agent-precog.org
FIVE RESEARCH QUESTIONS · 23 PAPERS SCORED AGAINST THEM

What PreCog sets out to answer

Each question is paired with falsifiable hypotheses, the literature pillars that feed it, and the gap in the included papers it addresses. Every paper in the review carries a 0–5 relevance score against each question; the counts below are papers scoring 3 or more.

RQ1 · ADDRESSES PROBLEMS 1 AND 2: THE LATENCY TRAP AND STATELESS EVALUATION

Sub-15ms latent world-model interception

9 papers scoring ≥ 3 →

Can a non-generative Hyperbolic Graph-JEPA, compiled via TensorRT on edge silicon, ingest a localised 3-hop multi-agent sub-graph and accurately halt emergent catastrophic trajectories in under 15 milliseconds?

Hypotheses
  1. End-to-end Fast Path latency ≤ 15 ms at p99 under nominal wireless conditions.
  2. Rejector reaches ROC-AUC ≥ 0.96 on held-out emergent-catastrophe trajectories, detecting before the time-to-irreversibility boundary.
  3. At matched recall, lower false-halt rate than a stateless per-call classifier and a 1-hop graph baseline.
Gap in the literature

No included paper couples a graph world model to a safety decision on edge hardware.

RQ2 · ADDRESSES PROBLEM 3: THE STATIC DEFERRAL COST

Jitter-adaptive sequential learning to defer

7 papers scoring ≥ 3 →

How can a two-stage sequential L2D surrogate loss adapt its deferral threshold in real time to 99th-percentile wireless jitter relative to a domain's physical or transactional time-to-irreversibility?

Hypotheses
  1. Under injected jitter, the adaptive threshold prevents a higher share of catastrophes before TTI than a fixed threshold calibrated on a clean network.
  2. The cost in additional false halts is bounded and degrades gracefully across the tested jitter range.
Gap in the literature

No deferral paper in the review treats signal latency or a deadline as an input to the decision.

RQ3 · ADDRESSES PROBLEMS 5 AND 6: THE OPAQUE HALT AND DEFERRAL UNDER MOTION

Dual-speed socio-technical handoff

6 papers scoring ≥ 3 →

Does decoupling safety into a Fast Path and an asynchronous Slow Path generalise across physical robotics and logical enterprise graphs while measurably reducing human out-of-the-loop recovery latency?

Hypotheses
  1. Operators given the Dual-Speed brief reach a correct release / re-plan / escalate decision faster than operators given raw telemetry.
  2. They make fewer incorrect decisions and report lower workload.
  3. The same pipeline, retrained but not re-architected, meets the RQ1 targets in both environments.
Gap in the literature

Situation-awareness research is rich but none of it is paired with sub-second automated interception.

RQ4 · ADDRESSES PROBLEMS 1 AND 2, AT THE LEVEL OF GENERALISATION

Compositional generalisation to unseen hazards

5 papers scoring ≥ 3 →

Can a rejector trained on a finite set of hazard motifs intercept catastrophes formed by motif combinations, agent configurations and neighbourhood depths it has never seen, and what is lost as the sub-graph is truncated from 3 hops to 2 or 1?

Hypotheses
  1. On the compositional test split the rejector retains at least 90% of its in-distribution prevention rate, where stateless and 1-hop baselines degrade substantially more.
  2. Hyperbolic embedding improves compositional generalisation on hazards spanning containment hierarchies.
  3. Prevention falls monotonically with neighbourhood depth, identifying the minimum context the architecture needs.
Gap in the literature

Only one learning-to-defer paper uses a multi-agent state; no JEPA paper is action-conditioned for safety.

RQ5 · ADDRESSES COMMERCIAL VALUE AND POSITIONING AGAINST CERTIFIED SAFETY SYSTEMS

Incremental value over existing safeguards

2 papers scoring ≥ 3 →

What fraction of catastrophic irreversible actions that pass existing safeguards (on-board protective fields, per-transaction rules, approval matrices, change approval) does PreCog intercept before irreversibility, and at what cost in false halts and throughput?

Hypotheses
  1. For the defined hazard classes PreCog prevents a majority of catastrophes the existing safeguard does not catch, in both environments.
  2. False-halt rate alongside the existing safeguard stays below a domain-defined nuisance threshold, and total emergency stops or reversals fall.
  3. Incremental value concentrates in hazards whose causal chain crosses two or more systems or agents.
Gap in the literature

The hazard class is defined as cross-system compositions, which single-node safeguards miss by construction.