Asoba Research  ·  July 2026

Bounded Intelligence: Coupling a JEPA World Model With a Retrieval-Grounded Reasoner

An anomaly score confirms that physical behavior has departed from normal operating boundaries. It doesn’t identify the cause, and it doesn’t tell an operator what to do about it. Closing that gap safely means refusing to let one model do both jobs.

6 sections 13 citations ~11 min read
I

The Gap Between Detection and Diagnosis

When an inverter’s behavior falls outside healthy operating parameters, a world model built on a Joint-Embedding Predictive Architecture (JEPA) flags the deviation and quantifies its severity — with a fraction of the false-alarm volume generated by traditional OEM thresholds. JEPA models continuous system dynamics in a latent space rather than generating raw pixels or time-series points, and it is the empirically stronger approach to physical-state prediction from telemetry.[0]He, Y., Wen, Y., Wang, X., Ma, T. (2026). MTS-JEPA: Multi-Resolution JEPA for Time-Series Anomaly Prediction. arXiv:2602.04643. Beats reconstruction, contrastive, and Transformer baselines on spacecraft and industrial telemetry.

Fig. 1a — Rule-Based Thresholds vs. JEPA World Model — INV-195, Sibaya (216 test faults)
DETECTION QUALITY BY METRIC (per-alert precision) 0% 25% 50% 75% 100% 23.8% 37.2% Precision ~100% 100% Recall 38.5% 54.2% F1 score Rule-based thresholds JEPA world model Precision alone understates it — below is what these percentages mean at actual alert volume. TOTAL ALERTS FIRED PER 100 REAL FAULTS RULE-BASED 9,862 JEPA 269 100 real 169 false -97% total alert volume  (9,862 → 269) 57.8x fewer false alarms  (9,762 → 169) 100% recall unchanged — same 100 real faults caught real fault caught false alarm
Fig. 1a — INV-195, the flagship Sibaya model (216 test faults, the largest evaluable set). Both methods catch the same 100 real faults; the true-scale bars show what that 56% precision gap actually costs in noise — rules bury every real fault under 98 false ones, JEPA under fewer than two. On the sparser INV-192 model (5 test faults), the same pattern holds at smaller scale: precision improves 0.006 → 0.022, a 3.5x reduction in false-alarm share. Source: Ona Platform Team, LeWorldModel deployment report, April 2026.

Within the OODA loop — observe, orient, decide, act — this predictive step completes the orientation phase. But by itself, continuous anomaly scoring is necessary and insufficient. A severity score of 0.87 across six consecutive observation windows says output has diverged from baseline in a way that correlates with an imminent production drop. It does not distinguish an overheating heat sink from a firmware handshake fault from an isolated component failure buried in a manufacturer troubleshooting manual.

Translating a deviation into a verified cause and a recommended check requires qualitative reasoning layered on top of the continuous output of the prediction-error model — and that reasoning has to be added without breaking the cost and determinism the world model was built to guarantee.


II

Bounding Fast and Slow Cognition

In Thinking, Fast and Slow, Daniel Kahneman split human cognition into System 1 — automatic, fast, always running — and System 2 — deliberate, effortful, invoked to evaluate what System 1 surfaces.[1]Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux. That split maps directly onto the monitoring architecture below.

System 1 / System 2 — Diagnostic Pipeline
Continuous Telemetry SYSTEM 1 — ONA JEPA WORLD MODEL Fast, low-cost forward pass Continuous physics-conditioned tracking, every inverter severity ≥ moderate  OR  streak ≥ 3 Deterministic Query Engine Translates payload into OEM-manual vocabulary Hybrid Retrieval Pipeline (RAG) Dense semantic + sparse keyword search over OEM corpus Cross-encoder re-ranking of candidate passages SYSTEM 2 — NEHANDA REASONER Bounded synthesis over retrieved context only Fine-tuned to refuse ungrounded guesses Cited, Actionable Diagnostic Payload  —  Orient
Fig. 1 — The JEPA world model runs continuously as System 1; Nehanda triggers only past an explicit severity/streak threshold, gated by cost and by the risk of ungrounded synthesis.

The JEPA world model executes at multi-minute intervals across every inverter, cheaply enough that compute cost does not constrain design. Nehanda v3, Asoba’s fine-tune of Qwen3.6 27B, functions as System 2 and does not run continuously — it triggers only when the world model’s output crosses an explicit threshold: a severity score at or above moderate, or a sustained streak of three consecutive anomalous windows.

Cheap, narrow intelligence must execute the vast majority of tasks, reserving general reasoning for conditions that have earned the compute expense.

Evaluating retrieved technical literature through a language model is expensive relative to a forward pass through a compact encoder. Invoking a large model on every telemetry packet destroys the economics of a first-pass filter. The gating threshold isn’t an operational compromise — it’s the architecture’s core thesis.


III

Structuring Context Through Retrieval

In 2020, Patrick Lewis and colleagues at Facebook AI Research formalized Retrieval-Augmented Generation — grounding language model output by retrieving relevant documents from an external corpus at query time and supplying them as explicit context, so responses trace back to verifiable sources rather than memorized training data.[2]Lewis, P. et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. arXiv:2005.11401. In this architecture, no human writes the prompt — the detection payload generated by the world model constructs the retrieval query itself.

Structured anomaly fields — fault codes, power loss relative to an irradiance-conditioned baseline, core temperatures, streak lengths — are converted into the explicit terminology used in technical documentation. An operating temperature exceeding 60°C translates into thermal-management terms; a persistent error code maps directly to a component-failure classification. The anomaly writes its own question, which then executes against a corpus of tens of thousands of OEM manuals, field engineering guides, and maintenance logs.

The novelty isn’t the retrieval mechanics — it’s the deterministic constraint on the input. Generating queries programmatically from verified telemetry fields eliminates free-text hallucination risk and constrains the search domain before a language model ever sees the question.


IV

Prioritizing Restraint Over Scale

Deploying an unconstrained frontier model in industrial operations introduces unacceptable risk. A confident, incorrect repair instruction is more damaging than a slow one — a hallucinated failure mechanism wastes a costly site visit, or worse, leaves the real fault undiagnosed while the asset stays exposed.

Nehanda was fine-tuned specifically against this failure mode: the training objective optimizes for source fidelity, strict adherence to retrieved context, and explicit refusal when documentation can’t establish cause. On FACTS Grounding — DeepMind’s 1,719-example benchmark for factual adherence to source documents[3]FACTS Grounding Leaderboard, Google DeepMind / Kaggle. 860 public examples across finance, law, medicine, technology. — Nehanda v3 outperformed significantly larger frontier models evaluated on the identical public set:

FACTS Grounding — Factual Adherence Score
ModelFACTS Grounding score
Nehanda v3 (27B, fine-tuned)88.7%
Gemini 2.5 Pro87.8%
Claude 3.5 Sonnet83.8%
GPT-4o79.8%

Fig. 2 — A 27B model fine-tuned for source fidelity outranks substantially larger general-purpose models on grounded factuality. Source: FACTS Grounding public leaderboard.

Domain-specific restraint outranks parameter count here. General capability was never the optimization target — restraint was: the trained habit of refusing to answer without evidence, which does not scale automatically with model size.


V

Epistemic Division of Labor

The architecture formulated in late 2025 excluded generative models entirely, for three reasons: proprietary cloud APIs expose critical infrastructure to deprecation and price risk; multi-billion-parameter inference is the most expensive compute tier, making continuous evaluation cost-prohibitive; and language models processing numerical time-series through text tokens lack an explicit representation of physical dynamics.

Those limits still hold. A 2025 study led by John Paparrizos found that autoregressive language models struggle with complex time-series anomalies and lack the structured reasoning needed for physical-state estimation — purpose-built forecasting models remain superior for continuous state tracking.[4]Paparrizos, J., Boniol, P., Liu, Q. et al. (2025). Advances in Time-Series Anomaly Detection. KDD '25. doi:10.1145/3711896.3736565 Language models should not perform anomaly detection over raw telemetry.

Broader reasoning benchmarks reinforce this. ARC-AGI-3, released by the ARC Prize Foundation in March 2026 to measure fluid problem-solving without linguistic memorization, showed frontier models scoring under 1% at launch.[5]ARC Prize Foundation. ARC-AGI-3 launched March 25, 2026, San Francisco.[6]ARC Prize Foundation. ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence. arXiv:2603.24621 By July 2026, pure language models had improved but remained bounded on abstract dynamic tasks:

ARC-AGI-3 Score — Unassisted Language Models, July 2026
ModelARC-AGI-3 score
Claude Opus 530.2%
Claude Fable 5~16–20%
Claude Mythos 5 (shares Fable 5’s base architecture)~16–20%
GPT-5.6 Sol (best variant)7.8%

Fig. 3 — Unassisted frontier language models remain bounded on interactive, dynamics-heavy reasoning even as scores improve off a near-zero March 2026 baseline.

By contrast, Sergey Rodionov at SingularityNET showed that pairing a language model with an explicit, executable world model yields far stronger performance: the language model writes and runs a Python program representing environment dynamics and plans through that symbolic representation rather than predicting raw state transitions.[7]Rodionov, S. (2026). Executable World Models for ARC-AGI-3 in the Era of Coding Agents. arXiv:2605.05138 That hybrid approach solved multiple public ARC-AGI-3 environments outright and posted a mean efficiency score well above what any unassisted language model reached on its own.

The Confirming Result

Couple an explicit, low-cost state model with a retrieval-grounded reasoner, rather than forcing a single model to do both jobs. That is the core architectural premise this system was built on — and it is now showing up independently in general reasoning benchmarks, not just in industrial telemetry.


VI

Where Decide Stays Deterministic

The Decide stage of the OODA loop remains entirely deterministic. Generative models do not trigger field dispatches, execute switching actions, or alter physical control loops — identical inputs yield identical control actions every time, a guarantee non-deterministic models cannot supply. Generative synthesis operates only in the Orient phase, as an additive diagnostic field: fully cited, threshold-gated, and built to say when data is insufficient.

End-to-End OODA Placement
Telemetry Observe JEPA World Model System 1 Orient (physical) RAG + Nehanda System 2 Orient (cited diagnosis) gated, not continuous Decide deterministic rules, not models Act dispatch
Fig. 4 — Generative models occupy Orient only. Decide and Act stay deterministic and reproducible regardless of what System 1 or System 2 report.

Where this generalizes. Any industrial environment with three properties needs this same split: a continuous, narrow telemetry stream cheap enough to evaluate constantly; a large corpus of technical documentation required to interpret deviations; and a high penalty for confident false diagnostics. Battery degradation, wind-turbine drivetrain vibration, and distribution-grid stability telemetry all fit. None of them need conversational general intelligence — they need a fast filter and a reasoner trained to tell a cited synthesis from a guess.

The core engineering challenge will rarely be about expanding model size. It will be about building precise architectural boundaries — knowing exactly where cheap, continuous intelligence must end, and where slow, verifiable reasoning must begin.

Forcing continuous physical tracking and diagnostic synthesis into one monolithic model compromises both execution speed and operational safety. Decoupling them achieves diagnostic precision without inheriting the cost, non-determinism, or epistemic frailty of an unconstrained frontier model.

Citations & Sources
[0]
He, Y., Wen, Y., Wang, X., Ma, T. (2026). MTS-JEPA: Multi-Resolution Joint-Embedding Predictive Architecture for Time-Series Anomaly Prediction. Benchmarked against reconstruction, contrastive, and Transformer baselines on NASA spacecraft, industrial control, and server telemetry. arxiv.org/abs/2602.04643 →
[1]
Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux — source for the System 1 / System 2 cognition framing.
[2]
Lewis, P. et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. arxiv.org/abs/2005.11401 →
[3]
FACTS Grounding benchmark, Google DeepMind. 1,719 examples (860 public). deepmind.google → / kaggle.com →
[4]
Paparrizos, J., Boniol, P., Liu, Q. et al. (2025). Advances in Time-Series Anomaly Detection: Algorithms, Benchmarks, and Evaluation Measures. KDD ’25. doi.org/10.1145/3711896.3736565 →
[5]
ARC Prize Foundation. ARC-AGI-3 launch, March 25, 2026, Y Combinator SF. arcprize.org →
[6]
ARC Prize Foundation. ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence. arxiv.org/abs/2603.24621 →
[7]
Rodionov, S. (2026), SingularityNET. Executable World Models for ARC-AGI-3 in the Era of Coding Agents. arxiv.org/abs/2605.05138 →
[8]
Samudzi, S. (2026). Per-Inverter World Models for Unsupervised Fault Detection in Distributed Solar Fleets. zenodo.org →
[9]
Asoba Corporation. Nehanda v3: RAG Synthesis on a Native Multimodal Architecture. Read on Asoba Research →
[10]
Asoba Corporation. Epistemic Robustness via Stacked Cognitive Training. Read on Asoba Research →
[11]
Asoba Corporation. Generative AI LLMs vs. Specialized Neural Networks for Decision-Making. Read on Asoba Research →
[12]
Asoba Corporation. Anomaly Detection for PV Solar Inverters with JEPA. Read on Asoba Research →