Skip to content

Enhancement: Automatic Rolling Reservoir & Prompt Retirement #58

@sharmaanchita

Description

@sharmaanchita

Problem
The system lacks an automated mechanism to manage prompt lifecycle (active → retired → public). Currently, prompts persist indefinitely unless manually revealed, undermining contamination resistance.

Basis of issue

  1. Automatic prompt retirement based on:
    • Age
    • Low sampling weight
    • Exposure thresholds
  2. Fixed-capacity rolling reservoir (k) for live prompts
  3. Automated state transitions: secret → retired → public
  4. Longitudinal contamination archive for retired prompts

Importance of issue

  1. Core requirement for contamination-resistant benchmarks (per paper)
  2. Prevents benchmark staleness and training leakage
  3. Maintains discriminative power over time

Current Implementation Gap

  1. isRevealed flag exists in DB
  2. No automated triggers or lifecycle orchestration

Implementation checklist

  1. Background job enforces reservoir capacity
  2. Prompts auto-retire based on defined policies
  3. State transitions are logged and auditable
  4. Retired prompts are excluded from live evaluation

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions