R/Python Scientific Parity for AOI Uncertainty

eyeprocess and eyeprocesspy share the same scientific contract for AOI perturbation and uncertainty analysis. Parity means the same analytical concepts, defaults, branch semantics, failure handling, and output meanings—not necessarily byte-for-byte numerical identity across different language backends.

Contract shared across both packages

Concept	R	Python	Contract
Baseline branch	baseline	baseline	Unchanged nominal geometry is explicit and retained
Outside label	__outside__	__outside__	No AOI contains the observation
Ambiguous label	__ambiguous__	__ambiguous__	More than one AOI contains the observation
Missing coordinates	NA	None/NaN	Preserved as missing, never converted to outside/zero
Units	px, deg	px, deg	Degree perturbations require explicit viewing geometry or degrees-per-pixel
Jitter	seeded	seeded	Deterministic within implementation and seed recorded
Geometry failure	audit row	audit row	Branch retained as failed, never silently dropped
Model failure	failure table	failure table	Callback error retained
Non-convergence	explicit flag	explicit flag	Not treated as a valid converged estimate
Stability frequency	descriptive	descriptive	Never interpreted as probability the conclusion is true

Frozen cross-language fixture

Both packages ship the same fixture for a two-AOI example. It checks exact baseline labels, dilation labels, outside membership, overlap ambiguity, missing-coordinate preservation, unchanged-assignment proportion, and newly-assigned proportion.

observation_id,x,y,baseline,dilate_1_px
1,1,1,a,a
2,11,5,__outside__,__ambiguous__
3,13,5,b,b
4,10.5,5,__outside__,a
5,,,,

Expected backend differences

Some implementation details are intentionally language-native:

• R uses base-R graphics; Python returns Matplotlib axes.
• Seeded jitter is reproducible within each implementation, but R and NumPy RNG streams are not required to generate identical coordinates from the same integer seed.
• Statistical callbacks are user supplied, so exact coefficient values depend on the estimator/backend chosen by the analyst.
• Floating-point results can differ at machine precision even when geometry semantics agree.

These differences must not alter labels, units, failure semantics, or interpretation rules.

Reporting parity

Methods sections should describe the analytical contract—not package-specific implementation quirks. Report the nominal AOIs, perturbation grid, coordinate/viewing geometry, overlap policy, assignment level, recomputed features, model specification, convergence/failure handling, and sensitivity summaries in the same way regardless of language.

Return to the AOI uncertainty handbook.