Files

Application composition corpus

Four authored tasks exercise the actual document loader, tree/bytecode evaluators and interactive history API. Each has separate Python and Node implementations of the producer and consumer. CASES.json is the eight-case regression subset; the reviewer driver generates 80 size/seed/direction instances before execution. They are four application families, not 80 independent real-user applications.

  • Invoice ledger: discard void rows, first-posted invoice deduplication, integer cents/discount accounting, signed refunds with floor division, regional totals.
  • Routes: minimum nonnegative parallel edge, zero-cost edges, disconnected nodes and weighted single-source distances. Python uses heap-based Dijkstra, Node a linear frontier; independent reference uses SQL and all-pairs Floyd-Warshall.
  • Concordance: NFC plus ASCII lowercase, explicit space/tab/LF/CR/NBSP separators, occurrence/document frequencies, UTF-8 byte order. This is not Unicode full case folding. Wide IDs remain strings; special object-key text, combining marks, BMP and astral characters are included.
  • Wine: UCI red-wine CSV, exact decimal density/pH scaling to integers and quality-group count/sum/min/max. Full input contains 1599 rows; no predictor accuracy or general machine-learning performance is claimed.

Independent SQL/Fraction/Floyd references live in the reviewer experiment application_cases.py; no foreign program output supplies expected answers. The raw/public-data digest, citation, license and retrieval source are preserved in data/PROVENANCE.json. Wine Quality is by Cortez, Cerdeira, Almeida, Matos and Reis (2009), UCI, DOI 10.24432/C56S3T, CC BY 4.0. CSV text fixture generation normalizes physical line endings and uses the requested first N data rows.

The composition copies input into a producer, gets a normalized value, copies it into the consumer, records an expected ordinary exception after saving answer, recovers it, replaces the consumer's bridge with null, verifies the producer's copy and answer remain, and closes both sessions. Interactive recordings repeat these actions and strictly replay both raw and losslessly packed history. Native and independent checks distinguish explicit null from display stdout.