M2.5: the reality gate passes — 33.0% empty verdicts on 32 real Logic projects

[sepehrsadri]

What and why

Issue #15's exit criterion was a decision, not a tool. If >50% of real Logic saves show no visible structural change, stop and map the volume-fader field before building M5 — because no UI copy survives a dominant "nothing changed" first impression. The tool shipped in #32; the measurement it exists to produce needed a real library and had never run. The ROADMAP and EXPERIMENTS.md §11 both still said the finding was open, with an n=1 (55.6%) data point standing in for it.

It has now run at the scale the issue specified.

Measured — 2026-08-16, 32 projects / 28 GB / 100 consecutive save pairs:

• 33.0% of save pairs show no structural change Wit can see (67.0% do)
• 28.0% of pairs are byte-different but structurally identical — M2's finding, reconfirmed at library scale
• Distribution is long-tailed, not bimodal: the mode above zero is a single change (12 pairs), 26 pairs sit in the 1–7 band, a thin tail runs to 235

33.0% is below the 50% threshold, so the gate passes and M5 proceeds. #18 is unblocked. Every ProjectData in the library walked cleanly — the report printed no read-error line — so all 100 pairs are real comparisons rather than partial ones.

The M2 volume-fader stretch goal stays worth doing: a save that moved only a fader sits in that 33%, and "no structural change Wit can see" remains a statement about the Structure honesty tier, not about the music. It just no longer has to come before M5.

For calibration: Ableton's semantically-empty rate is 24% (§1). Logic's 33% is higher, as expected from a narrower tier — but not the majority outcome the gate was written to catch.

Also closes M3's flagged wit dupes follow-up

M3 landed with wit dupes verified only against a deliberately-planted duplicate; the README's ~5.4 GB figure needed the full library. Run against it now, the Rust implementation reports 5.5 GB of 23.9 GB (23.2%) against §9's independent Python walk at 5.38 GB of 21.95 GB (24.5%) — agreement on a library that gained ~2 GB of audio in between. Recorded as a cross-check in §9 rather than as a rewrite of it, since they are different tools on different snapshots.

One bug, found while reconciling those two numbers

human_bytes divided by 1024 while printing decimal unit labels ("GB"). That understated its own totals by 7.4% and made wit dupes output silently incomparable to docs/EXPERIMENTS.md, which quotes decimal GB throughout and says so in §9's limits. Now decimal, with a test pinning the boundary.

Checklist

• Docs updated in this PR if behaviour changed
• No audio or project files committed (see .gitignore)
• Claims are labelled measured / cited / inferred

If this PR changes a number in docs/

Material: one person's real working Logic library — 32 projects, 28 GB, 100 consecutive save pairs. macOS, Apple Silicon. Real material, never committed; CI cannot reproduce any of this.

$ cargo run --release -p wit-cli -- logic-report ~/Music/Logic
  scanned 32 project(s), 32 alternative(s), 100 consecutive save pair(s)
  67.0% of save pairs show a structural change Wit can see (67 of 100)
  distribution of change counts per save pair (0 = no visible structural change):
    0 change(s): 33 pair(s)
    1 change(s): 12 pair(s)
    2 change(s): 4 pair(s)
    3 change(s): 6 pair(s)
    ... (full distribution in EXPERIMENTS.md §11)
    235 change(s): 1 pair(s)
  28 pair(s) (28.0%) are byte-different but structurally identical

$ cargo run --release -p wit-cli -- dupes ~/Music/Logic
  found 5.5 GB of duplicate audio (23.2% of 23.9 GB scanned)

$ WIT_LOGIC_LIBRARY=~/Music/Logic cargo test -p wit-index --test real_fixtures -- --nocapture --ignored
  test real_library_reports_the_three_m2_5_statistics ... ok
  test result: ok. 1 passed; 0 failed

Gates:

$ cargo test --workspace        127 passed, 0 failed
$ python3 -m pytest tests/ -q   293 passed, 13 skipped
$ cargo fmt --all --check       clean
$ cargo clippy --workspace --all-targets --locked -- -D warnings   clean
$ cargo deny check licenses advisories                             ok

Limits, stated plainly

One person's library, one machine — the same breadth weakness every number in this repo carries, and exactly what #4 exists to fix. It is a library that has never been used with a version-control tool, so its cadence reflects Logic's own backup behaviour rather than deliberate committing. And 33.0% is a rate over save pairs, not sessions: a project with 20 backups contributes 19 pairs and outweighs one with 2.

Closes #15. Unblocks #18.

🤖 Generated with Claude Code