Repository navigation
Limitations
APORIA reports evidence under declared assumptions and a finite experiment. The current implementation and measurements have limits at the language, search, representation, benchmark, and inference levels.
A region that receives no evaluations cannot be discovered from runtime evidence. The benchmark notes give synthetic/tenth_pct_3d as an example: at 640 evaluations, a region occupying roughly 0.1% of the three-dimensional domain is unlikely to be sampled reliably. The current ladder detected nothing there. Random and adaptive search share this basic density limit.
Atlas cells are axis-aligned and split only to a maximum depth of 12. On a unit interval this limits cell width to roughly 2.4e-4; the 2e-6-wide analytic/reciprocal_bound and synthetic/narrow_1d_unreachable regions lie below that resolution. A curved boundary such as ode/euler_decay_2d is approximated by boxes and can be detected without ever being localised by the current metric. More budget alone does not remove the representation's depth and shape limits.
Calibration is fitted to the run's own evidence and describes excess over that run's typical measurements. A campaign in which all points are similarly anomalous may therefore have little contrast. Fusion's empirical channel correlations are estimated from that same finite sample. Risk strength is a ranking measure, not a probability.
The minimisation risk oracle deliberately excludes behavioural relations because they are evaluated across records rather than at one point. Interval narrowing assumes a monotone failure boundary between the pinned value and the tested edge; for irregular regions, the returned interval is only what its verification samples support. The corpus ground-truth gate uses declared rules and divergence, so it cannot score a fault that exists only as a measured-channel violation without a suitable ground-truth definition. A verified reduced case preserves its selected oracle; it does not prove a globally minimal explanation.
The shipped front end accepts Aporia DSL files. The Rust Executor trait allows the campaign library to receive alternate execution results, but there is no packaged external-program adapter or foreign-program CLI workflow. Numerical and differential probes are only meaningful when an executor provides genuinely distinct precision or reference paths; unsupported paths are skipped.
The corpus is small, project-authored, and has known coverage gaps. It does not establish broad performance on external scientific applications, GPUs, or production systems. Adaptive search wins on only one entry and loses on two in the current comparison.
TRUSTED means no current evidence of a problem under the sampled points, declarations, enabled channels, calibration, and labeling policy. APORIA is not a formal verifier of a scientific model's correctness. It cannot validate assumptions that were never encoded, guarantee safety, or show that an unsampled region is sound.
See Experiments & Results for measured examples and Roadmap for work explicitly described as future in the repository.