EXP-0002 — Complexity Beats Volume (D0 vs D0+D1)

Status: Draft

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

Hypothesis

If we inject slightly more complex narrative structure (D1) into the baseline (D0) via a controlled mix, then late-stage learning will improve faster than simply training longer on D0 alone, because broader structure reduces saturation and encourages richer internal representations.

Setup / Test Plan

What stays fixed:

• Model config and training loop (same hyperparameters).
• Total training budget held constant across conditions.
• Prompt suite fixed:
  • File: organism/prompts/v1.json
  • prompt_set_id: month1_v1
• Eval cadence and deterministic settings fixed.

Conditions:

• Baseline: D0 only (fairy tales).
• Variant: D0 + D1 mix (weights fixed and recorded).

Data:

• D0: data/staging/phases/phase0a_early-childhood/fairy_tales.jsonl
• D1: data/staging/phases/phase0a_early-childhood/gutenberg_childrens_literature.jsonl
• Manifest (authoritative): data/manifests/month1_manifest_v1.yaml

Runs (example IDs):

• EXP-0002-D0-only
• EXP-0002-D0-D1-mix

Measurements (Pass/Fail)

Primary:

• Late slope improvement:
  • Compare late-slope (70–90%) vs baseline at equal tokens_seen.
• Plateau coefficient:
  • Variant should have a lower plateau coefficient (i.e., less flattening) than D0-only, at equal budget.

Secondary:

• Eval robustness:
  • Lower repetition rates on “play probes” prompts without loss of intelligibility.
• Qualitative behavior:
  • More consistent tense/POV in the “memory + consistency” prompts.

Confounders:

• D1 dataset quality/cleanliness can dominate results; keep D1 small at first (“baby steps”).

Results

Runs executed:

• (fill)

Observed:

• (fill)

Interpretation

• (fill)

Decision

• Adopt / Reject / Iterate
• Next actions: