Powered by

# DoD Agents

Distillation-on-Demand (DoD) agents are lightweight runtimes that live on user devices or edge servers. They detect knowledge gaps, run local reasoning and trigger new knowledge distillation when needed.

# Core Responsibilities

  • Gap detection
    Spot missing or outdated facts by monitoring local queries, cache misses and confidence thresholds.

  • Selective Contextual Reasoning
    Retrieve the smallest relevant context from local and gateway caches, then run on-device inference to answer immediately where possible.

  • Distillation orchestration
    When local reasoning is insufficient, compose a distilled prompt for a Big LLM, gather responses and synthesize the best answer.

  • Local self-training
    Use idle cycles (for example at night) to perform QLoRA fine-tuning, baking freshly distilled knowledge into the Tiny LM.

  • Privacy guard
    Ensure only distilled, context-safe snippets leave the device; raw private data never reaches a gateway.

# Workflow

  1. Receive user query.
  2. Attempt local answer via SCR and Tiny LM.
  3. If confidence is low, call Big LLM and aggregate multiple responses.
  4. Evaluate candidates with a Tiny reward model and choose the best.
  5. Forward distilled knowledge to a gateway for validation and caching.
  6. Schedule nightly QLoRA to update local weights.

# Incentives and Rewards

  • Earn token bounties by completing gateway distillation tasks first.
  • Save on external LLM costs through cache hits and local inference.
  • Higher reputation unlocks priority tasks and larger rewards.

# Resource Requirements

Component Minimum Spec Notes
CPU or GPU 4 cores / 6 GB VRAM Runs Tiny LM inference
Storage 10 GB Local KV cache and model deltas
Connectivity Intermittent Only needed for gateway sync

# Interaction with Gateways

DoD agents push distilled answers to gateways along with confidence scores and minimal context. Gateways validate, cache and reward the agent upon acceptance.

# Security and Trust

  • Agents sign every distillation packet with a local key pair.
  • Repeated low-quality submissions reduce reputation and throttle DoD call limits.
  • All local private data remains encrypted at rest; only distilled summaries are transmitted.

DoD agents turn passive edge devices into active knowledge contributors, accelerating network learning while protecting user privacy.