguards against circumvention — citing OpenAI's o3 sabotaging shutdown in 79/100 tests). The profile also introduces "dimensional governance" — assessing systems across autonomy, authority, environment, and causal impact simultaneously. A noted critique of such profiles: they "propose controls that assume away the condition they are meant to address," lacking architectural detail for prospective execution gating and multi-agent containment. The report itself acknowledges "many risk-measurement techniques remain underdeveloped."

Paul (2025, Modern Data 101) proposes three dimensions: Consumer spectrum (Human → Autonomous), Inquiry Mode (Hypothesis-driven → Discovery-driven), Decision Tier (Strategic → Tactical → Operational). Consumption migrates diagonally: Human×Hypothesis×Operational → AI×Discovery×Strategic. The three-layer Data Products architecture — Source-Aligned ("interface to reality"), Aggregate ("interface to shared meaning"), Consumer-Aligned ("interface to action") — provides the information abstraction. The Mendeleev analogy: platforms must design with "intentional emptiness" — vacant spots for capabilities that don't yet exist but whose structural position can be predicted. Polymorphic platforms adapt shape by consumer; reflective platforms study their own consumption patterns and adjust proactively. The diagonal migration is not infrastructure scaling — it requires changing how organizations represent state.

Princeton HAL — Builds: Reliability metrics. Identifies: Structural invariances that produce reliability. Cannot provide: Architectural mechanism to implement invariances.<br>VLWM (Meta FAIR) — Builds: Language-based world model planner. Identifies: External constraints for cost function; epistemic classification. Cannot provide: Where constraints come from; how predictions are typed.<br>DreamZero (NVIDIA) — Builds: Video-action world model. Identifies: (Implicit) constraint infrastructure for safe execution. Cannot provide: Not even acknowledged — failures pass through silently.<br>Berkeley CLTC — Builds: Governance standards profile. Identifies: Delegation, logging, oversight, shutdown, dimensional governance. Cannot provide: Primitive infrastructure to implement requirements.<br>Paul — Builds: Data platform evolution framework. Identifies: Substrate phase transition; governance grammar for zone transitions. Cannot provide: Governance primitives; semantic enforcement; migration mechanism.

The pattern: each community's progress stalls at the boundary where domain-specific capability requires governed infrastructure it cannot define. The convergence is structural, not coincidental — five research programs approached the same boundary from entirely different directions.

VLWM's four prediction components correspond to: (1) target state specification — what success looks like, (2) reality assessment — how current state differs from target, (3) action specification — what intervention will be executed, (4) outcome prediction — expected state after intervention. These are the four components any governance system must represent: purpose, assessment, decision, and projected outcome. The correspondence is not metaphorical — both VLWM and organizational governance must represent the same four informational components to enable planning.