docs: add deterministic multi-family admissibility benchmark documentation

docs/benchmarks/multi_family_admissibility_benchmark.md

@@ -0,0 +1,98 @@
+# Deterministic Multi-Family Admissibility Benchmark
+
+## Purpose
+
+The deterministic multi-family admissibility benchmark tracks operational admissibility degradation across fixture families registered in the manifest.
+
+Each manifest-registered fixture family contributes one deterministic degradation curve using the same standard levels, so contributors can compare progression behavior across families without changing scoring rules or artifact shape.
+
+## Pipeline
+
+```mermaid
+flowchart LR
+    A[fixtures/manifest.json]
+    B[DegradationCurveGenerator.fixtures_for_manifest_family(...)]
+    C[AdmissibilityScorer]
+    D[artifacts/multi_family_admissibility_results.json]
+    E[Reproducibility and progression tests]
+
+    A --> B --> C --> D --> E
+```
+
+### Pipeline notes
+
+1. `fixtures/manifest.json` is the source of truth for which fixture families participate.
+2. `DegradationCurveGenerator.fixtures_for_manifest_family(...)` resolves fixtures for each family from manifest registration.
+3. `AdmissibilityScorer` computes exact admissibility component outcomes for each level.
+4. Results are written to `artifacts/multi_family_admissibility_results.json` in a stable deterministic JSON layout.
+5. Reproducibility and progression tests validate that the committed artifact remains consistent and semantically protected.
+
+## Current fixture families
+
+The current multi-family benchmark includes these manifest-registered families:
+
+- `coding_workflow_pr_review`
+- `incident_response_page_triage`
+
+## Standard degradation levels
+
+Every included family is evaluated at exactly four standard levels in explicit order:
+
+1. `baseline`
+2. `mild`
+3. `moderate`
+4. `severe`
+
+## Determinism guarantees
+
+The benchmark is designed to remain deterministic across local runs and CI runs:
+
+- manifest-driven family selection
+- explicit level order (`baseline`, `mild`, `moderate`, `severe`)
+- exact rational score aggregation
+- stable JSON output structure and ordering
+- no timestamps or environment-dependent fields
+
+## Regeneration commands
+
+Use either command to regenerate the deterministic multi-family artifact:
+
+```bash
+python scripts/generate_multi_family_admissibility_artifact.py
+```
+
+```bash
+npm run generate:multi-family-admissibility
+```
+
+## Validation commands
+
+Run the targeted protections plus the repository-wide check entrypoint:
+
+```bash
+pytest tests/test_multi_family_admissibility_artifact.py -q
+pytest tests/test_artifact_reproducibility.py -q
+pytest tests/test_manifest_fixture_families.py -q
+npm run check
+```
+
+## Regression protections
+
+The benchmark is protected by deterministic regression checks that enforce:
+
+- committed artifact must match regenerated output
+- every family must expose all four standard levels
+- baseline and severe behavior is explicitly checked
+- mild and moderate behavior must be distinct
+- degradation must be progressive:
+  - `baseline > mild >= moderate > severe`
+
+## Non-goals
+
+This benchmark intentionally excludes:
+
+- LLM judging
+- embeddings
+- fuzzy semantic similarity
+- runtime orchestration
+- deployment/showcase dependencies