Context
The README's central finding is that architectures differ in dose-response shape: an interior optimum (effect peaks then reverses past the cliff), monotonic-to-cliff (best usable effect is the last coherent dose), or flat (a broken/low-SNR vector). Today a reader must eyeball the curve to tell these apart. A typed classifier on the analysis makes the shape a first-class, testable output.
Proposal
Add a small enum-like classifier that labels a dose curve as interior_optimum, monotonic_to_cliff, or flat, computed purely from the aggregated SweepPoint list plus the existing DoseAnalysis (sweet spot, cliff, coherent points). Attach the label to DoseAnalysis as a typed field and surface it in the summary lines.
Acceptance criteria
- Add
DoseShape = Literal["interior_optimum", "monotonic_to_cliff", "flat"] and populate a shape: DoseShape field on DoseAnalysis inside analyze_dose (pure, CPU-only).
- Classification rule documented in the docstring, e.g.:
flat when no coherent steered point clears baseline by a stated margin; interior_optimum when the max coherent effect shift occurs strictly before the last coherent dose; monotonic_to_cliff when the best coherent effect is the last coherent point.
- Add the shape label to
_effect_summary_lines so the card states which regime the vector is in.
- CPU-only unit tests in
tests/test_report.py covering all three shapes on hand-built two-sided curves (interior peak, monotone rise to cliff, flat). No model download.
ruff check ., ruff format --check ., mypy src, pytest -q all pass.
Context
The README's central finding is that architectures differ in dose-response shape: an interior optimum (effect peaks then reverses past the cliff), monotonic-to-cliff (best usable effect is the last coherent dose), or flat (a broken/low-SNR vector). Today a reader must eyeball the curve to tell these apart. A typed classifier on the analysis makes the shape a first-class, testable output.
Proposal
Add a small enum-like classifier that labels a dose curve as
interior_optimum,monotonic_to_cliff, orflat, computed purely from the aggregatedSweepPointlist plus the existingDoseAnalysis(sweet spot, cliff, coherent points). Attach the label toDoseAnalysisas a typed field and surface it in the summary lines.Acceptance criteria
DoseShape = Literal["interior_optimum", "monotonic_to_cliff", "flat"]and populate ashape: DoseShapefield onDoseAnalysisinsideanalyze_dose(pure, CPU-only).flatwhen no coherent steered point clears baseline by a stated margin;interior_optimumwhen the max coherent effect shift occurs strictly before the last coherent dose;monotonic_to_cliffwhen the best coherent effect is the last coherent point._effect_summary_linesso the card states which regime the vector is in.tests/test_report.pycovering all three shapes on hand-built two-sided curves (interior peak, monotone rise to cliff, flat). No model download.ruff check .,ruff format --check .,mypy src,pytest -qall pass.