Target Model Verification

Same layer-peeling approach as env-verify, but with the real target model that may require multi-GPU tensor parallelism. Runs the model twice (without and with multi-chip stack) and diffs results to isolate failures.

Skill Components

model-verify/
├── SKILL.md                            # This file — execution flow
├── scripts/
│   └── diff_analysis.py                # Compare Run A vs Run B, classify errors (JSON)
└── references/
    └── multichip-errors.md             # Multi-chip error patterns and diff truth table

Reused from env-verify:

• env-verify/scripts/run_offline_inference.py — Phase A test (parameterized)