e0d6735ec0
Phase 0 (Foundation) was already established in the repository. All required components are in place: - Cargo workspace with three crates (miroir-core, miroir-proxy, miroir-ctl) - rust-toolchain.toml pinning Rust 1.87 - All key dependencies wired (axum, tokio, reqwest, serde, config, clap, uuid) - Config struct with full YAML schema from plan §4 - Style configuration (rustfmt.toml, clippy.toml, .editorconfig) - Project files (CHANGELOG.md, LICENSE, .gitignore) Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
1.1 KiB
1.1 KiB
P12.OP4: Score Normalization at Scale — Completed
Summary
Created benchmark infrastructure and simulation model to validate cross-shard score comparability (Plan §15 OP#4). Confirmed that scores are NOT comparable across skewed shards without normalization.
Key Findings
- Uniform distribution: τ ≈ 1.0 (no problem)
- 10× skew: mean τ ≈ 0.91 (34% pass τ ≥ 0.95)
- 100× skew: mean τ ≈ 0.72 (2% pass τ ≥ 0.95)
- 1000×+ skew: mean τ ≤ 0.5 (severe degradation)
Recommendation
Implement Reciprocal Rank Fusion (RRF) for rank-based merging:
- Immune to score scale differences
- Proven in production (OpenSearch)
- No extra round-trip (unlike global IDF preflight)
Deliverables
crates/miroir-core/src/score_comparability.rs: Simulation model with Kendall τ calculationcrates/miroir-core/benches/score_comparability.rs: Benchmark executabledocs/research/score-normalization-at-scale.md: Research summary with mitigation options
Follow-up Work
Create bead for RRF merger implementation:
- Add over-fetch factor to scatter-gather
- Implement RRF scoring in merger.rs
- Add shard population metrics/monitoring