$ paw-test compare --help
exit code: 0
--- stdout ---

 Usage: paw-test compare [OPTIONS] {adapter_a} {adapter_b} {suite_path}

 Run every suite case through two compiled adapters and diff the results, per case.

 Read-only against both adapters: `--backend real` never compiles. A per-case diff is
 the point -- this is the technique that actually decided whether `paw-ft-bs48` was a
 different compile from `paw-4b-qwen3-0.6b` (132/134 byte-identical outputs, see
 `measurements/README.md`'s "Finetune compiler" section), where the aggregate pass-rate
 delta alone sat inside the LLM judge's own measured run-to-run noise.

 Both adapters are held in memory for the whole run (two ~600MB llama.cpp models
 under `--backend real`, never released) -- the first row's latencies include that
 cold load, not steady-state inference time.

╭─ Arguments ──────────────────────────────────────────────────────────────────────────────────────────────────────────╮
│ *    adapter_a       <path>  First compiled .paw adapter [required]                                                  │
│ *    adapter_b       <path>  Second compiled .paw adapter [required]                                                 │
│ *    suite_path      <path>  Path to declarative suite.yaml specification [required]                                 │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
╭─ Options ────────────────────────────────────────────────────────────────────────────────────────────────────────────╮
│ --backend  -b               <str>   Backend engine: mock | real [default: mock]                                      │
│ --fuzz         --no-fuzz            Also run the suite's fuzz cases through both adapters, not just standard_cases   │
│                                     -- on by default, so a default `compare` run covers the same cases `paw-test     │
│                                     check` does. Pass --no-fuzz for standard_cases only.                             │
│                                     [default: fuzz]                                                                  │
│ --json                      <file>  Write the full CompareReport as JSON to this path                                │
│ --help                              Show this message and exit.                                                      │
╰──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
