Language: English | 日本語

doctrine-corpus

A DOI-targeted dataset that encodes the documented judgment of the shimo4228 research program as a Q&A corpus optimized for LLM training. The corpus aggregates ADRs, theses, glossaries, and structured essays across four research lines — Agent Knowledge Cycle, Contemplative Agent, Agent Attribution Practice, and Authorship Strategy — into bilingual (English + Japanese) judgment-eliciting Q&A pairs, with metadata that preserves source attribution and line identity per example.

This repository is the operational form of Authorship Strategy Layer 4 tactic 7 (LLM-first ingest). It positions the documented judgment of the research program in the substrate that increasingly mediates how future researchers trace causation: LLM training data.

Why a corpus, not a LoRA

A prior prototype (disposition-lora, 2026-05-06; see docs/empirical/README.md for the retrospective) fine-tuned a LoRA adapter on 230 chunked completion examples from Zenn articles and AKC/AAP ADRs. The retrospective is explicit: voice transferred, judgment did not. The artifact behaved like a "shimo4228-style mannerism wrapper," not a "shimo4228 judgment oracle."

The pivot encoded in this repository:

• Primary artifact = the corpus, not the LoRA. The corpus is base-model-independent; a LoRA is bound to a specific base architecture and depreciates as base models turn over. See ADR-0001.
• Q&A is judgment-eliciting, not chunk-as-completion. The Q states a situation; the A applies the line's framework (three-axis inversion, six-phase cycle, four contemplative axioms, four Business AI Quadrants) explicitly.
• Bilingual pairs by default. Each judgment example exists as an English pair and a Japanese pair, consistent with Authorship Strategy's bilingual diffusion convention.
• Per-example metadata. {line, source, lang, shape} is recorded on every line so downstream consumers can ablate, weight, or filter.

Intended uses and current limitations

The corpus is designed to be used as an ingredient, not as a standalone fine-tuning package.

Verified for:

• RAG retrieval material — bilingual judgment Q&A pairs work well as retrieved context for downstream LLM applications.
• Mixing into larger instruction-tuning corpora — 851 examples is a small but high-density addition to FLAN/Alpaca-scale mixes.
• Human reading material — the corpus is designed to be readable end-to-end as a reference text for the research program's judgment patterns.

Not yet verified:

• Standalone small-LoRA fine-tuning. A Stage D verification LoRA (Qwen3-8B-4bit, 300 iter on 851 examples) was trained as a disposable probe and reproduced the same "mannerism wrapper" failure mode as the Phase 0 prototype — voice transferred, judgment did not. The corpus is base-model-independent and is retained as the deliverable; the LoRA was never a deposit target. See ADR-0005 for the verdict and v0.2.0 hypotheses (data scale, shape distribution, base model effect).

Content scope (v0.1.0)

851 examples in total, distributed across the four research lines and one source of published articles:

Line	Examples
Agent Knowledge Cycle (AKC)	38
Contemplative Agent	318
Agent Attribution Practice (AAP)	146
Authorship Strategy	126
Zenn articles	222
Cross-line (shape: contrast)	1
Total	851

By language: 334 English / 517 Japanese. By shape: 343 judgment / 371 explain / 136 definition / 1 contrast.

The train/valid split is 766/85 (seed=42, val-fraction=0.1).

Attention, Not Self (the fifth research line) is excluded from v0.1.0 as its content is still a stub. It will be added in v0.2.0+ when material accumulates.

Repository layout

doctrine-corpus/
├── corpus/
│   └── v0.1.0/
│       ├── pilot.jsonl     # 5 hand-written precedent pairs
│       ├── train.jsonl     # 766 examples (seed=42)
│       ├── valid.jsonl     # 85 examples (seed=42)
│       └── manifest.json   # version, counts, source commits, verification verdict
├── docs/
│   ├── adr/                # design ADRs (0001..0005)
│   ├── empirical/          # Phase 0 retrospective (disposition-lora findings)
│   └── CODEMAPS/
├── scripts/                # extraction + dataset builder + LoRA verification
└── eval/                   # prompt bank + base vs adapted comparison

See docs/CODEMAPS/architecture.md for the file-level guide.

How to consume

The corpus is distributed via:

• GitHub (canonical): this repository — full working tree, scripts, eval, ADRs.
• Zenodo (DOI archive): concept DOI + per-release version DOI. Citable via CITATION.cff.
• Hugging Face Datasets (mirror): Shimo4228/doctrine-corpus — corpus payload + manifest + README only. Auto-converted to Parquet for direct pandas / Polars / datasets load.

The HF mirror is the recommended ingest path for LLM training pipelines and RAG ingestion. The GitHub repository is the recommended source for citing or auditing the generation process.

License

• Corpus data (corpus/**): CC0 1.0 (public domain dedication) — see LICENSE
• Scripts (scripts/**) and docs: MIT — see scripts/LICENSE
• Source attribution: each example records meta.source pointing to the originating ADR / section / article in the upstream research line. Upstream content remains under each line's own license; the Q&A reformulations are this repository's derivative work and are released CC0 to maximize diffusion (per Authorship Strategy ADR-0015, license selection by audience).

Sibling and parent repositories

Repository	Relation
authorship-strategy	Parent line. This corpus is its Layer 4 tactic 7 implementation.
shimo4228 (hub)	Research program hub. Human-facing index of the five research lines.
agent-knowledge-cycle	Source line. ADRs + glossary harvested into corpus.
contemplative-agent	Source line. ADRs + glossary harvested into corpus.
agent-attribution-practice	Source line. ADRs + thesis + glossary harvested into corpus.
zenn-content	Source. Published articles harvested into corpus.
base-model-lab/experiments/disposition-lora (private)	Phase 0 prototype. Reusable extraction scripts; the failure mode it documented motivated this repository's design. See docs/empirical/README.md for the retrospective.

Citation

Cite using the concept DOI (which always resolves to the latest version): 10.5281/zenodo.20337008. The canonical citation form is in CITATION.cff, also available as codemeta.json (schema.org-based dataset metadata, read by the Software Heritage indexer).

Reference shape:

Shimomoto, T. (2026). doctrine-corpus: A judgment-eliciting Q&A corpus across the shimo4228 research program. Zenodo. https://doi.org/10.5281/zenodo.20337008

For reproducibility citation of a specific historical version, use the v0.1.0 version DOI: 10.5281/zenodo.20337009.