Skip to main content

[!FROZEN] MPLP Protocol v1.0.0 Frozen Specification Freeze Date: 2025-12-03 Status: FROZEN (no breaking changes permitted) Governance: MPLP Protocol Governance Committee (MPGC) License: Apache-2.0 Note: Any normative change requires a new protocol version.

Learning Feedback

[!NOTE] Duty Type: OS-Level Kernel Duty
SOT Reference: README v1.0.0 Section 8

Intent​

To ensure that every execution cycle contributes to the system's long-term improvement by capturing outcomes, evaluating performance, and feeding insights back into the model or policy.

Lifecycle Coverage​

  • Execution: Capture traces and results.
  • Post-Execution: Evaluate success/failure.
  • Planning: Use past learnings to improve future plans.

Agent Scope (SA / MAP)​

  • SA: Self-correction and few-shot example accumulation.
  • MAP: Shared learnings across agents, reputation scoring.

Required Events​

  • LearningSampleCreated
  • FeedbackReceived
  • ModelUpdateApplied

Compliance Requirements​

  1. Runtime MUST provide a mechanism to capture LearningSamples.
  2. Runtime MUST link outcomes (Success/Failure) to the original Plan and Context.
  3. Runtime SHOULD support persistent storage of feedback loops.

Implementation Details (Non-Normative)​

Learning Feedback is implemented via the Learning Module, which standardizes how execution history is captured for training.

Learning Sample Structure​

The LearningSample object (learning/mplp-learning-sample-core.schema.json) is the atomic unit of learning data:

  • sample_family: Classifies the type of learning (e.g., intent_resolution, pipeline_outcome, multi_agent_coordination).
  • input: Abstracted representation of the context before the action.
  • output: The actual result or decision made.
  • state: A snapshot of the relevant PSG state (roles, config) at the time of execution.
  • meta: Quality signals, human feedback labels (approved, rejected), and provenance IDs.

This structure allows any MPLP execution trace to be converted into a dataset for fine-tuning or few-shot prompting.

Schema Reference​

SchemaPurposeKey Fields
mplp-learning-sample-core.schema.jsonDefines a learning samplesample_id, input, output, state, meta

Examples​

  • Auto-Correction: An agent fails to generate valid JSON, receives an error, and adds that error to its negative constraints for the next attempt.
  • Golden Flow Extraction: Successful traces are promoted to "Golden Flows" for future reference.