Skip to content

AGENTIC SYSTEM 1 · DUAL-BRAIN ARCHITECTURE · docs

10. Operational Specs & Jagged Edge Catalog

To build resilient, deterministic systems, engineers must understand the exact operational boundaries, latency ceilings, and documented “jagged edges” of the underlying models.

1. Model Roster

Model Tag Identifier Intended Use Deployment Recommendation
laya-mlx Local Edge (Apple Silicon) On-device sub-10ms System 1 inference (MPS / MLX Metal Graph) Primary local engine for macOS workstations ($0 cost)
jev-latest Cloud API (Rolling) Evaluates latest patches and taxonomy updates Recommended for Dev / Staging environments
jev-1.13 Cloud API (Pinned) Guarantees immutable probability distributions and behavior Required for production compliance and regulated pipelines

2. Operational Specifications

  • Latency Profile (P50):
    • Local Edge (Laya-MLX): 6.53 ms
    • Cloud API (TypeSafe JEV): 796.8 ms
    • Cascade Router ($\tau = 0.30$): 243.2 ms (3.41x speedup over pure cloud)
  • Context Token Budgets:
    • Maximum context: ~8,192 tokens.
    • Recommended context: 50 – 4,000 tokens for optimal calibration.
  • Batching Capacity: Scales cleanly across dozens of questions per request ($\le 30$ recommended per single forward pass).

3. Documented Jagged Edges & Remediation

A. Double Negatives in Question Phrasing

  • ⚠️ Symptom: Using double negatives in Noul or Choice criteria (e.g., “This text does not contain content that is not in English”) degrades probability calibration.
  • ✅ Remediation: Always formulate questions affirmatively: “This text is written entirely in English”.

B. Over-Granular Score Criteria

  • ⚠️ Symptom: Supplying 10 to 20 granular levels in a Score prompt dilutes the probability mass across adjacent buckets, inflating entropy.
  • ✅ Remediation: Maintain 3 to 5 well-defined milestone levels (e.g., Low, Medium, High, or 0: None, 1: Moderate, 2: Critical). The engine computes continuous expectation (e.g., $1.42$) with mathematical precision.

C. Unlabeled Tabular Data

  • ⚠️ Symptom: Ingesting raw CSV rows without header rows forces the model to guess column relationships.
  • ✅ Remediation: Serialize tables as arrays of structured JSON objects or valid Markdown tables with explicit column headers.

D. Overlapping Semantic Boundaries in Choice

  • ⚠️ Symptom: Options with ambiguous overlap (e.g., Option A: “Billing defect”, Option B: “Banking issue”).
  • ✅ Remediation: Formulate mutually exclusive boundaries: Option A: “Card processor / Stripe gateway error”, Option B: “Direct wire transfer or ACH routing issue”.