Module blind 

Blindness: label assignment, attribution stripping, and leak detection.

A judge that knows which model wrote a candidate stops grading the patch and starts voting on the model’s reputation. Three things keep that from happening:

1. Labels. Candidates are presented as A/B/C, assigned by a seeded shuffle, and each judge sees them in its own order so position carries no signal either.
2. Stripping. Commit messages and candidate summaries lose their attribution trailers — both at write time (a per-worktree commit-msg hook) and at presentation time (this module). Belt and braces: the hook can be bypassed with --no-verify, the presentation filter cannot.
3. Leak detection. The patch body is scanned for vendor-identifying text. Blanket redaction there would corrupt the artifact under judgement, so the policy is configurable and defaults to recording the leak.

magi itself is the facilitator, which is the structural reason this works: there is no moderator agent that could leak an author, because the moderator is code that never learns anything it does not print.

Structs

Leak

A vendor token found in material shown to judges.

Functions

assign_labels

Label for candidate index i after the seeded shuffle.

commit_msg_hook

The commit-msg hook installed into every candidate worktree.

presentation_order

The order judge j sees the candidates in, as indices into the candidate list.

redact

Replace every vendor token with [REDACTED], case-insensitively.

sanitize_patch

Apply the configured leak policy to a patch body.

sanitize_prose

Sanitize prose written by a candidate (commit messages, summaries).

scan

Count occurrences of each vendor token in text.

strip_attribution

Drop every line containing one of patterns (case-insensitive substring).