Benchmark v1 · Protocol published

SemaMig-Bench

Does an AI migration preserve the behavior of the system it replaces?

SemaMig-Bench evaluates agentic COBOL-to-Java migration by running the executable source and generated Java candidate on the same held-out workloads. It scores contract-relevant behavior, not source similarity or compilation alone.

Release status

The method is public. The first benchmark run comes after freeze.

Protocol

Published

Task design, source-oracle evaluation, scoring, and release criteria.

Task corpus

In construction

50 practitioner-reviewed synthetic COBOL-to-Java tasks planned for v1.

Baseline results

Forthcoming

Scores publish only after the corpus and verifier admission gates are frozen.

Source-oracle evaluation

One hidden workload. Two isolated executions. One behavioral verdict.

01

Task

COBOL repository, contract, and public cases

02

Agent

Produce a native Java candidate

03

Paired run

Execute source and target on the same hidden input

04

Verdict

Pass the task or expose the first divergence

The candidate runs in a target-only image without the COBOL runtime, source files, or network access. Verifiers must accept alternate correct implementations and reject wrappers, public-case hard coding, and COBOL-specific semantic mutants.

Planned v1 scope

Repository migration across three horizons.

These are release targets, not completed corpus statistics. Every task must pass determinism, acceptance-breadth, verifier-strength, and independent-review gates before freeze.

TierTargetEvaluation unit
Diagnostic15Isolated semantic hazards
Workflow25Multi-file business workflows
System10Multi-program and batch boundaries

Development split

10 public tasks

Evaluation split

40 hosted tasks

Benchmark v1 release

Dataset and baseline results are in preparation.

The task corpus, verifier audit, full agent trajectories, and first leaderboard will be published together after the v1 suite is frozen. No placeholder scores will be presented as results.

01Public development tasks and a hosted evaluation split

02Reference and alternate-correct Java implementations

03COBOL-specific semantic mutants and verifier audit results

04Shared-harness baseline trajectories, cost, and failure analysis

05Versioned task cards, runner, comparator, and signed result bundles

Research and participation

Read the protocol, then bring a migration worth measuring.

The earlier STRUST-v1 study validates the differential verifier on controlled pairs. SemaMig-Bench turns that method into a shared evaluation for migration agents.