SKILL.md
# paper
A paper here is prose, its sources, the log of how each source was checked, and the marks that
say what was checked. The work documents itself as it documents its subject.
## The loop
1. **Read** what exists first: the design (a MIP), the notes and the code.
2. **Evaluate** each thing yourself: fetch it, run it, or mark it as not checked.
3. **Adopt** only what passed, with credit to a DOI or a commit.
4. **Write**, with a mark on each claim.
5. **Log** how each source was checked.
6. **Say** what is not known.
## Marks
- `(v)`: checked while writing. The log says how.
- `(v, title only)`: the record of the source was read, but not its text.
- `(MIP §n)`: the claim comes from that place of a design.
- `⚠`: not checked.
A mark belongs to its words. When you change the words, do the check again. A mark from an
earlier session tells you where to look. It is not a check.
"The source does not say it" is not "it is false". Write which of the two you found.
## Rules
- **Never invent a reference.** Copy the title, authors, venue and year from the registry
record (Crossref, DataCite, arXiv). A snippet, a summary, your memory or the reference list of
another paper is not a reading of the source.
- **Each number has its command** or its source, with its unit. A number copied by hand is `⚠`.
- **No result before parity.** Show that a thing is correct before you say that it is faster.
Never widen a tolerance to get a pass.
- **Keep facts apart from analysis.** Say which sentence has a source and which is your reading.
- **The design wins.** If the paper and its MIP do not agree, change the paper, or ask.
- **A private source stays private.** Say that it exists. Do not copy its content.
## Files
```text
<research>/
README.md what the research is, its rules, its status
paper/paper.md front matter (title, author, status: draft, version, date), then the prose
paper/refs.bib each reference, with a DOI where one exists
paper/REFERENCES.md one row for each entry: how it was checked, what the paper takes from it
notes/ one idea in one note
journal/ one entry for each problem that the work met
```
The text starts with a block that says what the text is and how to read the marks. The last
sections are "What is not known" and "Where the thesis comes from".
## Steps
1. Find the design. If a MIP exists, the paper is an account of it.
2. List the claims before the prose. Give each one its kind: a fact about a source, a
description, a cause, a forecast, a design choice or a claim of being first. A cause needs a
measurement. A claim of being first is narrowed to what was examined, and when.
3. Check each claim. A reference passes three checks: it exists, it is the work that you mean,
and it says what you say. The verdict is *verified*, *mismatch* or *not found*. Write the row
of `REFERENCES.md` as you go.
4. Write the body. Short sentences, active voice, numbers and not adjectives. Use the `ste100`
skill for English and the `cientifico` skill for pt-BR.
5. Remove or mark each claim that you could not check. Do not soften it.
6. Write the ends last: "What is not known", then the summary and the abstract.
7. Do the mechanical pass. A number in two places is the same in both. One word names one thing.
After a style edit, compare the facts with the earlier version. Render the PDF and read it.
8. Give the rendered paper, and nothing else, to a cold reader (a fresh agent). Ask five to ten
questions. Ask for faults: where, the quoted words, the rule, why it matters, what to do.
9. Open a PR. Paste what each check printed. List the decisions that belong to the owner.
## Release
Use the `cite-release` skill. An agent never makes the tag.
## Easy version
Use the `easy-translate` skill. Keep each claim and each mark. Add no fact.