Add a source

Point the extractor at a book other than Escoffier.

Check the licence first

Only sources you may redistribute go in corpus/. Anything published in the United States through 1930 is public domain, which covers most historical cookbooks worth parsing. Project Gutenberg texts are already cleared, and carry a licence wrapper the reader strips automatically.

If a source cannot be committed, it does not belong in corpus/ — write a fetch script instead, and keep the file untracked.

Add the text

$ curl -sL https://www.gutenberg.org/cache/epub/65061/pg65061.txt \
    -o corpus/farmer-1896.txt

Register it

In src/saucier/infrastructure/config.py, add the work name, the origin, and a Paths property for the file:

FARMER = "farmer"
FARMER_ORIGIN = "Project Gutenberg 65061"


@property
def farmer(self) -> Path:
    """The Fannie Farmer source text."""
    return self.corpus / "farmer-1896.txt"

The work name has no year in it. The edition year is read from the document's front matter, and the two together make the source_id. See ADR-0009.

Then add a factory in bootstrap.py, returning the port rather than the concrete class:

def farmer_source(paths: Paths | None = None) -> SourceText:
    resolved = paths or Paths.discover()
    return GutenbergText(
        path=resolved.farmer,
        work=FARMER,
        origin=FARMER_ORIGIN,
        fidelity=Fidelity.TRANSCRIPTION,
    )

fidelity is stated here because it is a fact about acquisition. A proofread transcription is Fidelity.TRANSCRIPTION. A machine-read scan is Fidelity.OCR, and every record the source yields carries that value.

Check that the source states an edition

$ uv run python -c "
from saucier.infrastructure.bootstrap import farmer_source
print(farmer_source().witness.source_id)
"

read_edition raises EditionUnstated when the head of the file names neither an edition nor a copyright year. That is an expected failure, not a bug. Rename nothing to work around it. A source with no stated identity needs a front-matter rule that reads whatever it does state.

Check whether the entry pattern fits

extract raises NoPreparationsFound when a source numbers its entries differently. It raises the same error when the entries parse but none of them is a sauce. Both are expected failures, not bugs. Escoffier's 22—BROWN SAUCE OR ESPAGNOLE form is specific to him.

$ uv run python -c "
from saucier.infrastructure.bootstrap import farmer_source
from saucier.services.extraction import extract
print(len(extract(farmer_source()).preparations))
"

If that raises, the source needs its own ENTRY pattern. Do not loosen the existing one to fit both. A looser pattern silently admits garbage from the source it was already working on.

A source that does not divide itself into titled chapters has no sauce chapter, so sauce_chapters returns nothing and only headings that say "sauce" qualify. That resolves less, and it is right. See ADR-0007.

For a non-Gutenberg source

PlainText already reads a file that carries no wrapper, which covers the Internet Archive _djvu.txt form. For a third format, write a driven adapter beside gutenberg.py satisfying SourceText: a witness property, a line_offset property, and a lines() method returning body lines with the format's packaging removed. line_offset is the count of file lines the adapter stripped before the body, so a recorded line number names a line in the file. Nothing above the adapter layer changes.

Do not loosen GutenbergText to accept a file with no Gutenberg markers. That refusal is what tells you a second adapter is needed.

For a scanned source

Wrap it in NormalisedText, which collapses double spacing and removes the space before punctuation. Without it the chapter and mothers patterns fail on whitespace alone.

NormalisedText(inner=PlainText(path=..., work=..., origin=..., fidelity=Fidelity.OCR))

The wrapper never repairs a character. velout^ stays velout^, because repairing one witness toward another manufactures agreement between them. See ADR-0011.