{
  "description": "Repeated policy-overhead benchmark runs, identical code and methodology, sequential requests on a shared development machine. The 2026-08-16 run is the published headline figure and is deliberately conservative: later runs measured lower. Runs whose service-overhead p95 exceeds 25ms carry stall outliers with max values many times their own p99 (host contention, not software). Reproduce with policy_overhead_bench.py from this directory.",
  "service_overhead_p95_range_ms": [
    13.55,
    26.2
  ],
  "service_overhead_p50_range_ms": [
    10.0,
    16.0
  ],
  "runs": [
    {
      "label": "2026-08-16 headline run (the published claim: 16ms p50 / 21ms p95)",
      "methodology": {
        "connector": "pgvector (local)",
        "vectors": 1000,
        "embedding_dim": 1536,
        "resources_registered": 100,
        "chunks_per_resource": 10,
        "policy": "1 active RBAC policy, 5 rules (2 group-allow, 1 classification deny, 1 department allow, 1 sensitivity deny)",
        "metadata_resolution": "sidecar",
        "top_k": 10,
        "requests": 150,
        "warmup": 10,
        "pace_rps": 8.0,
        "query_embedding": "client-supplied (server-side embedding excluded)",
        "notes": "Run against a DEBUG=true dev backend (verbose logging and relaxed rate limits) \u2014 overhead numbers are conservative relative to production settings."
      },
      "errors": 0,
      "series": [
        {
          "series": "connector_leg_ms (raw vector query)",
          "n": 150,
          "p50_ms": 126.0,
          "p95_ms": 161.1,
          "p99_ms": 278.52,
          "mean_ms": 133.67,
          "max_ms": 319
        },
        {
          "series": "service_total_ms (retrieval service end-to-end)",
          "n": 150,
          "p50_ms": 141.5,
          "p95_ms": 177.55,
          "p99_ms": 308.21,
          "mean_ms": 150.45,
          "max_ms": 354
        },
        {
          "series": "service_overhead_ms (policy layer, service-side)",
          "n": 150,
          "p50_ms": 16.0,
          "p95_ms": 21.0,
          "p99_ms": 37.04,
          "mean_ms": 16.79,
          "max_ms": 49
        },
        {
          "series": "full_overhead_ms (policy layer incl. HTTP middleware)",
          "n": 150,
          "p50_ms": 46.5,
          "p95_ms": 65.41,
          "p99_ms": 96.91,
          "mean_ms": 48.9,
          "max_ms": 134.48
        }
      ]
    },
    {
      "label": "2026-08-21 run A",
      "methodology": {
        "connector": "pgvector (local)",
        "vectors": 1000,
        "embedding_dim": 1536,
        "resources_registered": 100,
        "chunks_per_resource": 10,
        "policy": "1 active RBAC policy, 5 rules (2 group-allow, 1 classification deny, 1 department allow, 1 sensitivity deny)",
        "metadata_resolution": "sidecar",
        "top_k": 10,
        "requests": 150,
        "warmup": 10,
        "pace_rps": 8.0,
        "query_embedding": "client-supplied (server-side embedding excluded)",
        "notes": "Run against a DEBUG=true dev backend (verbose logging and relaxed rate limits) \u2014 overhead numbers are conservative relative to production settings."
      },
      "errors": 0,
      "series": [
        {
          "series": "connector_leg_ms (raw vector query)",
          "n": 150,
          "p50_ms": 108.0,
          "p95_ms": 156.55,
          "p99_ms": 266.58,
          "mean_ms": 116.09,
          "max_ms": 375
        },
        {
          "series": "service_total_ms (retrieval service end-to-end)",
          "n": 150,
          "p50_ms": 118.0,
          "p95_ms": 184.5,
          "p99_ms": 357.27,
          "mean_ms": 133.51,
          "max_ms": 1007
        },
        {
          "series": "service_overhead_ms (policy layer, service-side)",
          "n": 150,
          "p50_ms": 11.0,
          "p95_ms": 18.65,
          "p99_ms": 30.65,
          "mean_ms": 17.42,
          "max_ms": 895
        },
        {
          "series": "full_overhead_ms (policy layer incl. HTTP middleware)",
          "n": 150,
          "p50_ms": 31.43,
          "p95_ms": 57.33,
          "p99_ms": 132.37,
          "mean_ms": 40.43,
          "max_ms": 918.93
        }
      ]
    },
    {
      "label": "2026-08-21 run B (cleanest recorded: max 26ms)",
      "methodology": {
        "connector": "pgvector (local)",
        "vectors": 1000,
        "embedding_dim": 1536,
        "resources_registered": 100,
        "chunks_per_resource": 10,
        "policy": "1 active RBAC policy, 5 rules (2 group-allow, 1 classification deny, 1 department allow, 1 sensitivity deny)",
        "metadata_resolution": "sidecar",
        "top_k": 10,
        "requests": 150,
        "warmup": 10,
        "pace_rps": 8.0,
        "query_embedding": "client-supplied (server-side embedding excluded)",
        "notes": "Run against a DEBUG=true dev backend (verbose logging and relaxed rate limits) \u2014 overhead numbers are conservative relative to production settings."
      },
      "errors": 0,
      "series": [
        {
          "series": "connector_leg_ms (raw vector query)",
          "n": 150,
          "p50_ms": 104.0,
          "p95_ms": 163.55,
          "p99_ms": 226.4,
          "mean_ms": 111.0,
          "max_ms": 257
        },
        {
          "series": "service_total_ms (retrieval service end-to-end)",
          "n": 150,
          "p50_ms": 114.5,
          "p95_ms": 178.1,
          "p99_ms": 245.28,
          "mean_ms": 121.96,
          "max_ms": 269
        },
        {
          "series": "service_overhead_ms (policy layer, service-side)",
          "n": 150,
          "p50_ms": 10.0,
          "p95_ms": 13.55,
          "p99_ms": 20.04,
          "mean_ms": 10.96,
          "max_ms": 26
        },
        {
          "series": "full_overhead_ms (policy layer incl. HTTP middleware)",
          "n": 150,
          "p50_ms": 31.15,
          "p95_ms": 44.38,
          "p99_ms": 100.37,
          "mean_ms": 33.34,
          "max_ms": 105.97
        }
      ]
    },
    {
      "label": "2026-08-21 run C",
      "methodology": {
        "connector": "pgvector (local)",
        "vectors": 1000,
        "embedding_dim": 1536,
        "resources_registered": 100,
        "chunks_per_resource": 10,
        "policy": "1 active RBAC policy, 5 rules (2 group-allow, 1 classification deny, 1 department allow, 1 sensitivity deny)",
        "metadata_resolution": "sidecar",
        "top_k": 10,
        "requests": 150,
        "warmup": 10,
        "pace_rps": 8.0,
        "query_embedding": "client-supplied (server-side embedding excluded)",
        "notes": "Run against a DEBUG=true dev backend (verbose logging and relaxed rate limits) \u2014 overhead numbers are conservative relative to production settings."
      },
      "errors": 0,
      "series": [
        {
          "series": "connector_leg_ms (raw vector query)",
          "n": 150,
          "p50_ms": 103.5,
          "p95_ms": 174.95,
          "p99_ms": 399.63,
          "mean_ms": 116.78,
          "max_ms": 681
        },
        {
          "series": "service_total_ms (retrieval service end-to-end)",
          "n": 150,
          "p50_ms": 114.5,
          "p95_ms": 218.3,
          "p99_ms": 590.39,
          "mean_ms": 134.64,
          "max_ms": 1091
        },
        {
          "series": "service_overhead_ms (policy layer, service-side)",
          "n": 150,
          "p50_ms": 10.0,
          "p95_ms": 21.2,
          "p99_ms": 272.8,
          "mean_ms": 17.86,
          "max_ms": 444
        },
        {
          "series": "full_overhead_ms (policy layer incl. HTTP middleware)",
          "n": 150,
          "p50_ms": 30.44,
          "p95_ms": 46.55,
          "p99_ms": 624.17,
          "mean_ms": 49.43,
          "max_ms": 1152.97
        }
      ]
    },
    {
      "label": "2026-08-21 run D (host contention: 307ms outlier, 12x its own p99)",
      "methodology": {
        "connector": "pgvector (local)",
        "vectors": 1000,
        "embedding_dim": 1536,
        "resources_registered": 100,
        "chunks_per_resource": 10,
        "policy": "1 active RBAC policy, 5 rules (2 group-allow, 1 classification deny, 1 department allow, 1 sensitivity deny)",
        "metadata_resolution": "sidecar",
        "top_k": 10,
        "requests": 150,
        "warmup": 10,
        "pace_rps": 8.0,
        "query_embedding": "client-supplied (server-side embedding excluded)",
        "notes": "Run against a DEBUG=true dev backend (verbose logging and relaxed rate limits) \u2014 overhead numbers are conservative relative to production settings."
      },
      "errors": 0,
      "series": [
        {
          "series": "connector_leg_ms (raw vector query)",
          "n": 150,
          "p50_ms": 118.0,
          "p95_ms": 199.95,
          "p99_ms": 245.57,
          "mean_ms": 127.59,
          "max_ms": 384
        },
        {
          "series": "service_total_ms (retrieval service end-to-end)",
          "n": 150,
          "p50_ms": 130.0,
          "p95_ms": 232.1,
          "p99_ms": 267.04,
          "mean_ms": 143.41,
          "max_ms": 691
        },
        {
          "series": "service_overhead_ms (policy layer, service-side)",
          "n": 150,
          "p50_ms": 12.0,
          "p95_ms": 26.2,
          "p99_ms": 57.55,
          "mean_ms": 15.82,
          "max_ms": 307
        },
        {
          "series": "full_overhead_ms (policy layer incl. HTTP middleware)",
          "n": 150,
          "p50_ms": 33.43,
          "p95_ms": 88.22,
          "p99_ms": 203.25,
          "mean_ms": 43.35,
          "max_ms": 418.15
        }
      ]
    }
  ]
}