{
  "status": "proposed_unstarted_not_a_locked_registration",
  "created_date": "2026-10-10",
  "purpose": "Concrete expanded cross-modal sensitivity matrix for the next extraction pilot",
  "strict_full_paper_family_holdout": false,
  "vision_candidates": [
    {
      "checkpoint": "microsoft/resnet-50",
      "source": "https://huggingface.co/microsoft/resnet-50",
      "objective": "supervised ImageNet-1k classification",
      "readout": "global average pooled features before classifier"
    },
    {
      "checkpoint": "facebook/vit-mae-base",
      "source": "https://huggingface.co/facebook/vit-mae-base",
      "objective": "masked pixel reconstruction on ImageNet-1k",
      "readout": "disable masking and mean-pool patch tokens"
    }
  ],
  "language_candidates": [
    {
      "checkpoint": "facebook/contriever",
      "source": "https://huggingface.co/facebook/contriever",
      "objective": "unsupervised contrastive retrieval",
      "readout": "published attention-mask-aware mean pooling",
      "previous_paper_use": "generation experiments"
    },
    {
      "checkpoint": "sentence-transformers/gtr-t5-base",
      "source": "https://huggingface.co/sentence-transformers/gtr-t5-base",
      "training_reference": "https://arxiv.org/abs/2112.07899",
      "objective": "semantic retrieval with multistage training including supervised MS MARCO",
      "readout": "published SentenceTransformer pipeline",
      "previous_paper_use": "Natural Questions experiments"
    }
  ],
  "proposed_pilot": {
    "all_four_crossings_required": true,
    "bridge_pair": [
      "dinov2_vit-b14@224_mean",
      "mpnet"
    ],
    "fit_seeds": [
      0,
      1,
      2
    ],
    "total_fits": 15,
    "data_arm": "fully_dedup_known_metadata_source_identity_only",
    "settings": "same reduced pipeline pilot profile as PROTOCOL.md",
    "extract_encoders_sequentially": true,
    "freeze_before_extraction": [
      "checkpoint revisions",
      "readouts",
      "image preprocessing",
      "text truncation",
      "selected rows",
      "metric definitions"
    ],
    "report_all_combinations_and_failures": true,
    "outcome_based_selection_allowed": false
  },
  "limits": [
    "Contriever and GTR are held out from the main 7x3 grid, not from the full paper; this does not fulfill phase 3's strict unused-family criterion.",
    "Both vision models use ImageNet; differences in architecture and objective do not establish independent data provenance.",
    "Failure under reduced optimization can reflect solver effort rather than intrinsic geometric incompatibility.",
    "The matrix varies several model attributes together and does not causally isolate architecture, objective, or corpus.",
    "No new encoder was extracted or trained in the completed overlap pipeline pilot."
  ],
  "optional_separate_base_model_stress_candidates": [
    {
      "checkpoint": "answerdotai/ModernBERT-base",
      "source": "https://huggingface.co/answerdotai/ModernBERT-base",
      "caveat": "Masked language base model; retrieval competence and pooling are confounds. Verify absence from complete paper/config inventory before registration."
    },
    {
      "checkpoint": "openai-community/gpt2",
      "source": "https://huggingface.co/openai-community/gpt2",
      "caveat": "Raw generative hidden states are not trained sentence embeddings; complete pretraining corpus unavailable. Verify absence from full inventory before registration."
    }
  ]
}