Quantization Saliency Agreement Study — Frozen Analysis Specification

Frozen analysis specification and pre-registered decision rules for the study "Do post-training quantization (PTQ) saliency criteria identify the same quantization-sensitive channels?" This deposit fixes the content of the pre-registered specification for replication. Contents: design-spec.md — study design and pre-registered research questions and decision rules (RQ1–RQ5), corrected metric definitions, ground-truth definitions, and analysis plan (dated 2026-06-03). config_phase1_frozen.py — frozen run configuration for Phase 1 (Qwen2.5-0.5B): seeds, calibration/eval sizes, bits, damping grid, sweep layers, and RQ1–RQ5 numeric thresholds. config_phase2_frozen.py — frozen run configuration for Phase 2 (EleutherAI/pythia-410m) and the pre-registered third-model contingency (SmolLM2-360M). AMENDMENT_fp64_recheck_threshold.md — one documented pre-registration amendment (fp64 recheck threshold 1e-4 → 1e-6). This deposit fixes the content of the specification at deposit time so the released analysis can be checked against the rules as written; the chronological ordering of "rules frozen before data" is evidenced by the project's version-control history. The design spec names "TinyLlama-1.1B" only as an illustrative second model; the binding rule was "≥2 models of differing family," and the actual Phase-2 model (Pythia-410m) was selected and frozen before the Phase-2 ground-truth run.v2 adds config_phase3_frozen.py — the Phase-3 (SmolLM2-360M) pre-registration, deposited while the Phase-3 ground-truth run was in progress.