Generated by python tests/tools/trial.py report --write from tests/cobol_mainframe/trials.json. Do not edit by hand. check (in CI) fails when this page or the chart is stale.

first-try proof rate: no trial yet (0 fresh estates trialled)

Per program: did the porting loop's FIRST attempt produce Java the equivalence harness proves identical to the COBOL? A trial freezes the system (engine commit, porting rules, model, harness), takes an estate the system has never seen, and iterates each eligible program to a proof. Each failed attempt carries its cause: the model's (a retry, same system) or ours (a fix commit with a Trial-Cause: trailer: ticket, generator, engine, harness). A trialled estate is development data for good: its numbers stay here but never count as first-try again. The rate counts programs whose first attempt has a verdict (and every program of a closed trial).

• Public code may be in a model's training data; only private estates escape that.
• Eligibility is declared up front, at trial start, so 'out of scope' cannot hide a failure.

No trial yet.