← ResearchGitHub
RESEARCH / BENCHMARKS

Measure the engineering effect, don't assume it.

BOUND's benchmark program should measure whether explicit boundaries, contracts, independent execution, and continuous verification change observable engineering outcomes.

Evidence status: this page defines proposed measurements and experimental structure. It does not report validation results, adoption, or causal effects that have not been established by published experiments.

Candidate measurements

MeasureOperational definitionWhy it matters
Integration defectsDefects discovered when independently developed components are integrated.Captures failures at coordination boundaries.
Contract violationsObserved deviations from a declared interface or behavioral contract.Separates contract drift from implementation defects.
Coordination eventsHuman synchronization events required to resolve cross-boundary ambiguity.Measures coordination load rather than implementation volume.
Time to integrationElapsed time from independently executable change to verified integration.Provides a concrete integration-cycle measure.
Rollback frequencyChanges reverted because integration or verification failed.Captures downstream cost of divergence.
Agent scope violationsAgent actions outside the explicitly permitted execution context.Measures bounded-agent compliance.
Verification failuresRequired checks that fail before acceptance.Shows where contracts or implementations diverge.

Experimental structure

  1. Define the engineering task and population.
  2. Specify the boundary and contract artifacts before execution.
  3. Record baseline and BOUND conditions using the same task definition.
  4. Capture raw events and verification results rather than only summary metrics.
  5. Predefine exclusions, failure handling, and stopping conditions.
  6. Report limitations and alternative explanations alongside results.

Minimum benchmark record

{
  "task": "...",
  "condition": "baseline | bound",
  "participants": [],
  "boundaries": [],
  "contracts": [],
  "coordination_events": [],
  "integration_events": [],
  "verification_results": [],
  "rollbacks": [],
  "scope_violations": [],
  "duration": "...",
  "limitations": []
}

The structure is illustrative rather than a normative benchmark schema. A future benchmark implementation should version its data model before collecting comparable datasets.

What a benchmark must not claim by itself

  • A reduction in one measured defect class does not establish universal superiority.
  • One project or team does not establish general adoption or industry effect.
  • Correlation between explicit contracts and an outcome does not by itself establish causation.
  • Agent scope-violation measurements depend on the enforcement model used by the experiment.