{
  "status": "Recovered user decisions; this file does not claim a new 439-image run has completed.",
  "source_chat": {
    "title": "Image analysis experiment with DINO x AnyUp",
    "thread_id": "01a0cf41-dcd4-7bc1-914d-d2e241bbcbf1",
    "final_decision_user_text": "OK now we conclude that we keep lambda of 4 as our decision. we keep both k-means and GMM, also both pca + pca-varimax as visualization. We keep 5% and 10% cropping as the main approach. we just need to conclude here.",
    "prior_explicit_user_decision": "OK, now I conclude that 75% overlap is necessary process in the workflow. I discard AnyUp idea."
  },
  "photo_scope": {
    "unique_images": 439,
    "source_entries": 444,
    "archive": "KICS_ZERT_2025.zip",
    "source_folder": "2025/Annotation Indicator/Plot view/all_farms_plot/",
    "identity": "source SHA-256; preserve source IDs and duplicate aliases"
  },
  "adopted_local_workflow": {
    "feature_model": "facebook/dinov3-vitb16-pretrain-lvd1689m",
    "revision": "5931719e67bbdb9737e363e781fb0c67687896bc",
    "local_crop_fractions": [
      0.05,
      0.1
    ],
    "crop_reference": "shorter side of the chosen working image",
    "crop_encoder_pixels": [
      384,
      384
    ],
    "orientation_averaging": [
      "identity",
      "horizontal",
      "vertical",
      "both"
    ],
    "overlap_fraction": 0.75,
    "stride_fraction": 0.25,
    "blending": "separable Hann, per-axis floor 0.05; normalize accumulated weights",
    "upsampling": "bilinear token interpolation; no AnyUp",
    "whitening": false,
    "additional_unit_vector_normalization": false,
    "spatial_coordinates_in_cluster_distance": false
  },
  "adopted_clustering": {
    "algorithms": [
      "automatic_kmeans",
      "full_covariance_gmm"
    ],
    "kmeans": {
      "candidate_k": [
        3,
        4,
        5,
        6,
        7,
        8,
        9,
        10,
        11,
        12
      ],
      "selection": "three spatial folds, held-out silhouette; choose K closest to 5 among candidates within one SE of best; ties prefer smaller K",
      "final_fit_samples_maximum": 8192,
      "n_init": 10,
      "seed": 42
    },
    "gmm": {
      "candidate_k": [
        1,
        2,
        3,
        4,
        5,
        6,
        7,
        8,
        9,
        10,
        11,
        12
      ],
      "covariance_type": "full",
      "lambda": 4,
      "criterion": "-2 * log_likelihood + lambda * parameter_count * log(n)",
      "parameter_count": "K * (d + d*(d+1)/2) + K - 1",
      "selection": "minimum converged score; smaller K for exact ties",
      "reg_covar": 1e-05,
      "n_init": 3,
      "max_iter": 300,
      "retry_max_iter": 600,
      "tol": 0.001,
      "seed": 42,
      "hard_assignment": "maximum responsibility; retain area fractions and confidence"
    }
  },
  "adopted_display": [
    "PCA",
    "PCA-varimax"
  ],
  "dimensions_tested": [
    3,
    16
  ],
  "dimension_decision_note": "Both 3 and 16 were tested. The final user conclusion did not fix one dimensionality as the exclusive primary representation.",
  "current_full_cohort_baseline": {
    "taxonomic_threshold": 0.7,
    "taxonomic_run": "/home/masahiro/kics-zert2/runs/moin_plot_all_20260928/taxonomic_threshold_0p7",
    "taxonomic_region_source": "older AnyUp/cue-policy pipeline",
    "visual_source": "whole-image raw 768D L2-normalized patches; Rao Q, L2N2 and Leiden persistent groups; 10x250 samples, 5 graph replicates",
    "unnamed_taxa": "excluded from all named diversity calculations",
    "semantic_vegetation_scoring_mask": false
  },
  "required_extensions": [
    "Campaign and season family accumulation",
    "Taxonomic beta diversity and NMDS",
    "Shared cross-image DINO PCA",
    "Explicit structural alpha/beta/gamma definitions in a common representation",
    "Apply finalized 5%/10% K-means and lambda4 GMM protocol to the full target cohort if replacing old region maps"
  ],
  "cross_image_constraint": "Per-photo PCA axes and cluster IDs have no shared biological or numerical meaning. Use raw common DINO coordinates or a common fitted representation for cross-photo PCA and structural beta/gamma.",
  "input_extent_constraint": "Latest benchmark MOIN inputs were native 2048x2048 center windows; older full cohort used complete frames downscaled to longest side1280. Record and distinguish these extents before transferring settings."
}