# ForgeBench v1 — Verification Report - Date: 2026-10-27T08:16:51 - Scope: dataset (100 tasks) + evaluation harness (offline tests) - Result: **19/19 checks passed** ## Checks | # | Check | Status | Detail | |---|-------|--------|--------| | 1 | V1 strict schema (field set/order/types) | PASS | 100 rows | | 2 | V2 unique task_ids | PASS | | | 3 | V2 id pattern ^[A-Z]+_\d{3}$ | PASS | | | 4 | V2 20 tasks per module | PASS | {'api': 20, 'mixin': 20, 'migration': 20, 'debug': 20, 'build': 20} | | 5 | V2 module order api=80 chars) were searched in sampled corpus shards of ComerX/forge-neoforge-sft-v2.1.0-rc2 and -rc1 (3 shards each): 0 hits. - **V9** harness offline tests: sandbox workspace preparation (--dry-run), ast_eval passing its own golden, ast_eval catching poisoned answers (Forge namespace + 1.16 text API), GameTest template generation, and the run_benchmark driver. - **V10** Dockerfile/template coherence, judge rubric presence, and the no-raw-javac/no-Mockito rule (executable use). ## Known limitations (honest scope) 1. The Gradle compile path (sandbox_runner) was verified structurally (workspace preparation, command assembly, output parsing) but NOT executed end-to-end in this sandbox: no Docker daemon and no 8 GB RAM budget for NeoForm decompile are available here. The same limitation is documented for the main SFT audit (compile_verified unreachable). 2. javalang is a Java-8 parser; Java 14+ syntax (records, pattern matching) is shimmed for validation only. The real gate is the Gradle compile inside the MDK. 3. Contamination screen is sampled (6 shards), not exhaustive; it screens exact-prefix containment, not paraphrase-level overlap. 4. Plugin versions in build-module goldens (NeoGradle 7.0.145, ModDevGradle 1.0.24, Minotaur 2.+, curseforgegradle 1.1.24) reflect mid-2024/2025 releases; syntax is the graded property, not the pinned version.