{
  "generated_at": "2026-09-25T04:55:32Z",
  "study": {
    "title": "Vector Retrieval Drift under Corpus Growth",
    "venue_target": "IEEE Access",
    "bench": "EnterpriseRAG-Bench",
    "primary_n": 200,
    "ladder": [
      5000,
      10000,
      15000,
      20000,
      25000,
      40000,
      50000,
      75000,
      100000
    ],
    "seed": 42,
    "goal": "Measure how dense retrieval Hit@k declines as an enterprise RAG corpus grows from 5k to 100k documents, quantify metadata rescue (\u0394_meta), and test whether the drift pattern replicates on a second embedding model (Amazon Titan Text Embeddings V2) under the same gold-pinned ladder."
  },
  "openai_primary": {
    "model": "text-embedding-3-small",
    "raw_hit_at_1_5k_to_100k": [
      0.6,
      0.305
    ],
    "raw_hit_at_10_5k_to_100k": [
      0.795,
      0.51
    ],
    "raw_mrr_5k_to_100k": [
      0.658,
      0.365
    ],
    "delta_meta_hit10_at_100k": 0.13,
    "n_star_tau_0_10": null,
    "n_star_last_ge": 100000,
    "fit_hit_at_10": {
      "a": 1.4740926693219225,
      "b": 0.08578610511454804,
      "form": "a - b*log(N)"
    }
  },
  "titan_a8": {
    "model": "amazon.titan-embed-text-v2:0",
    "status": "complete",
    "all_three_raw_metrics_replicate": true,
    "endpoint_raw": {
      "hit_at_1": {
        "openai_5k": 0.6,
        "openai_100k": 0.305,
        "openai_delta": -0.295,
        "titan_5k": 0.62,
        "titan_100k": 0.385,
        "titan_delta": -0.235,
        "replicates": true
      },
      "hit_at_5": {
        "openai_5k": 0.74,
        "openai_100k": 0.46,
        "openai_delta": -0.28,
        "titan_5k": 0.73,
        "titan_100k": 0.565,
        "titan_delta": -0.165,
        "replicates": true
      },
      "hit_at_10": {
        "openai_5k": 0.795,
        "openai_100k": 0.51,
        "openai_delta": -0.285,
        "titan_5k": 0.785,
        "titan_100k": 0.61,
        "titan_delta": -0.175,
        "replicates": true
      },
      "mrr": {
        "openai_5k": 0.6576,
        "openai_100k": 0.3647,
        "openai_delta": -0.2929,
        "titan_5k": 0.6714,
        "titan_100k": 0.465,
        "titan_delta": -0.2064,
        "replicates": true
      }
    },
    "exact_control": {
      "titan_identical_top10_at_5k_50k_100k": true,
      "titan_hit10_disagreements": 0,
      "openai_hit10_exact_minus_hnsw_at_50k": 0.085,
      "openai_hit10_exact_minus_hnsw_at_100k": 0.075,
      "note": "Titan HNSW matches exact cosine top-10 for 200/200 questions at 5k/50k/100k. Most of the OpenAI\u2013Titan HNSW gap is ANN recall loss on the OpenAI index; under exact search the embedder drops are statistically indistinguishable."
    },
    "bootstrap": {
      "B": 5000,
      "seed": 42,
      "hnsw_hit10_drop_titan_minus_openai": {
        "estimate": 0.11,
        "ci95": [
          0.04,
          0.175
        ]
      },
      "exact_hit10_drop_titan_minus_openai": {
        "estimate": 0.035,
        "ci95": [
          -0.025,
          0.1
        ]
      }
    }
  },
  "static_json_catalog": [
    {
      "path": "data/published/landing_summary.json",
      "role": "This file \u2014 compact study + Titan headline numbers"
    },
    {
      "path": "data/published/CANONICAL_METRICS.json",
      "role": "OpenAI integrity-clean promotion record + fit coeffs"
    },
    {
      "path": "data/published/erb_full_primary200_to100k_fit.json",
      "role": "OpenAI ladder fit + bootstrap CIs"
    },
    {
      "path": "data/published/openai_vs_titan_primary200.json",
      "role": "Full OpenAI vs Titan ladder comparison"
    },
    {
      "path": "data/published/openai_vs_titan_bootstrap.json",
      "role": "Paired bootstrap (B=5000) HNSW vs exact"
    },
    {
      "path": "data/published/exact_control_titan.json",
      "role": "Titan exact cosine vs HNSW identity check"
    },
    {
      "path": "data/published/exact_control_mvp.json",
      "role": "OpenAI exact cosine vs HNSW (ANN gap)"
    },
    {
      "path": "data/published/primary_questions_200.json",
      "role": "Primary-200 question bank"
    },
    {
      "path": "data/published/paper_metrics_for_tex.json",
      "role": "Numbers wired into Access LaTeX"
    },
    {
      "path": "data/verification/",
      "role": "Per-query verify bundle for /verify/ UI"
    }
  ],
  "artifacts_not_on_pages": [
    {
      "path": "artifacts/published/erb_full_primary200_to100k.json",
      "role": "Full OpenAI sweep (~4MB; copied to Pages when present in clone)"
    },
    {
      "path": "artifacts/published/erb_titan_primary200_to100k.json",
      "role": "Full Titan sweep (~3.6MB; copied to Pages when present)"
    },
    {
      "path": "GitHub Releases erb_tables.dump.gz",
      "role": "Postgres+pgvector dump with embeddings (~1GB)"
    }
  ]
}