Hypotheses (frozen)

• H1 · Domain adaptation QLoRA Gemma beats the same model with a fixed zero-shot prompt on held-out macro F1.
• H2 · Prompting versus fine-tuning QLoRA beats a fixed five-class few-shot prompt.
• H3 · Generalization Measurable improvement remains on the challenge set.
• H4 · Data scaling Performance rises with train size, then diminishes.
• H5 · Remaining weaknesses Hard boundaries persist, especially PARTIAL vs SUFFICIENT and PARTIAL vs INSUFFICIENT.

Within the completed Gemma v1.0 experiments, few-shot is strongest and QLoRA doesn't beat it. Cross-family hypotheses requiring a common dataset version aren't treated as controlled tests.

Evaluation contract

• Deterministic inference do_sample=false, temperature 0 where supported.
• Strict parsing Ambiguous output becomes UNPARSEABLE, never silently coerced to gold.
• Uncertainty Bootstrap CIs; McNemar for paired comparisons where appropriate.
• Slices Domain, artifact type, difficulty, failure tags.
• Prompt freeze Canonical template in src/controlsift/prompting/templates.py. Don't retune from test failures after lock.

Sealed v1.1 test hash

Full seal object: PROTOCOL_SEAL.json (mirrored from governance).