Experiment — Gutenberg Curriculum Baseline

Hypothesis (Falsifiable)

Training on curriculum-aligned Gutenberg narrative text will reduce loss steadily without early collapse into repetition, and will improve short prompt continuation coherence relative to a minimal smoke dataset.

Variables

Independent (Changed)

• Dataset: Gutenberg curriculum subset (vs prior minimal samples).

Controlled (Fixed)

• Model architecture (locked via organism/configs/models/*).
• Prompt suite (organism/prompts/*).
• Training loop implementation (organism/training/train.py).

Explicitly Not Tested

• Tokenizer changes beyond the current codec/tokenization module.
• Any RLHF/preference shaping.

Dataset Specification

• Manifest: data/manifests/phase0_curriculum_training.json
• Local staging: data/staging/phases/phase0_curriculum_training/

Runs

• (Add run IDs here after execution.)

Results Summary (1 page max)

(Write after runs complete.)

Interpretation

(Write after results summary.)

Next Actions

• Decide whether the curriculum baseline is good enough to proceed to the next curriculum stage.