Deterministic CSR Regression Gate¶
Report-only CI gate — not a leaderboard metric
The Constraint Satisfaction Rate (CSR) harness is a deterministic, per-PR regression gate (bit-identical every run — no LLM judge, no embeddings, no Redis module; identical on Redis and Valkey). Its RSR (standard/adversarial) and Adversarial Gap numbers are report-only signals for catching regressions, not a "higher is better" leaderboard score.
Run date: 2026-07-03T09:37:37.974131+00:00
Python: 3.12.13
Platform: macOS-26.3.1-arm64-arm-64bit
Summary¶
| Metric | Value |
|---|---|
| Cases | 8 |
| Errors | 0 |
| RSR (standard) | 0.8750 |
| RSR (adversarial) | 0.2750 |
| Adversarial Gap | 0.6000 |
Gap alert: Adversarial Gap >= 0.3 — likely keyword-dependent retrieval. Expected of the pure-BM25 lexical path (healthy signature); this is NOT the #409 query-blindness signature.
Query-blindness flags (the #409 signature — identical rankings + standard CSR < 0.5): composite_control_409
Per primitive¶
| Primitive | n | CSR (std) | CSR (adv) |
|---|---|---|---|
| coverage | 1 | 1.0000 | 0.3333 |
| exclusion | 1 | 1.0000 | 0.3333 |
| lexical_recall | 1 | 1.0000 | 0.2500 |
| ordering | 1 | 1.0000 | 0.3333 |
| precision | 1 | 1.0000 | 0.5000 |
| query_independence | 2 | 0.5000 | 0.1000 |
| recency | 1 | 1.0000 | 0.2500 |
Per case¶
| Case | Primitive | Status | CSR (std) | CSR (adv) | Rankings identical | Query-blind |
|---|---|---|---|---|---|---|
| lexical_recall_basic | lexical_recall | ok | 1.0000 | 0.2500 | False | False |
| ordering_score_separated | ordering | ok | 1.0000 | 0.3333 | False | False |
| recency_none_older | recency | ok | 1.0000 | 0.2500 | False | False |
| coverage_all_topic_facts | coverage | ok | 1.0000 | 0.3333 | False | False |
| exclusion_unrelated | exclusion | ok | 1.0000 | 0.3333 | False | False |
| precision_multi_topic | precision | ok | 1.0000 | 0.5000 | False | False |
| query_blind_409 | query_independence | ok | 1.0000 | 0.2000 | False | False |
| composite_control_409 | query_independence | ok | 0.0000 | 0.0000 | True | True |
Notes¶
- Deterministic: lexical/BM25 + composite control only — no LLM judge, no embedding model, no Redis module. Identical on Redis and Valkey.
- A LARGE Adversarial Gap on the lexical path = keyword dependence (healthy). Query-blindness (#409) = identical rankings for both queries AND a low standard CSR.
executed_pathis recorded per run in the JSON; a composite fallback can never masquerade as a lexical number.