EXP-0004 — Guided Choice / Preference Signal (Adaptive Sampling)

Status: Draft

Created: 2025-12-25 Last Updated: 2025-12-25

Hypothesis

If we allow the organism to bias its dataset sampling using a simple adaptive rule, then training and behavior will diverge measurably from fixed mixing weights, because the organism’s learning dynamics will express a stable preference signal (as measured by loss improvement and drift in outputs).

Setup / Test Plan

What stays fixed:

• Dataset pool (D0, D1, optionally D2 when ready).
• Total training budget.
• Prompt suite: organism/prompts/v1.json (prompt_set_id: month1_v1).

What changes:

• Sampling policy:
  • Control: fixed weights.
  • Variant: adaptive weights using one explicit rule (choose only one):
    • Upsample dataset with best recent loss delta, or
    • Target a mid-entropy / non-collapse band (requires entropy proxy support).

Measurements (Pass/Fail)

Primary:

• Dataset learning curves:
  • Compare per-dataset dataset_loss_ma improvement rate and resulting weights over time.
• Behavioral drift:
  • Tag changes in style/content in eval prompts.

Secondary:

• Overall late-slope and plateau coefficient vs control.

Results

Runs executed:

• (fill)

Observed:

• (fill)

Interpretation

• (fill)

Decision

• Adopt / Reject / Iterate
• Next actions: