35 added, 27 removed. Audit A to A.
# sci-paper plugin — project working directory instructions
## What this project is
- A standalone Claude Code plugin for scientific paper writing + review,
- extracted and generalized from the `weak-gravitational-lensing` project's
- `.claude/skills/{paper, paper-review, figure-review}` skills, plus a new
- `paper-style` skill that adds **corpus-driven style distillation** (RAG-for-
- style; not actual model fine-tuning — Anthropic does not offer Claude
- fine-tuning, see EVALUATION.md).
+ A standalone Claude Code plugin for scientific paper writing, rewriting,
+ review, figure review, adversarial critique, narrative analysis, and research
+ ideation. `docs/SCIPAPER_STANDARD.md` is the single normative writing standard.
+ Corpus profiles, structural measurements, UID evidence, and learned
+ field-similarity models are descriptive or calibrational evidence; they do not
+ create an authorship verdict or a separate paper gate.
## Working principles for THIS project
- 1. **The three ported skills are v0.** They contain weak-gravitational-lensing
- project specifics (NFW, ACSDM, S/N maps, kappa, escnn library refs).
- Generalization is a separate task — do NOT silently rewrite them while
- working on something else. Mark project-specific anchors with `[WGL]`
- comments when you find them.
+ 1. **One normative contract.** Writing and review skills must implement
+ `docs/SCIPAPER_STANDARD.md` and `sci-paper.feedback.v1`. Do not introduce a
+ competing consequence vocabulary, universal prose PASS/FAIL gate, or
+ authorship claim. Field-specific scientific guidance may remain explicitly
+ marked `[WGL]`; do not silently generalize domain facts while changing the
+ shared writing contract.
- 2. **`paper-style` is the headline feature.** The corpus is user-supplied and
+ 2. **Corpus evidence is read-only input.** The corpus is user-supplied and
copyright-sensitive; never commit PDFs from `style-corpus/` (`.gitignore`
- already excludes them). Treat any contents of `style-corpus/<field>/tier-*`
- as read-only inputs; write only to `style-profile/<field>/`.
+ excludes them). Treat `style-corpus/<field>/tier-*` as read-only inputs;
+ generated evidence belongs under `style-profile/<field>/`. Corpus statistics
+ cannot redefine consequence classes or prove authorship.
- **Field-aware layout.** v0.1 ships only `wgl`. Tools/skills auto-detect a
- single-field corpus; multi-field corpora require explicit `--field <name>`.
- When extending, add subdirs under `style-corpus/<new-field>/{tier-1-top,
- tier-2-mentor,tier-3-reference}/` and run
- `python tools/extract_style.py --field <new-field>`.
+ **Field-aware layout.** Never assume a specific field exists. Tools/skills
+ auto-detect only when exactly one field is present; multi-field corpora
+ require explicit `--field <name>`. Add new sources under
+ `style-corpus/<new-field>/{tier-1-top,tier-2-mentor,tier-3-reference}/` and
+ run `python tools/extract_style.py --field <new-field>`.
- 3. **Honest about limits.** When asked "can we fine-tune Claude?", answer
- "no, but here's what we can actually do" — see EVALUATION.md for the
- reasoning. Do not pretend the LoRA path works.
+ 3. **Honest measurement states.** Missing calibration, optional dependencies,
+ complete-document corpora, or human labels remain `unmeasured` or
+ `degraded`. Never convert unavailable evidence into zero findings. Current
+ metrics and limitations belong in `EVALUATION.md`, not the normative standard.
4. **Plugin manifest format is verified.** `.claude-plugin/plugin.json` per
the official Claude Code plugins reference; skills live at
`skills/<name>/SKILL.md` (NOT inside `.claude-plugin/`). When invoked
from another project via `--plugin-dir`, skills are namespaced as
`/sci-paper:paper`, etc.
## Cross-project references
- **Source project:** the sibling `weak-gravitational-lensing` project
(sci-paper was extracted from its `.claude/skills/{paper, paper-review,
figure-review}`). Those existing skills are the v0 inputs to this plugin.
Record your local clone path in `CLAUDE.local.md` (gitignored) if you
want an absolute reference handy.
- **User's global writing-quality memory** lives in the user-level
`~/.claude/CLAUDE.md` (Linux/macOS) or `%USERPROFILE%\.claude\CLAUDE.md`
(Windows). The rule that matters here, verbatim:
> **Paper writing: never quote from memory.** Every number, date,
> coefficient, figure caption, and citation in a manuscript edit must
> be re-read from its source in the same turn it's pasted.
## Toolchain
Python ≥ 3.11 on PATH (developed against 3.13; nothing in `tools/` uses
3.12-only syntax).
- Required for the full feature build (pinned in `requirements.txt`):
- - `pymupdf` — PDF text extraction
- - `sentence-transformers` — exemplar retrieval embeddings
- - `scikit-learn` — optional sentence classifier
- - `regex` (the third-party one, not stdlib `re`) — Unicode-aware sentence segmentation
+ Optional capabilities are listed in `requirements.txt`:
+ - `pymupdf` — PDF corpus extraction and compiled-page inspection
+ - `sentence-transformers` — exemplar retrieval and embedding features
+ - `scikit-learn` + `joblib` — legacy and learned field-similarity models
+ - `transformers` + `torch` — optional token-surprisal / UID measurement
+ - `numpy` — learned feature, cache, and rewrite-score utilities
+
+ The shared schema, deterministic L0 linter, model-free L1/L2 axes, and validator
+ remain standard-library paths. Missing optional dependencies must stay visible as
+ `unmeasured` or `degraded`.
## Personal / machine-specific notes
Put paths, virtualenv locations, or anything else that shouldn't go to
the public repo in `CLAUDE.local.md` next to this file. That filename
is gitignored on purpose — Claude Code reads both, but only this file
is shared.