As a digitizer maintainer, I want to know whether a DP-level curve-identity prior would catch S/M-crossing cases that #1170's post-DP candidate-walk does not, so that we have a documented next-step if a future chart fails both detectors.
What
Extend `_ridge_dp_two_paths` so each path's cost includes a slope-continuity term against its own history (track A = path whose slope stays continuous from its starting direction; track B = the other). Output the produced tracks side-by-side with the post-DP candidate-walk detector's output on:
- `ttartisan-af-75mm-f2-0` stopped freq30 (the in-the-wild bug case)
- `ttartisan-50mm-f1-2` stopped freq30 (also affected — see ADR-055)
- `ttartisan-tilt-50mm-f1-4` stopped freq30 (control: must not regress)
- 4 Tier 1 anchors
Why
ADR-055 (post-DP candidate-walk + both-reverse + swap-left) fixes the af-75 and 50/1.2 cases without regressing the tilt-50 control. The S151 probe data showed a DP-level curve-identity prior was not needed for this cohort. Path A stays filed in case a future chart surfaces where the post-DP fix-up cannot get a reliable verdict — e.g. when both DP tracks have severely curtailed coverage in the convergence window.
Risk noted in S151 probe: slope-stability across the field is NOT a reliable discriminator (af-75 t1=0.82 > t2=0.78 but tilt-50 t1=0.70 < t2=0.84 — direction-inverted). The spike must design a different prior shape.
Acceptance criteria
Out of scope
Per-column raw-mask discriminator is tracked separately as Path B.
Related
As a digitizer maintainer, I want to know whether a DP-level curve-identity prior would catch S/M-crossing cases that #1170's post-DP candidate-walk does not, so that we have a documented next-step if a future chart fails both detectors.
What
Extend `_ridge_dp_two_paths` so each path's cost includes a slope-continuity term against its own history (track A = path whose slope stays continuous from its starting direction; track B = the other). Output the produced tracks side-by-side with the post-DP candidate-walk detector's output on:
Why
ADR-055 (post-DP candidate-walk + both-reverse + swap-left) fixes the af-75 and 50/1.2 cases without regressing the tilt-50 control. The S151 probe data showed a DP-level curve-identity prior was not needed for this cohort. Path A stays filed in case a future chart surfaces where the post-DP fix-up cannot get a reliable verdict — e.g. when both DP tracks have severely curtailed coverage in the convergence window.
Risk noted in S151 probe: slope-stability across the field is NOT a reliable discriminator (af-75 t1=0.82 > t2=0.78 but tilt-50 t1=0.70 < t2=0.84 — direction-inverted). The spike must design a different prior shape.
Acceptance criteria
Out of scope
Per-column raw-mask discriminator is tracked separately as Path B.
Related