Skip to content

Honest low-confidence grid display (Stage 10) - #44

Open
robmorgan wants to merge 1 commit into
mainfrom
feat/low-confidence-grid-display
Open

Honest low-confidence grid display (Stage 10)#44
robmorgan wants to merge 1 commit into
mainfrom
feat/low-confidence-grid-display

Conversation

@robmorgan

Copy link
Copy Markdown
Owner

Closes the last code item in ROADMAP Stage 10's non-adopter work: an untrustworthy grid is now visibly untrustworthy.

The problem the naive version couldn't solve

Grid confidence scores internal consistency, not ground truth: Somebody To Love's wandering DP grid measures confidence 0.845 with beat F 0.31 — squarely inside the healthy corpus cluster (0.79–0.94). No display threshold on raw confidence can dim it without dimming correct grids.

Three layers

  1. Library (rigid_grid.rs): when both adoption gates decline a fit — phase_lock < 0.3 AND tracked-beat corroboration < 0.6 — the two independent estimators genuinely disagree, which is positive evidence the surviving tracked grid's phase is suspect on quantized material. refine_grid_rigid now caps reported confidence at 0.5. Verified on corpus: somebody 0.772 → 0.500; hot_stuff/cold_heart (adopted) unchanged; ramps/live material unaffected (sanity-floor declines and adoptions don't touch the cap). Unit-tested via the drifting-tracker control.
  2. Desktop: GridMarks.low_confidence (threshold 0.6, documented against the measured 0.79 healthy-corpus floor); beat/downbeat ticks dim to 40% in both painters; the counter row shows a small "grid: low confidence" hint.
  3. Annotation: somebody-to-love marked phase_indecisive with the 2026-08-07 measurements recorded in its meta.

Also: desktop/Cargo.toml gains an empty [workspace] table so desktop cargo commands work inside git worktrees (cargo otherwise walks past the worktree root and claims the outer repo's workspace; main-checkout behavior unchanged).

Root + desktop checks green (fmt, clippy -D warnings, tests, docs); 2 new desktop unit tests.

Visual check for review: load Somebody To Love on this branch — its markers should draw dimmed with the hint, while Hot Stuff draws full-strength.

🤖 Generated with Claude Code

Three layers make an untrustworthy grid visibly untrustworthy:

- Library: when BOTH rigid-adoption gates decline a fit (phase_lock
  below 0.3 AND tracked-beat corroboration below 0.6), the estimators
  genuinely disagree - positive evidence the surviving tracked grid's
  phase is suspect on quantized material - so refine_grid_rigid caps
  the reported confidence at 0.5. The raw metric scores internal
  consistency, not ground truth: Somebody To Love measured beat F 0.31
  with confidence 0.845. Verified on corpus: somebody 0.772 -> 0.500,
  adopted tracks unchanged. Ramps/live material unaffected (declined by
  the sanity floor, or adopted).

- Desktop: GridMarks carries low_confidence (threshold 0.6, documented
  against the measured 0.79-0.94 healthy-corpus floor); beat/downbeat
  ticks dim to 40% in the zoomed and overview painters and the counter
  row shows a 'grid: low confidence' hint.

- Annotation: somebody-to-love marked phase_indecisive with the
  2026-08-07 measurements recorded.

Also: desktop/Cargo.toml gains an empty [workspace] table so desktop
cargo commands work in git worktrees (cargo otherwise walks past the
worktree root and claims the outer repo's workspace; behavior in the
main checkout is unchanged).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant