pandoras-box/monkey
8
0
Fork 0
Code Issues Pull requests Projects Releases 2 Packages 1 Wiki Activity Actions

feat(classify): gate halftone-scan to a peak band (MK-7) #28

Merged
nrupard merged 3 commits from feat/classify-halftone-ceiling-MK-7 into main 2026-05-21 18:01:30 +02:00
Conversation 0 Commits 3 Files changed 2 +88 -7
nrupard commented 2026-05-21 17:46:21 +02:00
Member
Copy link

What

MK-7: gate halftone-scan to a peak BAND instead of just a floor. Adds a configurable halftone_max ceiling; the halftone score is now soft_ge(peak, halftone_min) * soft_le(peak, halftone_max) * low_sat. Scoring change only, no change to the autocorrelation detector.

Changes

  • halftone_max field in FeatureThresholds + classify.toml, same XDG override path as the rest.
  • Halftone scoring uses both floor and ceiling.
  • Synthetic halftone_dots fixture previously autocorrelated at exactly 1.0 (a digital screen, not a scan) which the ceiling correctly rejects. It now jitters a fraction of dot cells so the peak lands at ~0.80, inside the realistic 0.70-0.95 band, and still classifies as halftone-scan (conf >= 0.5).
  • New test ceiling_rejects_perfect_periodic_grid: the un-jittered grid (peak ~1.0) is kept out of halftone-scan.

Acceptance status

  • halftone_max loads from classify.toml alongside halftone_min.
  • Halftone scoring uses both floor and ceiling.
  • Synthetic classifies_halftone still passes (now with conf >= 0.5).
  • Signature-scanned PNGs stop classifying as halftone-scan — blocked on fixtures (see below).
  • Real halftone fixtures classify as halftone-scan with conf >= 0.5 — blocked on fixtures.
  • Real-halftone unit test under tests/fixtures/classify/halftone-scan/ — blocked on fixtures.

Blocked / needs input

The default band (0.55..0.95) is a placeholder. The observed signature-scan peak from MK-6 validation was 0.641, which is above the 0.55 floor, so the current band does not yet reject it. Excluding it needs the floor raised (~0.68), but the real-halftone low-end peak is unknown without the user's fixtures, so I have not guessed the final values.

To finish: please drop the real halftone scan(s) (newspaper/magazine) and the signature-scanned PNG into the repo. I will measure their peaks, set the band to admit real halftone and reject the signature, and add the checked-in fixture test. Tuning lands as a follow-up commit on this branch.

🤖 Generated with Claude Code

## What MK-7: gate `halftone-scan` to a peak BAND instead of just a floor. Adds a configurable `halftone_max` ceiling; the halftone score is now `soft_ge(peak, halftone_min) * soft_le(peak, halftone_max) * low_sat`. Scoring change only, no change to the autocorrelation detector. ## Changes - `halftone_max` field in `FeatureThresholds` + `classify.toml`, same XDG override path as the rest. - Halftone scoring uses both floor and ceiling. - Synthetic `halftone_dots` fixture previously autocorrelated at exactly 1.0 (a digital screen, not a scan) which the ceiling correctly rejects. It now jitters a fraction of dot cells so the peak lands at ~0.80, inside the realistic 0.70-0.95 band, and still classifies as halftone-scan (conf >= 0.5). - New test `ceiling_rejects_perfect_periodic_grid`: the un-jittered grid (peak ~1.0) is kept out of halftone-scan. ## Acceptance status - [x] `halftone_max` loads from `classify.toml` alongside `halftone_min`. - [x] Halftone scoring uses both floor and ceiling. - [x] Synthetic `classifies_halftone` still passes (now with conf >= 0.5). - [ ] Signature-scanned PNGs stop classifying as halftone-scan — **blocked on fixtures** (see below). - [ ] Real halftone fixtures classify as halftone-scan with conf >= 0.5 — **blocked on fixtures**. - [ ] Real-halftone unit test under `tests/fixtures/classify/halftone-scan/` — **blocked on fixtures**. ## Blocked / needs input The default band (0.55..0.95) is a placeholder. The observed signature-scan peak from MK-6 validation was 0.641, which is **above** the 0.55 floor, so the current band does not yet reject it. Excluding it needs the floor raised (~0.68), but the real-halftone low-end peak is unknown without the user's fixtures, so I have not guessed the final values. To finish: please drop the real halftone scan(s) (newspaper/magazine) and the signature-scanned PNG into the repo. I will measure their peaks, set the band to admit real halftone and reject the signature, and add the checked-in fixture test. Tuning lands as a follow-up commit on this branch. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
feat(classify): gate halftone-scan to a peak band with halftone_max ceiling (MK-7)
All checks were successful
Check / fmt + clippy + build + tests (pull_request) Successful in 17s
Details
a1b50bc76a 
Add a configurable `halftone_max` ceiling so the halftone-scan class fires only when the autocorrelation peak falls inside the [halftone_min, halftone_max] band, not merely above a floor. The floor rejects photos and gradients (peak < 0.3); the ceiling curves the score back down past true screened-print levels, so a flawless digital periodic pattern (peak ~1.0) is no longer mistaken for a scanned halftone. Scoring change only: the autocorrelation detector is untouched.

`halftone_max` loads from classify.toml alongside `halftone_min`, with the same XDG override path. Default band is 0.55..0.95, documented as a starting point pending real halftone fixtures.

The synthetic `halftone_dots` test fixture previously autocorrelated at exactly 1.0 (a digital screen, not a scan), which the ceiling correctly rejects. It now jitters a fraction of dot cells so the peak lands at ~0.80, in the realistic 0.70-0.95 band a real halftone scan produces, and still classifies as halftone-scan with confidence >= 0.5. A new test asserts the perfect (un-jittered, peak ~1.0) grid is kept out of halftone-scan by the ceiling.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
test(classify): add synthetic line-art rejection, raise floor to 0.70 (MK-7)
All checks were successful
Check / fmt + clippy + build + tests (pull_request) Successful in 19s
Details
571d2e6b47 
Cover the line-art / signature acceptance criterion synthetically, without needing real image fixtures. Adds a `line_hatching` generator (signature-style strokes) whose autocorrelation peak lands ~0.69, the mid-range that mis-classified as halftone during MK-6 validation (observed 0.641 for a real signature). `line_art_is_not_halftone` asserts every pitch variant stays below the floor and out of halftone-scan.

Raises `halftone_min` from 0.55 to 0.70 so the band rejects this line-art range (and the observed 0.641 signature) while still admitting the synthetic halftone (peak 0.80) as halftone-scan with confidence >= 0.5. The 0.70 floor matches the bottom of the true-screened-print range in the MK-7 data table.

Real-image tuning (a checked-in newspaper/magazine fixture) remains a follow-up; the synthetic tests now exercise the full floor+ceiling band.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
fix(classify): widen ceiling scale, de-fragilize halftone tests (MK-7 review)
All checks were successful
Check / fmt + clippy + build + tests (pull_request) Successful in 17s
Details
Create release / Create release from merged PR (pull_request) Has been skipped
Details
9d4f489a8d 
Address review of the halftone band:

The ceiling now uses its own `soft_scale_halftone_max` (0.10) instead of reusing the tight floor scale (0.04). With one shared tight scale, both sigmoids eroded the band edges and a legitimately strong halftone just below `halftone_max` got its score halved. The wider ceiling scale keeps true high-periodicity print scoring fully while still curving down past ~1.0 digital-perfect patterns.

Test cleanup: `classifies_halftone` no longer asserts a thin `conf >= 0.5` margin on the synthetic (that bound is scoped to real fixtures in the acceptance criteria, and the synthetic margin was fragile); `line_art_is_not_halftone` drops the over-constrained `peak < halftone_min` assertion and keeps the real intent (`class != HalftoneScan`), and hoists `cfg()` out of the loop. Doc comment corrected to the generator's actual ~0.69 peak.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
nrupard deleted branch feat/classify-halftone-ceiling-MK-7 2026-05-21 18:01:31 +02:00
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
pandoras-box/monkey!28
No description provided.