Repository navigation
Home
APORIA is a local instrument for studying where a numerical computation becomes difficult to trust. You describe a model and its parameter domain, state any rules you know, and APORIA samples the domain, gathers evidence, and builds a three-label Trust Atlas with findings tied to the executions that produced them.
The current implementation includes an Aporia DSL compiler, a scalar interpreter, a batch evaluator, an adaptive campaign driver, evidence calibration and fusion, counterexample minimisation, and run archives with replay. The benchmark corpus and its measurements are part of the repository; they show useful cases, failures, and open questions rather than a general performance claim. See Experiments & Results and Limitations.
TRUSTED means no current evidence of a problem under the tested assumptions and evidence model. It does not mean proven correct.
- Architecture — crates and execution flow
- How APORIA Works — from model to findings
- Aporia DSL — the model language
- Evidence Model — five channels, calibration, and fusion
- Trust Atlas — partitioning and labels
- Adaptive Search — where the evaluation budget goes
- Counterexample Minimisation — reducing a finding
- Archives and Replay — durable run records
- Benchmarking — corpus and measurement method
- Experiments & Results — current measured results
- Limitations · Roadmap · FAQ
To run the shipped front end, build from software/ and use cargo run --release -p aporia-cli -- run path/to/model.ap. The command accepts a DSL model and reports an analysis; it does not create an archive or minimise findings. See the source README for build requirements and benchmark commands.