Skip to content

Accuracy and scope

arche is alpha software. A successful example is not evidence that a matching rule is safe for a new population, domain, or decision.

What to measure

Evaluate against labelled pairs that represent the data and risks of your use case. Report at least:

  • precision and false merges
  • recall and missed matches
  • review rate and the outcomes of reviewed cases
  • blocking recall, when candidate generation is used
  • performance by meaningful slices such as language, region, source system, or product category

False merges deserve separate attention. They are often more costly than missed matches and can compound when pairwise links are clustered.

Current limits

  • A crosswalk score is not a calibrated probability.
  • Review thresholds and review capacity must be set for the deployment, not inferred from a documentation example.
  • The current product lane is experimental. It should not be presented as a general product-catalogue solution until it has an independently reproducible benchmark.
  • No released arche MCP server exists. Agent integrations remain the responsibility of the calling system.
  • Statute mappings are software artifacts, not legal advice or a determination of the law governing a deployment.

What would make the thesis credible

The central hypothesis is that inspectable representation data, such as equivalence packs, frequency tables, identifier grammars, and vocabularies, improves record resolution in the cases where generic comparisons fail.

That hypothesis needs predeclared, independently reproducible evaluations on the populations it claims to serve. It should be compared with plain string matching, established probabilistic linkage, and model-assisted matching under the same candidate set and decision policy.

See review and evaluate a crosswalk for the minimum review artifact and evaluation conditions.

See places and products in action for a runnable place-resolution tour and an evaluation against the public Abt-Buy electronics benchmark. The notebook release evidence records the exercised notebook set, data paths, and current scope.