data

data/ — Per-subject inputs and generated artifacts

This directory contains the inputs and generated artifacts for every subject in the "Beyond Recall" study. Model outputs, judge scores, retrieval logs, and memory-system ingestion records for each subject live under ../results/, which is the per-subject evidence tree.

For the authoritative numbers the paper reports, see ../docs/DATA_REFERENCE.md. For provenance (which file produced which claim), see ../docs/PROVENANCE_INDEX.md.

Directory layout

data/
├── source_corpora/       Source autobiographies (raw text + provenance) — the inputs the pipeline consumed
│   ├── augustine/        raw.txt + provenance.md per subject
│   ├── babur/
│   ├── ... (16 subjects total: 14 main-study + franklin_autobiography + franklin_letters)
│   └── MANIFEST.md       Top-level index with bibliographic record + SHA-256 hashes
├── global_subjects/      13 low-to-mid pretraining subjects (the gradient sample)
│   ├── augustine/
│   ├── babur/
│   ├── bernal_diaz/
│   ├── cellini/
│   ├── ebers/
│   ├── equiano/
│   ├── fukuzawa/
│   ├── keckley/
│   ├── rousseau/
│   ├── seacole/
│   ├── sunity_devee/
│   ├── yung_wing/
│   └── zitkala_sa/
├── hamerton/             high-depth single-subject study (reference subject)
├── franklin/             known-figure control (high pretraining baseline)
└── franklin_obscure/     Franklin counterfactual (obscure-scenario battery)

The source_corpora/ directory was added 2026-05-05 to mirror the original Project Gutenberg / archive.org source texts (~13 MB total) so the pipeline is reproducible end-to-end from the same inputs the study consumed. Generated by scripts/migrate_source_corpora.py. Per-subject provenance.md records the bibliographic edition, eBook ID, translator/editor, file integrity hash, and pipeline-consumption notes.

The 14 subjects in the paper are Hamerton + the 13 global subjects. Franklin is the known-figure control; franklin_obscure is a secondary counterfactual battery on Franklin with more obscure scenarios.

File schema — global subjects (richest, most uniform)

Each data/global_subjects/<subject>/ contains:

File	Purpose
facts.json	Extracted behavioral facts (47-predicate vocabulary) with provenance back to source text. Input to the specification pipeline and to every memory system.
battery.json	Held-out behavioral prediction battery. Each item has a scenario, a ground-truth excerpt from the held-out half of the corpus, and metadata. This is the test set.
spec.md	Early short-form spec (pre-final-layer).
spec_production.md	The production behavioral specification used in the paper (the "~5,000 tokens" spec).
anchors_v4.md	Anchors layer — axiom-style behavioral anchors authored blind from the facts.
core_v4.md	Core layer — ~800-word behavioral narrative.
predictions_v4.md	Predictions layer — behavioral patterns and decision heuristics.
brief_v5.md	Unified brief composed from the three layers.
judgments.json	Subject-level judge output (per-item scores, rubric).
results.json	Subject-level aggregated results (condition means, CIs, raw responses).

The pipeline stages correspond to files as follows: extract → facts.json, author → anchors_v4.md + core_v4.md + predictions_v4.md, compose → brief_v5.md + spec_production.md.

Schema variation — Hamerton, Franklin, Franklin obscure

Not every subject has every file. This reflects the study history (Hamerton was the deep-dive reference subject, Franklin is a known-figure control).

Subject	facts.json	battery.json	Spec files	Notes
hamerton/	yes (also shared_facts.json)	yes (also questions_80.json)	hamerton/spec/ contains anchors_v4.md, core_v4.md, predictions_v4.md, brief_v5_clean.md	No spec_production.md; use brief_v5_clean.md. Empty analysis/ placeholder.
franklin/	yes (also franklin_shared_facts.json)	yes (also questions_80_franklin.json)	None stored here	Franklin is the known-figure control. Empty analysis/ placeholder.
franklin_obscure/	yes	yes	None	Obscure-scenario counterfactual battery on Franklin.

Canonical filenames vs. legacy aliases

Some files exist under two names (same content, different names kept for scripts that reference the older paths):

• hamerton/facts.json == hamerton/shared_facts.json
• hamerton/battery.json == hamerton/questions_80.json
• franklin/facts.json == franklin/franklin_shared_facts.json
• franklin/battery.json == franklin/questions_80_franklin.json

Prefer the facts.json / battery.json names when reading programmatically.

Where the model outputs live

All condition responses, retrieval logs, memory-system ingestion records, and per-judge judgment files live under ../results/<subject>/ — not here. data/ holds inputs and pipeline-generated artifacts; results/ holds evaluation outputs.