@1mbrain/benchmarks 0.1.1

package/fixtures/graph-stress-hard/README.md

@@ -0,0 +1,43 @@
+# Graph Stress Hard Benchmark
+
+This fixture is a diagnostic benchmark for graph-aware memory retrieval. It is
+intentionally harder and less provider-neutral than the Claude balanced and
+realistic fixtures.
+
+## Purpose
+
+Use this dataset to tune and compare:
+
+- Multi-hop association recall.
+- Conflict and current-state resolution.
+- Graph traversal with weak lexical overlap.
+- Near-entity distractor resistance.
+- Abstention when only a similar entity has the requested fact.
+
+## Shape
+
+- 12 conversations.
+- 144 memory records.
+- 60 questions.
+- Deterministic generation from `generate_graph_stress_hard.js`.
+
+## Category Mix
+
+- 24 `multi_hop_association`.
+- 12 `contradiction_resolution`.
+- 12 `graph_traversal`.
+- 7 `entity_disambiguation`.
+- 5 `abstention`.
+
+## Usage
+
+```powershell
+node packages/benchmarks/fixtures/graph-stress-hard/generate_graph_stress_hard.js
+$env:BENCH_DATASET = "graph-stress-hard"
+$env:BENCH_PROVIDERS = "1mbrain_graph_full,1mbrain_vector_only,vector_baseline"
+node packages/benchmarks/dist/runner.js
+```
+
+This is best interpreted alongside provider-neutral datasets. A graph-enabled
+provider should separate itself most clearly on `multi_hop_recall` and
+`memory_update` metrics.