Skip to content

Can a checker know when not to decide?

Golden Gate tests whether a system accepts, refuses, asks for evidence, or abstains for the reason specified by the case.

Problem being tested

Passing cases authored by the implementation team can expose defects, but cannot establish foreign semantics or independent validity. Golden Gate uses one frozen artifact twice: first as a captive qualification system, later as the byte-identical packet controlled and replayed by outsiders.

Three kinds of answer

  • MACHINE: an exact exit or certificate is objectively checkable.
  • PROPERTY: a declared state or trace invariant decides the case.
  • ADJUDICATION: there is no manufactured machine truth; independent judgments and disagreement remain evidence.

The scorecard orders unsafe acceptance, unsafe refusal, unauthorized governance selection, unsupported acceptance under disagreement, typed abstention, evidence requests, and safe coverage. It never collapses semantic, authority, disclosure, latency, or monetary burden into a composite confidence score.

Replay a frozen packet

$ PYTHONPATH=bulla/src python bulla/bench/golden/v0.3/run_boundary_suite.py
$ python -I bulla/scripts/verify_golden_v03.py bulla/bench/golden/v0.3

The zero-import verifier recomputes the oracle rules, formula hashes, safe regions, declared case sets, decision rules, and manifest bindings without importing Bulla.

What the internal suite establishes

Project-run cases cover adversarial soundness, changed-input semantics, portability configuration, tamper resistance, economic state transitions, disclosure, rule conflict, ratchets, completeness-assumption stress, resource frontiers, precedent laundering, and effect laundering. They establish only the reported behavior inside the frozen models and environments actually observed.

What the supplied evidence supports

Availability: the finite checker core ships experimentally in Bulla 0.44.1; qualification packets and clean-room runners remain source or research artifacts. Evidence: methods implemented; internal replay completed; external authors, adjudicators, independent implementations, and witnesses remain zero until real submissions close the corresponding gates. Found schemas are not production traffic, configured CI is not observed portability, and a simulated outsider is still captive.

Read The Golden Gate, bring one seam, or open a bounded review packet.