What craft-crw decodes a CRW from, and the evidence that it is right. The full record was produced in LightCraft (clean-room/crw/, branch clean-room-crw-format-decode, revision 6ca439b5155820d2befeacd81ec0c6a80b2bd3bc; the record is identical at 70c755e4, where the decoder was ported) between 2026-10-12 and 2026-10-25; the documents below moved here verbatim when the decoder became this shared crate, and the derivation trail stays at that revision. The documents are the evidence: they are not rewritten after the fact, so paths they mention are LightCraft's paths of that time (the mapping is below).

The derivation trail and frozen table-generation materials remain in LightCraft PR #3021. This crate retains the format specification, verification summary, colour metadata notes, source log and public-facing decoder documentation.

Documents

File	What it is
`sources.md`	Every URL opened and search run, with the time, what was asked and what was taken. The rules in force are at its top.
`02-crw-format.md`	The frozen format (revision 2) and its reopening rule. The decoder implements this document.
`03-verification.md`	Full-frame exact comparison against the reference, zero tuning: 28 of 29 files, then 1,516 of 1,516 D60 files.
`04-color.md`	CFA phase, active area, black and white levels, white-balance records, orientation and time: what the decoder reports beyond the samples. §3's colour matrices concern LightCraft (below).
`05-private-corpus.md`	A summary of private-corpus decoder verification; no image data or identifying file list is included.

What stayed local

The Python reference model and probe harness (plan/crw/model/ in LightCraft, gitignored), by the maintainer's decision #4. Its findings are the entropy-coding derivation, which is not part of this record.
The reference decoder: LibRaw 0.22.2, official pre-built Windows binaries installed by the maintainer; only unprocessed_raw.exe was ever run, as a black box. The source tree shipped alongside it was never opened, read or searched.
Canon's CIFF specification PDF (copyrighted; archive URL and SHA-256 in sources.md).
The reference TIFFs (<name>.crw.tiff) the verification compared against; CRAFT_CRW_ORACLE points the test at a local copy.

Path mapping (LightCraft at the time → craft-libs)

In the documents	Now
`clean-room/crw/…`	`crates/crw/clean-room/…`
`crates/raw/src/vendor/crw.rs`	`crates/crw/src/lib.rs`
`crates/raw/src/vendor/crw_tables.rs`	`crates/crw/src/tables.rs`
`crates/raw/tests/crw_oracle.rs`	`crates/crw/tests/samples.rs` (`CRAFT_CRW_ORACLE`, formerly `LIGHTCRAFT_CRW_ORACLE`)
`white_from_data` (LightCraft's shared estimate)	`saturation_plateau` in `crates/crw/src/lib.rs`
`xtask/src/main.rs` `RAW_SAMPLES` (corpus URLs)	`crates/crw/tests/samples.json` + `scripts/fetch-crw-samples.py`
`plan/crw/model/gen_tables.py`	LightCraft PR #3021
`docs/arw6-compression.md` (the house model for a derivation)	LightCraft's document; the method is summarised in `docs/clean-room.md` of this repository
`crates/raw/src/chart_table.rs`, `spectral_table.rs`, `camera_preview`, `assets/camera-profiles/`	LightCraft (colour stays in the app)

Rules that were in force

From sources.md: no raw-decoder source may be opened, fetched, searched inside or requested (dcraw, LibRaw, rawspeed, rawloader/rawler, dnglab, darktable, RawTherapee, libopenraw, ExifTool's Perl code, including licence headers, comments and mirrors). Allowed: published format documentation in prose, published standards, the sample files, and a reference decoder's binary run as a black box with every probe logged. No tuning after the freeze: a mismatch reopens the derivation explicitly, with its cause and a new freeze commit recorded (02-crw-format.md, revision 2, is the one reopening). The general standard this followed is docs/clean-room.md.