{ "_meta": { "generated": "2026-07-06", "generator": "site-content/build_metrics.py", "hard_rules": [ "Every number carries a source path. Never adjust, estimate, or fill gaps \u2014 missing = \"TBD\".", "Statuses are ratification states, not quality claims. Visual verdicts belong to the user only.", "kid_hp_pooled_x1e3 (2.5 m suite) and loop 'kid' (0.5 m fixture) are different scales/planes \u2014 never compare or convert." ], "status_vocabulary": { "ratified": "user decision recorded in reports/README.md", "pending_signoff": "measured; awaiting user panel verdict before any re-ranking", "measured": "computed fact; no decision attached", "historical_proxy": "re-run stand-in for a lost artifact", "external_baseline": "public model run on our inputs for comparison", "baseline_control": "classical floor / sanity control", "calibration": "GT / null-floor calibration arm" }, "metric_definitions": { "kid_hp_pooled_x1e3": "KID \u00d710\u00b3 between DINOv2 features of high-pass-filtered 192\u00b2 patches and the stratified real-HR crop bank; lower = texture statistics closer to real aerial (reports/phase0_baseline_table.md)", "dists": "DISTS-official perceptual structure+texture distance vs GT (reports/phase0_baseline_table.md)", "lpips": "LPIPS perceptual distance vs GT \u2014 recorded but demoted from decision metric (date-confounded; CLAUDE.md goal statement)", "psd_dist": "radial power-spectral-density distance r\u2208[10,220) vs GT \u2014 texture spectrum shape (reports/phase0_baseline_table.md)", "hf_frac": "pinned def-A top-third-band spectral energy fraction, Nyquist-truncated; tile-set-dependent \u2014 state the tile set (reports/metrology_pinned.md)", "block_p16": "16-px-pitch blockiness score; 1.0 = no periodic block artifact (reports/phase0_baseline_table.md)", "roundtrip_corr": "correlation of output downsampled back to 10 m vs the input \u2014 LR-consistency / anti-hallucination (reports/metrology_pinned.md pinned protocol)", "judge_2afc_opus": "LLM-judge 2AFC real-vs-output; J\u21920.5 = indistinguishable; UNVALIDATED (reports/suite_v2/validity_gate.json)", "user_2afc_selftest": "the user's own forced-choice accuracy; 0.5 = goal (docs/plans/plane25_indistinguishability.md)", "opensr_correctness_triad": "opensr-test hallucination/omission/improvement decomposition (reports/suite_v2/opensr_triad*.json)" }, "validity_gate": { "status": "METRIC_HALF_VALID; JUDGE_UNVALIDATED_STOPPED", "source": "reports/suite_v2/validity_gate.json" } }, "models": { "HR": { "label": "Real HR aerial (GT) \u2014 calibration / null floor", "kind": "calibration", "status": "calibration", "plane": "2.5m", "metrics": { "lpips": 0.0, "dists": 1e-05, "hf_frac": 0.00187, "psd_dist": 2e-05, "block_p16": 0.98796, "kid_hp_pooled_x1e3": -3.3493, "kid_hp_pooled_std": 43.2532 }, "per_stratum_kid_hp": { "residential": { "n_crops": 123, "kid_hp_x1e3": -32.8335 }, "industrial_port": { "n_crops": 93, "kid_hp_x1e3": 2.8793 }, "ag_row": { "n_crops": 91, "kid_hp_x1e3": -83.9315 }, "forest": { "n_crops": 19, "kid_hp_x1e3": -908.5978, "small_n_caveat": true }, "water_coast": { "n_crops": 39, "kid_hp_x1e3": -165.4324 } }, "n_tiles": 73, "provenance": { "arm": "HR", "ckpt": "hr", "wkey": "G_ema", "input": "synth", "n": 73, "generated": "suite_v2 gen" }, "date": "2026-07-05", "sources": [ "reports/suite_v2/scores/HR.json", "reports/phase0_baseline_table.md" ] }, "bicubic_norm": { "label": "Bicubic \u00d74, radiometry-normalized (honest classical floor)", "kind": "baseline_control", "status": "baseline_control", "plane": "2.5m", "metrics": { "lpips": 0.72003, "dists": 0.34212, "hf_frac": 1e-05, "psd_dist": 1.59374, "block_p16": 0.90022, "kid_hp_pooled_x1e3": 18296.8515, "kid_hp_pooled_std": 0.0009 }, "per_stratum_kid_hp": { "residential": { "n_crops": 123, "kid_hp_x1e3": 23560.5316 }, "industrial_port": { "n_crops": 93, "kid_hp_x1e3": 20354.5206 }, "ag_row": { "n_crops": 91, "kid_hp_x1e3": 16365.8543 }, "forest": { "n_crops": 19, "kid_hp_x1e3": 17422.5241, "small_n_caveat": true }, "water_coast": { "n_crops": 39, "kid_hp_x1e3": 21171.2402 } }, "n_tiles": 73, "provenance": { "arm": "bicubic_norm", "ckpt": "bicubic", "wkey": "G_ema", "input": "synth", "n": 73, "generated": "suite_v2 gen" }, "date": "2026-07-05", "sources": [ "reports/suite_v2/scores/bicubic_norm.json", "reports/phase0_baseline_table.md" ], "judge_2afc_opus": { "J": 0.775, "J_ci95": [ 0.662, 0.887 ], "fool_pts": 22.5, "J_per_stratum": { "residential": 1.0, "industrial_port": 0.625, "ag_row": 0.75, "forest": 0.75, "water_coast": 0.75 }, "hr_hr_fpr": 0.333, "repeat_agreement": 0.438, "judge_model": "claude-opus-4-8", "prompt_v": 2, "status": "UNVALIDATED \u2014 recorded facts, not decision inputs. Repeat-agreement failed both prompt versions (v1 0.75 / v2 0.625 < 0.8) and the v2 prompt destroyed blur sensitivity; judge redesign goes to human review (reports/suite_v2/validity_gate.json, reports/phase0_baseline_table.md).", "source": "reports/suite_v2/judge/bicubic_norm.json" }, "pinned_protocol_metrics": { "lpips_honest_floor": { "value": 0.72, "note": "one-time recompute with TRAIN-split affine normalization; the historically recorded 0.841 conflated blur with a radiometric domain offset", "source": "reports/metrology_pinned.md" } } }, "difffusr_rgb": { "label": "DiffFuSR (public fusion SR) on real S2, RGB", "kind": "external_baseline", "status": "external_baseline", "caveat": "real-input LPIPS/PSD content-confounded (different-date GT)", "plane": "2.5m", "metrics": { "lpips": 0.44703, "dists": 0.29607, "hf_frac": 0.00048, "psd_dist": 0.31574, "block_p16": 0.99488, "kid_hp_pooled_x1e3": 22757.4959, "kid_hp_pooled_std": 0.0 }, "per_stratum_kid_hp": { "residential": { "n_crops": 123, "kid_hp_x1e3": 29680.3131 }, "industrial_port": { "n_crops": 93, "kid_hp_x1e3": 25423.3786 }, "ag_row": { "n_crops": 91, "kid_hp_x1e3": 18622.4829 }, "forest": { "n_crops": 19, "kid_hp_x1e3": 21170.6377, "small_n_caveat": true }, "water_coast": { "n_crops": 39, "kid_hp_x1e3": 26254.8939 } }, "n_tiles": 73, "provenance": { "arm": "difffusr_rgb", "model": "DiffFuSR RGB SR (DiffBlindSR3AtomJIFModel: RRDBNet encoder + SR3-atom diffusion, EMA weights, fast DDIM skip=50 (20 steps), eta=0), x4 10m->2.5m", "code": "/ceph/grid/home/jd3099/projects/satellite_upscaling_new/data/third_party/DiffFuSR (git commit 48a087b5345392ba50939109594dc81d53b642ba)", "ckpt": { "last.ckpt": { "path": "/ceph/grid/home/jd3099/projects/satellite_upscaling_new/data/hf_cache/hub/models--NorskRegnesentralSTI--DiffFuSR/snapshots/c907d3da7b32bff40ab5f8958b1d79670f83318e/logs/blindsrsnf_aniso_naip_degraded_harm_large/version_1/checkpoint/last.ckpt", "sha256": "60535465a1ec44c80c7f426001eb110a42966f577cdb08ee3e0822b71de1a3df" }, "choice": "naip_degraded_harm v1 (blindsrsnf_aniso_naip_degraded_harm_large/version_1): NAIP-degraded training WITH harmonization to Sentinel-2 radiometry (hparams train.harmonization=true) -> matches our S2 reflectance input; per task instruction over not_harm / worldstrat variants." }, "input_transform": "npz lr_coreg [3,128,128] RGB (B04,B03,B02) float32 reflectance ~[0,1], nan_to_num(0), clip>=0; rgb_range=1; model-internal norm mean=std=[0.5]*3", "output_transform": "test_step_lr_only()['sr_raw'] [3,512,512] denormalized reflectance; arm dir = pinned S2->HR affine then clip[0,1] float16; rawdom dir = raw reflectance", "command": "CUDA_VISIBLE_DEVICES=1 python reports/suite_v2/drivers/run_difffusr_rgb.py", "tiles": [ "brezice_00", "brezice_01", "koper_00", "koper_01", "koper_02", "koper_04", "nat_s_012_078", "nat_s_020_088", "nat_s_020_119", "nat_s_021_082", "nat_s_021_087", "nat_s_026_115", "nat_s_026_118", "nat_s_027_116", "nat_s_027_117", "nat_s_028_116", "nat_s_028_117", "nat_s_028_118", "nat_s_029_119", "nat_s_110_084", "nat_s_110_094", "nat_s_111_094", "nat_s_111_095", "nat_s_112_094", "nat_s_112_095", "nat_s_112_116", "nat_s_113_096", "nat_s_113_115", "nat_s_114_093", "nat_s_120_109", "nat_s_121_092", "nat_s_122_092", "nat_s_122_093", "nat_s_127_092", "nat_s_129_086", "nat_s_130_086", "nat_s_131_084", "nat_s_132_084", "nat_s_133_088", "nat_s_133_089", "nat_s_134_087", "nat_s_134_088", "nat_s_137_085", "nat_s_138_087", "nat_s_138_088", "nat_s_139_088", "nat_s_141_088", "nat_u_016_121", "nat_u_020_082", "nat_u_024_118", "nat_u_024_119", "nat_u_025_117", "nat_u_026_117", "nat_u_110_095", "nat_u_111_096", "nat_u_111_097", "nat_u_112_096", "nat_u_112_097", "nat_u_113_116", "nat_u_130_084", "nat_u_131_083", "novagorica_00", "novomesto_00", "novomesto_01", "novomesto_02", "novomesto_03", "novomesto_04", "novomesto_05", "novomesto_06", "novomesto_07", "tail_crop_08", "tail_tree_00", "tail_tree_01" ], "n_tiles": 73, "subset_rule": "full 73 val tiles unless first-tile time > 90 s (then first 24, sorted)", "sec_per_tile_mean": 0.5565418935801885, "sec_per_tile_median": 0.5442178249359131, "device": "cuda (L40S, CUDA_VISIBLE_DEVICES=1)", "written": "2026-07-05T12:59:06+02:00", "post_transform_affine": { "formula": "y = clip((x - mean_synth)/std_synth * std_hr + mean_hr, 0, 1), per band RGB", "constants_source": "reports/assets/sean/metrology/bicubic_domain.json corpus_stats_train (synth, hr)", "mean_synth": [ 0.0872078612446785, 0.09296771883964539, 0.06185305491089821 ], "std_synth": [ 0.05349373072385788, 0.04091332480311394, 0.04054270312190056 ], "mean_hr": [ 0.4506443440914154, 0.473564088344574, 0.4195850193500519 ], "std_hr": [ 0.19198717176914215, 0.16378414630889893, 0.16234740614891052 ] }, "rawdom_note": "_rawdom/.npz holds the model's raw S2-reflectance-domain RGB output (float16, nan->0, clipped at 0 only, NO affine, may exceed 1.0). /.npz holds the affine-normalized version that gets scored." }, "date": "2026-07-05", "sources": [ "reports/suite_v2/scores/difffusr_rgb.json", "reports/phase0_baseline_table.md" ], "opensr_correctness_triad": { "hallucination": 0.20925110143919787, "omission": 0.6133131061991056, "improvement": 0.17743579049905142, "n_tiles": 24, "caveat": "opensr-test v1.3.3 correctness triad, 24 tiles, real-LR arms only; absolutes carry the HR-date confound \u2014 read within-group (reports/phase0_baseline_table.md)", "source": "reports/suite_v2/opensr_triad.json" } }, "hpf100": { "label": "SEAN + high-pass-masked spectral fine-tune \u03bb=100 (EMA 3.8k ft)", "kind": "internal", "status": "pending_signoff", "caveat": "Best pooled KID-HP in all five strata, but the 2026-07-02 verdict for HPF was NEGATIVE on the PSD/hf axis; re-ranking is PENDING USER SIGN-OFF (reports/phase0_baseline_table.md finding 2). USER verdict on ab_hpf100 panels 2026-07-06 traced remaining geometry defects to the conditioning map, not the model (docs/plans/s_struct_v2.md)", "plane": "2.5m", "metrics": { "lpips": 0.22429, "dists": 0.14597, "hf_frac": 0.0009, "psd_dist": 0.12208, "block_p16": 1.00509, "kid_hp_pooled_x1e3": 2878.1424, "kid_hp_pooled_std": 0.0015 }, "per_stratum_kid_hp": { "residential": { "n_crops": 123, "kid_hp_x1e3": 2981.9107 }, "industrial_port": { "n_crops": 93, "kid_hp_x1e3": 3728.289 }, "ag_row": { "n_crops": 91, "kid_hp_x1e3": 4173.0885 }, "forest": { "n_crops": 19, "kid_hp_x1e3": 5328.4365, "small_n_caveat": true }, "water_coast": { "n_crops": 39, "kid_hp_x1e3": 3372.8443 } }, "n_tiles": 73, "provenance": { "arm": "hpf100", "ckpt": "outputs/sean/hpf100/sean_best.pt", "wkey": "G_ema", "input": "synth", "n": 73, "generated": "suite_v2 gen" }, "date": "2026-07-05", "sources": [ "reports/suite_v2/scores/hpf100.json", "reports/phase0_baseline_table.md" ], "judge_2afc_opus": { "J": 0.762, "J_ci95": [ 0.675, 0.85 ], "fool_pts": 23.8, "J_per_stratum": { "residential": 0.812, "industrial_port": 0.812, "ag_row": 0.875, "forest": 0.625, "water_coast": 0.688 }, "hr_hr_fpr": 0.167, "repeat_agreement": 0.562, "judge_model": "claude-opus-4-8", "prompt_v": 2, "status": "UNVALIDATED \u2014 recorded facts, not decision inputs. Repeat-agreement failed both prompt versions (v1 0.75 / v2 0.625 < 0.8) and the v2 prompt destroyed blur sensitivity; judge redesign goes to human review (reports/suite_v2/validity_gate.json, reports/phase0_baseline_table.md).", "source": "reports/suite_v2/judge/hpf100.json" } }, "hpf300": { "label": "SEAN + high-pass-masked spectral fine-tune \u03bb=300 (EMA 5.4k ft)", "kind": "internal", "status": "measured", "plane": "2.5m", "metrics": { "lpips": 0.23548, "dists": 0.15185, "hf_frac": 0.00065, "psd_dist": 0.16907, "block_p16": 1.00316, "kid_hp_pooled_x1e3": 3344.3779, "kid_hp_pooled_std": 0.0019 }, "per_stratum_kid_hp": { "residential": { "n_crops": 123, "kid_hp_x1e3": 3405.6232 }, "industrial_port": { "n_crops": 93, "kid_hp_x1e3": 3882.6687 }, "ag_row": { "n_crops": 91, "kid_hp_x1e3": 5760.898 }, "forest": { "n_crops": 19, "kid_hp_x1e3": 6888.9296, "small_n_caveat": true }, "water_coast": { "n_crops": 39, "kid_hp_x1e3": 4010.2604 } }, "n_tiles": 73, "provenance": { "arm": "hpf300", "ckpt": "outputs/sean/hpf300/sean_best.pt", "wkey": "G_ema", "input": "synth", "n": 73, "generated": "suite_v2 gen" }, "date": "2026-07-05", "sources": [ "reports/suite_v2/scores/hpf300.json", "reports/phase0_baseline_table.md" ] }, "ldsr_s2": { "label": "LDSR-S2 (S2-native public model) on real S2, RGBN", "kind": "external_baseline", "status": "external_baseline", "caveat": "real-input LPIPS/PSD content-confounded (different-date GT)", "plane": "2.5m", "metrics": { "lpips": 0.58633, "dists": 0.34333, "hf_frac": 5e-05, "psd_dist": 1.03329, "block_p16": 1.01914, "kid_hp_pooled_x1e3": 19174.9884, "kid_hp_pooled_std": 0.0009 }, "per_stratum_kid_hp": { "residential": { "n_crops": 123, "kid_hp_x1e3": 24309.8095 }, "industrial_port": { "n_crops": 93, "kid_hp_x1e3": 21390.0427 }, "ag_row": { "n_crops": 91, "kid_hp_x1e3": 17287.5633 }, "forest": { "n_crops": 19, "kid_hp_x1e3": 17564.1794, "small_n_caveat": true }, "water_coast": { "n_crops": 39, "kid_hp_x1e3": 22738.3896 } }, "n_tiles": 73, "provenance": { "arm": "ldsr_s2", "model": "LDSR-S2 (opensr-model SRLatentDiffusion, VAE z=4ch + UNet 160ch), RGBN x4 10m->2.5m, DDIM 100 steps, eta 0.95, temperature 1.0, encode_conditioning=True, apply_normalization=False (v1.0.0), histogram_matching(spe_cor)=True (default)", "code": "/ceph/grid/home/jd3099/projects/satellite_upscaling_new/data/third_party/opensr-model (git commit 10f4c01cc8172586841ea9e78c6de9939da47337), config /ceph/grid/home/jd3099/projects/satellite_upscaling_new/data/third_party/opensr-model/opensr_model/configs/config_10m.yaml", "ckpt": { "opensr-ldsrs2_v1_0_0.ckpt": { "path": "/ceph/grid/home/jd3099/projects/satellite_upscaling_new/data/hf_cache/hub/models--simon-donike--RS-SR-LTDF/snapshots/106e1ba972fdcf16443b2ecbfe934cf67235bfbf/opensr-ldsrs2_v1_0_0.ckpt", "sha256": "e2621e3912eb7c14867c3d20c9029607ba941be8e166dc09621860fcac27dc3a" }, "load": "state_dict via torch.load(weights_only=False), 'loss' keys dropped, strict=True into model.model (mirrors SRLatentDiffusion.load_pretrained)" }, "input_transform": "npz lr_coreg [3,128,128] RGB + lr_nir_coreg B08 -> RGBN [B04,B03,B02,B08] [1,4,128,128] float32 reflectance ~[0,1], nan_to_num(0), clip>=0; no rescale", "output_transform": "model out [1,4,512,512] reflectance (clipped >=0 + per-band histogram match to LR inside model); keep ch 0:3 (B04,B03,B02 -> R,G,B); arm dir = pinned S2->HR affine then clip[0,1] float16; rawdom dir = raw reflectance", "command": "CUDA_VISIBLE_DEVICES=1 python reports/suite_v2/drivers/run_ldsr_s2.py", "tiles": [ "brezice_00", "brezice_01", "koper_00", "koper_01", "koper_02", "koper_04", "nat_s_012_078", "nat_s_020_088", "nat_s_020_119", "nat_s_021_082", "nat_s_021_087", "nat_s_026_115", "nat_s_026_118", "nat_s_027_116", "nat_s_027_117", "nat_s_028_116", "nat_s_028_117", "nat_s_028_118", "nat_s_029_119", "nat_s_110_084", "nat_s_110_094", "nat_s_111_094", "nat_s_111_095", "nat_s_112_094", "nat_s_112_095", "nat_s_112_116", "nat_s_113_096", "nat_s_113_115", "nat_s_114_093", "nat_s_120_109", "nat_s_121_092", "nat_s_122_092", "nat_s_122_093", "nat_s_127_092", "nat_s_129_086", "nat_s_130_086", "nat_s_131_084", "nat_s_132_084", "nat_s_133_088", "nat_s_133_089", "nat_s_134_087", "nat_s_134_088", "nat_s_137_085", "nat_s_138_087", "nat_s_138_088", "nat_s_139_088", "nat_s_141_088", "nat_u_016_121", "nat_u_020_082", "nat_u_024_118", "nat_u_024_119", "nat_u_025_117", "nat_u_026_117", "nat_u_110_095", "nat_u_111_096", "nat_u_111_097", "nat_u_112_096", "nat_u_112_097", "nat_u_113_116", "nat_u_130_084", "nat_u_131_083", "novagorica_00", "novomesto_00", "novomesto_01", "novomesto_02", "novomesto_03", "novomesto_04", "novomesto_05", "novomesto_06", "novomesto_07", "tail_crop_08", "tail_tree_00", "tail_tree_01" ], "n_tiles": 73, "subset_rule": "full 73 val tiles unless first-tile time > 90 s (then first 24, sorted)", "sec_per_tile_mean": 1.190396246844775, "sec_per_tile_median": 1.1838843822479248, "device": "cuda (L40S, CUDA_VISIBLE_DEVICES=1)", "written": "2026-07-05T13:02:09+02:00", "post_transform_affine": { "formula": "y = clip((x - mean_synth)/std_synth * std_hr + mean_hr, 0, 1), per band RGB", "constants_source": "reports/assets/sean/metrology/bicubic_domain.json corpus_stats_train (synth, hr)", "mean_synth": [ 0.0872078612446785, 0.09296771883964539, 0.06185305491089821 ], "std_synth": [ 0.05349373072385788, 0.04091332480311394, 0.04054270312190056 ], "mean_hr": [ 0.4506443440914154, 0.473564088344574, 0.4195850193500519 ], "std_hr": [ 0.19198717176914215, 0.16378414630889893, 0.16234740614891052 ] }, "rawdom_note": "_rawdom/.npz holds the model's raw S2-reflectance-domain RGB output (float16, nan->0, clipped at 0 only, NO affine, may exceed 1.0). /.npz holds the affine-normalized version that gets scored." }, "date": "2026-07-05", "sources": [ "reports/suite_v2/scores/ldsr_s2.json", "reports/phase0_baseline_table.md" ], "opensr_correctness_triad": { "hallucination": 0.1433683419600129, "omission": 0.7448902676502863, "improvement": 0.11174140156557162, "n_tiles": 24, "caveat": "opensr-test v1.3.3 correctness triad, 24 tiles, real-LR arms only; absolutes carry the HR-date confound \u2014 read within-group (reports/phase0_baseline_table.md)", "source": "reports/suite_v2/opensr_triad.json" } }, "osediff_zs": { "label": "OSEDiff zero-shot (one-step Real-ISR) on real S2", "kind": "external_baseline", "status": "external_baseline", "caveat": "real-input LPIPS/PSD content-confounded (different-date GT)", "plane": "2.5m", "metrics": { "lpips": 0.46465, "dists": 0.24277, "hf_frac": 0.0003, "psd_dist": 0.42511, "block_p16": 0.98386, "kid_hp_pooled_x1e3": 9301.6468, "kid_hp_pooled_std": 0.0014 }, "per_stratum_kid_hp": { "residential": { "n_crops": 123, "kid_hp_x1e3": 12290.9159 }, "industrial_port": { "n_crops": 93, "kid_hp_x1e3": 12746.1214 }, "ag_row": { "n_crops": 91, "kid_hp_x1e3": 8847.3276 }, "forest": { "n_crops": 19, "kid_hp_x1e3": 8359.6193, "small_n_caveat": true }, "water_coast": { "n_crops": 39, "kid_hp_x1e3": 8919.8713 } }, "n_tiles": 73, "provenance": { "arm": "osediff_zs", "date": "2026-07-05", "model": "OSEDiff (NeurIPS 2024, one-step Real-ISR; SD-2.1-base + VAE-encoder/UNet LoRA, RAM+DAPE prompt extraction), fp16, align_method=adain", "code": "data/third_party/OSEDiff (shallow clone cdd3b7b, Apache-2.0) with local-only compat shims, base venv untouched: (1) models/{autoencoder_kl,unet_2d_condition}.py replaced by diffusers-0.38 subclass shims mirrored from the PiSA-SR clone (.orig kept); (2) my_utils/vaehook.py tiled-VAE fp16 zeros dtype fix (mirrored); (3) ram/models/utils.py BertTokenizer dead hardcoded path -> staged google-bert/bert-base-uncased snapshot 86b5e0934494bd15c9632b12f734a8a67f723594; (4) ram/models/{bert,bert_lora}.py: apply_chunking_to_forward/find_pruneable_heads_and_indices/prune_linear_layer imports moved to transformers.pytorch_utils (transformers 4.57); (5) ram/models/vit.py fairscale import guarded (only needed for training-time grad checkpointing); (6) osediff.py set_timesteps(1) -> set_timesteps(timesteps=[999]) at 3 call sites (diffusers 0.38 previous_timestep requires t in scheduler.timesteps; custom [999] reproduces the diffusers-0.25 one-step prev_t=-1 semantics, zero added variance)", "venv": "opensr_eval overlay venv (system-site-packages over base satellite_upscaling venv; torch 2.11+cu130 from base); loralib installed into the OVERLAY only (pip --no-deps), required by ram/models/{bert_lora,swin_transformer_lora}.py", "weights": { "sd21_base": "data/hf_cache/hub/models--RedbeardNZ--stable-diffusion-2-1-base/snapshots/c6a5e9bab8d874d081de76fa270ae0aefa5410ff (official stabilityai repo delisted; mirror verified 2026-07-03)", "osediff_lora": "data/third_party/OSEDiff/preset/models/osediff.pkl (20072053 B, bundled in upstream repo)", "ram_base": "data/hf_cache/hub/models--xinyu1205--recognize_anything_model/snapshots/ebc52dc741e86466202a5ab8ab22eae6e7d48bf1/ram_swin_large_14m.pth (5625634877 B)", "dape": "data/third_party/OSEDiff/preset/models/DAPE.pth (7194489 B, bundled)" }, "input": { "source": "REAL Sentinel-2 LR: key 'lr_coreg' float32 [3,128,128] in [0,1] S2-reflectance domain, from /ceph/grid/home/jd3099/projects/slovenia_s2dof_dataset/data/sat_data/slovenia_corpus/pairs/.npz", "tiles": "73 val tiles (sidecar json split=='val', skipping names with 'coreg', 'anchor_', '__S_core_v0')", "transform": "project pinned deployment affine (reports/assets/sean/metrology/bicubic_domain.json corpus_stats_train): x_norm = clip((lr - mean_synth)/std_synth * std_hr + mean_hr, 0, 1) per band; synth mean=[0.08720786420037976,0.09296771687732303,0.0618530544542574] std=[0.053493731443232585,0.040913325125573205,0.04054270127190933]; hr mean=[0.45064433408573007,0.4735640734750075,0.4195850191671701] std=[0.19198717263852538,0.16378415178390907,0.16234740020402846]; then round(x*255) 8-bit PNG 128x128", "staging_dir": "reports/suite_v2/inputs_real_norm/ (same staging as pisasr_zs arm)" }, "command": "cd data/third_party/OSEDiff && source /ceph/grid/home/jd3099/venvs/opensr_eval/bin/activate && CUDA_VISIBLE_DEVICES=0 HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 HF_HOME=$REPO/data/hf_cache PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True TMPDIR=/tmp python test_osediff.py -i $REPO/reports/suite_v2/inputs_real_norm -o --osediff_path preset/models/osediff.pkl --pretrained_model_name_or_path $REPO/data/hf_cache/hub/models--RedbeardNZ--stable-diffusion-2-1-base/snapshots/c6a5e9bab8d874d081de76fa270ae0aefa5410ff --ram_path $REPO/data/hf_cache/hub/models--xinyu1205--recognize_anything_model/snapshots/ebc52dc741e86466202a5ab8ab22eae6e7d48bf1/ram_swin_large_14m.pth --ram_ft_path preset/models/DAPE.pth --upscale 4 --process_size 512 --align_method adain --mixed_precision fp16", "runner_behavior": "test_osediff.py resizes 128x128 input x4 LANCZOS to 512x512; RAM+DAPE extracts tag prompt at 384x384; one-step SD-2.1 LoRA UNet at t=999; output 512x512 8-bit RGB; adain color fix vs the LANCZOS-upsampled input; seed 42 (default); per-tile RAM prompts saved by the runner in the raw dir txt/ (not copied into arm dir)", "output_format": ".png (512x512 uint8 RGB, direct model output) + .npz key 'sr' float16 [3,512,512] in [0,1] (= png/255, CHW)", "sr_domain": "display/HR radiometric domain (the affine above maps S2 reflectance to HR display domain; outputs are NOT mapped back)" }, "date": "2026-07-05", "sources": [ "reports/suite_v2/scores/osediff_zs.json", "reports/phase0_baseline_table.md" ], "opensr_correctness_triad": { "hallucination": 0.23510913737118244, "omission": 0.6044419954220454, "improvement": 0.16044886813809475, "n_tiles": 24, "caveat": "opensr-test v1.3.3 correctness triad, 24 tiles, real-LR arms only; absolutes carry the HR-date confound \u2014 read within-group (reports/phase0_baseline_table.md)", "source": "reports/suite_v2/opensr_triad.json" } }, "pisasr_zs": { "label": "PiSA-SR zero-shot (SD2.1 one-step Real-ISR) on real S2", "kind": "external_baseline", "status": "external_baseline", "caveat": "real-input LPIPS/PSD content-confounded (different-date GT)", "plane": "2.5m", "metrics": { "lpips": 0.42629, "dists": 0.23757, "hf_frac": 0.00038, "psd_dist": 0.33653, "block_p16": 0.98486, "kid_hp_pooled_x1e3": 7777.2966, "kid_hp_pooled_std": 0.0009 }, "per_stratum_kid_hp": { "residential": { "n_crops": 123, "kid_hp_x1e3": 11291.6756 }, "industrial_port": { "n_crops": 93, "kid_hp_x1e3": 10426.8126 }, "ag_row": { "n_crops": 91, "kid_hp_x1e3": 6237.4506 }, "forest": { "n_crops": 19, "kid_hp_x1e3": 9306.2399, "small_n_caveat": true }, "water_coast": { "n_crops": 39, "kid_hp_x1e3": 7933.0343 } }, "n_tiles": 73, "provenance": { "arm": "pisasr_zs", "date": "2026-07-05", "model": "PiSA-SR (CVPR 2025, one-step Real-ISR, SD-2.1-base + dual pixel/semantic LoRA), --default knob (merged dual-LoRA, lambda_pix=lambda_sem=1.0)", "code": "data/third_party/PiSA-SR (shallow clone, with 2026-07-03 diffusers-0.38 compat shims in src/models/)", "weights": { "sd21_base": "data/hf_cache/hub/models--RedbeardNZ--stable-diffusion-2-1-base/snapshots/c6a5e9bab8d874d081de76fa270ae0aefa5410ff (via symlink preset/models/stable-diffusion-2-1-base)", "pisa_sr_lora": "data/third_party/PiSA-SR/preset/models/pisa_sr.pkl (33MB dual-LoRA, SHA256 13502f7b...91f6)" }, "input": { "source": "REAL Sentinel-2 LR: key 'lr_coreg' float32 [3,128,128] in [0,1] S2-reflectance domain, from /ceph/grid/home/jd3099/projects/slovenia_s2dof_dataset/data/sat_data/slovenia_corpus/pairs/.npz", "tiles": "73 val tiles (sidecar json split=='val', skipping names with 'coreg', 'anchor_', '__S_core_v0')", "transform": "project pinned deployment affine (reports/assets/sean/metrology/bicubic_domain.json corpus_stats_train): x_norm = clip((lr - mean_synth)/std_synth * std_hr + mean_hr, 0, 1) per band; synth mean=[0.08720786420037976,0.09296771687732303,0.0618530544542574] std=[0.053493731443232585,0.040913325125573205,0.04054270127190933]; hr mean=[0.45064433408573007,0.4735640734750075,0.4195850191671701] std=[0.19198717263852538,0.16378415178390907,0.16234740020402846]; then round(x*255) 8-bit PNG 128x128", "staging_dir": "reports/suite_v2/inputs_real_norm/" }, "command": "cd data/third_party/PiSA-SR && CUDA_VISIBLE_DEVICES=0 HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 HF_HOME=$REPO/data/hf_cache PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True TMPDIR=/tmp python test_pisasr.py --pretrained_model_path preset/models/stable-diffusion-2-1-base --pretrained_path preset/models/pisa_sr.pkl --process_size 512 --upscale 4 --default --input_image $REPO/reports/suite_v2/inputs_real_norm --output_dir ", "runner_behavior": "test_pisasr.py resizes 128x128 input x4 LANCZOS to 512x512 before the one-step model; output 512x512 8-bit RGB; align_method=adain (default) color fix vs the LANCZOS-upsampled input; fp16; seed 42 (default)", "output_format": ".png (512x512 uint8 RGB, direct model output) + .npz key 'sr' float16 [3,512,512] in [0,1] (= png/255, CHW)", "sr_domain": "display/HR radiometric domain (the affine above maps S2 reflectance to HR display domain; outputs are NOT mapped back)" }, "date": "2026-07-05", "sources": [ "reports/suite_v2/scores/pisasr_zs.json", "reports/phase0_baseline_table.md" ], "opensr_correctness_triad": { "hallucination": 0.23520570496718088, "omission": 0.5971496949593226, "improvement": 0.16764460814495882, "n_tiles": 24, "caveat": "opensr-test v1.3.3 correctness triad, 24 tiles, real-LR arms only; absolutes carry the HR-date confound \u2014 read within-group (reports/phase0_baseline_table.md)", "source": "reports/suite_v2/opensr_triad.json" } }, "sean22k": { "label": "SEAN v5 @ ~21k steps (proxy re-run; original snapshot lost)", "kind": "internal", "status": "historical_proxy", "plane": "2.5m", "metrics": { "lpips": 0.2362, "dists": 0.15034, "hf_frac": 0.00108, "psd_dist": 0.09959, "block_p16": 1.00557, "kid_hp_pooled_x1e3": 4684.9056, "kid_hp_pooled_std": 0.0008 }, "per_stratum_kid_hp": { "residential": { "n_crops": 123, "kid_hp_x1e3": 5044.6503 }, "industrial_port": { "n_crops": 93, "kid_hp_x1e3": 5375.5939 }, "ag_row": { "n_crops": 91, "kid_hp_x1e3": 6186.1294 }, "forest": { "n_crops": 19, "kid_hp_x1e3": 8458.6475, "small_n_caveat": true }, "water_coast": { "n_crops": 39, "kid_hp_x1e3": 7268.4435 } }, "n_tiles": 73, "provenance": { "arm": "sean22k", "ckpt": "outputs/sean/v5_sweep/sean_022000.pt", "wkey": "G_ema", "input": "synth", "n": 73, "generated": "suite_v2 gen" }, "date": "2026-07-05", "sources": [ "reports/suite_v2/scores/sean22k.json", "reports/phase0_baseline_table.md" ] }, "sen2srlite": { "label": "SEN2SRLite (S2-native public model) on real S2, RGBN", "kind": "external_baseline", "status": "external_baseline", "caveat": "real-input LPIPS/PSD content-confounded (different-date GT)", "plane": "2.5m", "metrics": { "lpips": 0.73593, "dists": 0.37111, "hf_frac": 1e-05, "psd_dist": 1.4334, "block_p16": 1.10083, "kid_hp_pooled_x1e3": 19117.1389, "kid_hp_pooled_std": 0.0015 }, "per_stratum_kid_hp": { "residential": { "n_crops": 123, "kid_hp_x1e3": 25142.3054 }, "industrial_port": { "n_crops": 93, "kid_hp_x1e3": 21236.4017 }, "ag_row": { "n_crops": 91, "kid_hp_x1e3": 16903.4557 }, "forest": { "n_crops": 19, "kid_hp_x1e3": 17308.4528, "small_n_caveat": true }, "water_coast": { "n_crops": 39, "kid_hp_x1e3": 21153.3291 } }, "n_tiles": 73, "provenance": { "arm": "sen2srlite", "model": "SEN2SRLite NonReference_RGBN_x4 (CNNSR(4,4,24,4,True,False,6) + HardConstraint), 10m->2.5m x4", "code": "/ceph/grid/home/jd3099/projects/satellite_upscaling_new/data/third_party/SEN2SR (git commit 8e21bb669bcc6e8eb953a1fb24dfc5bb59dc18a4)", "ckpt": { "model.safetensor": { "path": "/ceph/grid/home/jd3099/projects/satellite_upscaling_new/data/hf_cache/hub/models--tacofoundation--SEN2SR/snapshots/469e9b3a049f9320ba66b3a39b9655dbd31480b3/SEN2SRLite/NonReference_RGBN_x4/model.safetensor", "sha256": "479aa796d5068d0b1206118ccbca27bd3223df0214db1a9b31a1e18349ed1c7e" }, "hard_constraint.safetensor": { "path": "/ceph/grid/home/jd3099/projects/satellite_upscaling_new/data/hf_cache/hub/models--tacofoundation--SEN2SR/snapshots/469e9b3a049f9320ba66b3a39b9655dbd31480b3/SEN2SRLite/NonReference_RGBN_x4/hard_constraint.safetensor", "sha256": "fbad981519066387c413ead1d6af7ef3e0d2947c34147ba90163fc79ae539239" } }, "input_transform": "npz lr_coreg [3,128,128] RGB + lr_nir_coreg B08 -> RGBN [B04,B03,B02,B08] [1,4,128,128] float32 reflectance ~[0,1], nan_to_num(0), clip>=0; no rescale", "output_transform": "model out [1,4,512,512] reflectance; keep ch 0:3 (B04,B03,B02 -> R,G,B); arm dir = pinned S2->HR affine then clip[0,1] float16; rawdom dir = raw reflectance", "command": "CUDA_VISIBLE_DEVICES=1 python reports/suite_v2/drivers/run_sen2srlite.py", "tiles": [ "brezice_00", "brezice_01", "koper_00", "koper_01", "koper_02", "koper_04", "nat_s_012_078", "nat_s_020_088", "nat_s_020_119", "nat_s_021_082", "nat_s_021_087", "nat_s_026_115", "nat_s_026_118", "nat_s_027_116", "nat_s_027_117", "nat_s_028_116", "nat_s_028_117", "nat_s_028_118", "nat_s_029_119", "nat_s_110_084", "nat_s_110_094", "nat_s_111_094", "nat_s_111_095", "nat_s_112_094", "nat_s_112_095", "nat_s_112_116", "nat_s_113_096", "nat_s_113_115", "nat_s_114_093", "nat_s_120_109", "nat_s_121_092", "nat_s_122_092", "nat_s_122_093", "nat_s_127_092", "nat_s_129_086", "nat_s_130_086", "nat_s_131_084", "nat_s_132_084", "nat_s_133_088", "nat_s_133_089", "nat_s_134_087", "nat_s_134_088", "nat_s_137_085", "nat_s_138_087", "nat_s_138_088", "nat_s_139_088", "nat_s_141_088", "nat_u_016_121", "nat_u_020_082", "nat_u_024_118", "nat_u_024_119", "nat_u_025_117", "nat_u_026_117", "nat_u_110_095", "nat_u_111_096", "nat_u_111_097", "nat_u_112_096", "nat_u_112_097", "nat_u_113_116", "nat_u_130_084", "nat_u_131_083", "novagorica_00", "novomesto_00", "novomesto_01", "novomesto_02", "novomesto_03", "novomesto_04", "novomesto_05", "novomesto_06", "novomesto_07", "tail_crop_08", "tail_tree_00", "tail_tree_01" ], "n_tiles": 73, "sec_per_tile_mean": 0.0160675048828125, "sec_per_tile_median": 0.0071752071380615234, "device": "cuda", "written": "2026-07-05T12:53:40+02:00", "post_transform_affine": { "formula": "y = clip((x - mean_synth)/std_synth * std_hr + mean_hr, 0, 1), per band RGB", "constants_source": "reports/assets/sean/metrology/bicubic_domain.json corpus_stats_train (synth, hr)", "mean_synth": [ 0.0872078612446785, 0.09296771883964539, 0.06185305491089821 ], "std_synth": [ 0.05349373072385788, 0.04091332480311394, 0.04054270312190056 ], "mean_hr": [ 0.4506443440914154, 0.473564088344574, 0.4195850193500519 ], "std_hr": [ 0.19198717176914215, 0.16378414630889893, 0.16234740614891052 ] }, "rawdom_note": "_rawdom/.npz holds the model's raw S2-reflectance-domain RGB output (float16, nan->0, clipped at 0 only, NO affine, may exceed 1.0). /.npz holds the affine-normalized version that gets scored." }, "date": "2026-07-05", "sources": [ "reports/suite_v2/scores/sen2srlite.json", "reports/phase0_baseline_table.md" ], "opensr_correctness_triad": { "hallucination": 0.13314056287830076, "omission": 0.7692302713791529, "improvement": 0.09762917133048177, "n_tiles": 24, "caveat": "opensr-test v1.3.3 correctness triad, 24 tiles, real-LR arms only; absolutes carry the HR-date confound \u2014 read within-group (reports/phase0_baseline_table.md)", "source": "reports/suite_v2/opensr_triad.json" } }, "v5_best": { "label": "SEAN v5 (EMA @ 48.7k) \u2014 the ratified Stage-A baseline", "kind": "internal", "status": "ratified", "plane": "2.5m", "metrics": { "lpips": 0.22529, "dists": 0.14524, "hf_frac": 0.00106, "psd_dist": 0.09999, "block_p16": 1.00442, "kid_hp_pooled_x1e3": 3262.0203, "kid_hp_pooled_std": 0.0026 }, "per_stratum_kid_hp": { "residential": { "n_crops": 123, "kid_hp_x1e3": 3462.4737 }, "industrial_port": { "n_crops": 93, "kid_hp_x1e3": 4340.2382 }, "ag_row": { "n_crops": 91, "kid_hp_x1e3": 4233.7938 }, "forest": { "n_crops": 19, "kid_hp_x1e3": 5945.4987, "small_n_caveat": true }, "water_coast": { "n_crops": 39, "kid_hp_x1e3": 3861.3335 } }, "n_tiles": 73, "provenance": { "arm": "v5_best", "ckpt": "outputs/sean/v5/sean_best.pt", "wkey": "G_ema", "input": "synth", "n": 73, "generated": "suite_v2 gen" }, "date": "2026-07-05", "sources": [ "reports/suite_v2/scores/v5_best.json", "reports/phase0_baseline_table.md" ], "judge_2afc_opus": { "J": 0.75, "J_ci95": [ 0.662, 0.838 ], "fool_pts": 25.0, "J_per_stratum": { "residential": 0.875, "industrial_port": 0.562, "ag_row": 0.812, "forest": 0.688, "water_coast": 0.812 }, "hr_hr_fpr": 0.25, "repeat_agreement": 0.562, "judge_model": "claude-opus-4-8", "prompt_v": 2, "status": "UNVALIDATED \u2014 recorded facts, not decision inputs. Repeat-agreement failed both prompt versions (v1 0.75 / v2 0.625 < 0.8) and the v2 prompt destroyed blur sensitivity; judge redesign goes to human review (reports/suite_v2/validity_gate.json, reports/phase0_baseline_table.md).", "source": "reports/suite_v2/judge/v5_best.json" }, "pinned_protocol_metrics": { "roundtrip_corr_10m": { "value": 0.898, "n_tiles": 73, "min_tile": 0.707, "definition": "avg-pool \u00d74 back to 10 m, zpearson per band, RGB mean \u2014 pinned protocol", "source": "reports/metrology_pinned.md", "raw": "reports/assets/sean/metrology/roundtrip_reconcile.json" }, "hf_frac_defA_12tile": { "value": 0.000757, "hr_reference": 0.000914, "definition": "pinned def-A (Nyquist-truncated radial PSD, top-third band r\u2208[170,256) on 512\u00b2); 12-tile set \u2014 NOT comparable to the 73-tile suite hf_frac", "source": "reports/metrology_pinned.md", "raw": "reports/assets/sean/metrology/hf_reconcile.json" }, "top_octave_band_power_ratio_vs_HR": { "value": 0.81, "note": "SEAN 38244 vs HR 47104, def-A band, 12-tile mean", "source": "reports/metrology_pinned.md" } } }, "v5_best_sub20": { "label": "v5_best on the 20-tile subset (control for 20-tile arms)", "kind": "internal", "status": "control_subset", "caveat": "20-tile arm: compare ONLY against other 20-tile arms; the subset alone shifts pooled KID +24% (reports/phase0_baseline_table.md)", "plane": "2.5m", "metrics": { "lpips": 0.2412, "dists": 0.15776, "hf_frac": 0.00084, "psd_dist": 0.09678, "block_p16": 1.00612, "kid_hp_pooled_x1e3": 4044.7475, "kid_hp_pooled_std": 0.003 }, "per_stratum_kid_hp": { "residential": { "n_crops": 40, "kid_hp_x1e3": 4203.0376 }, "industrial_port": { "n_crops": 33, "kid_hp_x1e3": 5229.2847 }, "ag_row": { "n_crops": 16, "kid_hp_x1e3": 4486.0146 }, "forest": { "n_crops": 5, "kid_hp_x1e3": 4663.0491, "small_n_caveat": true }, "water_coast": { "n_crops": 6, "kid_hp_x1e3": 7008.8636 } }, "n_tiles": 20, "provenance": null, "date": "2026-07-05", "sources": [ "reports/suite_v2/scores/v5_best_sub20.json", "reports/phase0_baseline_table.md" ] }, "v5_grain": { "label": "v5 + fitted sensor-grain overlay (probe 6e)", "kind": "internal_probe", "status": "measured", "caveat": "grain overlay makes KID-HP 3\u00d7 WORSE; \u0394-judge queued before the probe verdict is recorded (reports/phase0_baseline_table.md)", "plane": "2.5m", "metrics": { "lpips": 0.25795, "dists": 0.16843, "hf_frac": 0.00104, "psd_dist": 0.09972, "block_p16": 1.00436, "kid_hp_pooled_x1e3": 9185.8538, "kid_hp_pooled_std": 0.0006 }, "per_stratum_kid_hp": { "residential": { "n_crops": 123, "kid_hp_x1e3": 10004.3808 }, "industrial_port": { "n_crops": 93, "kid_hp_x1e3": 11500.2174 }, "ag_row": { "n_crops": 91, "kid_hp_x1e3": 10350.0084 }, "forest": { "n_crops": 19, "kid_hp_x1e3": 11634.1816, "small_n_caveat": true }, "water_coast": { "n_crops": 39, "kid_hp_x1e3": 10379.0459 } }, "n_tiles": 73, "provenance": { "arm": "v5_grain", "base": "v5_best", "overlay": { "grain_std_rgb": [ 0.02346, 0.01742, 0.01471 ], "residual_lag1_autocorr": 0.707, "blur_sigma": 1.202, "n_tiles_fit": 17, "hp_sigma": 1.5, "flat_thresh": 0.015 } }, "date": "2026-07-05", "sources": [ "reports/suite_v2/scores/v5_grain.json", "reports/phase0_baseline_table.md" ], "judge_2afc_opus": { "J": 0.875, "J_ci95": [ 0.787, 0.95 ], "fool_pts": 12.5, "J_per_stratum": { "residential": 1.0, "industrial_port": 1.0, "ag_row": 0.812, "forest": 0.812, "water_coast": 0.75 }, "hr_hr_fpr": 0.25, "repeat_agreement": 0.938, "judge_model": "claude-opus-4-8", "prompt_v": 2, "status": "UNVALIDATED \u2014 recorded facts, not decision inputs. Repeat-agreement failed both prompt versions (v1 0.75 / v2 0.625 < 0.8) and the v2 prompt destroyed blur sensitivity; judge redesign goes to human review (reports/suite_v2/validity_gate.json, reports/phase0_baseline_table.md).", "source": "reports/suite_v2/judge/v5_grain.json" } }, "v5_osediff_lf": { "label": "v5 + OSEDiff under LF hard constraint (probe T7, 20 tiles)", "kind": "internal_probe", "status": "measured", "caveat": "20-tile arm: compare ONLY against v5_best_sub20 (reports/phase0_baseline_table.md)", "plane": "2.5m", "metrics": { "lpips": 0.28776, "dists": 0.18156, "hf_frac": 0.00125, "psd_dist": 0.09053, "block_p16": 1.15133, "kid_hp_pooled_x1e3": 7437.602, "kid_hp_pooled_std": 0.0015 }, "per_stratum_kid_hp": { "residential": { "n_crops": 40, "kid_hp_x1e3": 10639.3456 }, "industrial_port": { "n_crops": 33, "kid_hp_x1e3": 8441.4688 }, "ag_row": { "n_crops": 16, "kid_hp_x1e3": 4092.7855 }, "forest": { "n_crops": 5, "kid_hp_x1e3": 5929.7621, "small_n_caveat": true }, "water_coast": { "n_crops": 6, "kid_hp_x1e3": 7516.5649 } }, "n_tiles": 20, "provenance": { "arm": "v5_osediff_lf", "base": "v5_best", "enh_dir": "reports/suite_v2/osediff_lf/enh", "lock": "avg-pool \u00d74 (10 m band)" }, "date": "2026-07-05", "sources": [ "reports/suite_v2/scores/v5_osediff_lf.json", "reports/phase0_baseline_table.md" ], "opensr_correctness_triad": { "hallucination": 0.2603218749165535, "omission": 0.22072902023792268, "improvement": 0.5189490973949432, "n_tiles": 20, "caveat": "opensr-test v1.3.3 correctness triad, 24 tiles, real-LR arms only; absolutes carry the HR-date confound \u2014 read within-group (reports/phase0_baseline_table.md)", "source": "reports/suite_v2/opensr_triad_realv5.json" } }, "v5_real_harm": { "label": "SEAN v5 on REAL Sentinel-2 input, inference-harmonized", "kind": "internal", "status": "measured", "caveat": "harmonization does NOT recover realism (KID-HP 4613 \u2248 4631) while shifting radiometry \u22120.043 mean luma (reports/phase0_baseline_table.md finding 3); real-input LPIPS/PSD content-confounded", "plane": "2.5m", "metrics": { "lpips": 0.35857, "dists": 0.19672, "hf_frac": 0.00083, "psd_dist": 0.16163, "block_p16": 1.00322, "kid_hp_pooled_x1e3": 4612.9963, "kid_hp_pooled_std": 0.0015 }, "per_stratum_kid_hp": { "residential": { "n_crops": 123, "kid_hp_x1e3": 5187.8052 }, "industrial_port": { "n_crops": 93, "kid_hp_x1e3": 6636.4979 }, "ag_row": { "n_crops": 91, "kid_hp_x1e3": 5073.9349 }, "forest": { "n_crops": 19, "kid_hp_x1e3": 11472.1798, "small_n_caveat": true }, "water_coast": { "n_crops": 39, "kid_hp_x1e3": 4741.7686 } }, "n_tiles": 73, "provenance": { "arm": "v5_real_harm", "ckpt": "outputs/sean/v5/sean_best.pt", "wkey": "G_ema", "input": "real_harm", "n": 73, "generated": "suite_v2 gen" }, "date": "2026-07-05", "sources": [ "reports/suite_v2/scores/v5_real_harm.json", "reports/phase0_baseline_table.md" ] }, "v5_real_raw": { "label": "SEAN v5 on REAL Sentinel-2 input (deployment condition, raw)", "kind": "internal", "status": "measured", "caveat": "real-input LPIPS/PSD are vs different-date GT \u2014 content-confounded; KID-HP/DISTS-vs-bank and hf are the meaningful columns (reports/phase0_baseline_table.md)", "plane": "2.5m", "metrics": { "lpips": 0.36323, "dists": 0.20147, "hf_frac": 0.00097, "psd_dist": 0.16499, "block_p16": 1.00465, "kid_hp_pooled_x1e3": 4630.7247, "kid_hp_pooled_std": 0.0017 }, "per_stratum_kid_hp": { "residential": { "n_crops": 123, "kid_hp_x1e3": 4918.4494 }, "industrial_port": { "n_crops": 93, "kid_hp_x1e3": 6188.9662 }, "ag_row": { "n_crops": 91, "kid_hp_x1e3": 5661.0813 }, "forest": { "n_crops": 19, "kid_hp_x1e3": 9061.479, "small_n_caveat": true }, "water_coast": { "n_crops": 39, "kid_hp_x1e3": 5287.5248 } }, "n_tiles": 73, "provenance": { "arm": "v5_real_raw", "ckpt": "outputs/sean/v5/sean_best.pt", "wkey": "G_ema", "input": "real", "n": 73, "generated": "suite_v2 gen" }, "date": "2026-07-05", "sources": [ "reports/suite_v2/scores/v5_real_raw.json", "reports/phase0_baseline_table.md" ], "opensr_correctness_triad": { "hallucination": 0.2837322416404883, "omission": 0.5059078757961591, "improvement": 0.210359879458944, "n_tiles": 24, "caveat": "opensr-test v1.3.3 correctness triad, 24 tiles, real-LR arms only; absolutes carry the HR-date confound \u2014 read within-group (reports/phase0_baseline_table.md)", "source": "reports/suite_v2/opensr_triad_realv5.json" } }, "ref_v5_refiner25": { "label": "2.5m HF-adversarial refiner on v5_best (checkpoint ladder)", "kind": "internal", "plane": "2.5m", "status": "measured \u2014 KID-HP regressed at every checkpoint in both arms; AWAITING USER VISUAL VERDICT (reports/README.md 2026-07-06)", "date": "2026-07-06", "kid_hp_pooled_x1e3_by_step": { "20000": 4608.26, "40000": 6415.7028, "60000": 5751.3494, "80000": 5609.6762, "100000": 5446.2412, "120000": 5555.1182 }, "sources": [ "reports/phase0_baseline_table.md", "reports/suite_v2/scores/ref_v5_020000.json", "reports/suite_v2/scores/ref_v5_040000.json", "reports/suite_v2/scores/ref_v5_060000.json", "reports/suite_v2/scores/ref_v5_080000.json", "reports/suite_v2/scores/ref_v5_100000.json", "reports/suite_v2/scores/ref_v5_120000.json", "reports/README.md", "reports/sean_part_progress_report.md" ] }, "ref_hpf_refiner25": { "label": "2.5m HF-adversarial refiner on hpf100 (checkpoint ladder)", "kind": "internal", "plane": "2.5m", "status": "measured \u2014 KID-HP regressed at every checkpoint in both arms; AWAITING USER VISUAL VERDICT (reports/README.md 2026-07-06)", "date": "2026-07-06", "kid_hp_pooled_x1e3_by_step": { "20000": 4541.2804, "40000": 5773.5609, "60000": 5304.0655, "80000": 5255.2021, "100000": 5141.0112, "120000": 5284.6926 }, "sources": [ "reports/phase0_baseline_table.md", "reports/suite_v2/scores/ref_hpf_020000.json", "reports/suite_v2/scores/ref_hpf_040000.json", "reports/suite_v2/scores/ref_hpf_060000.json", "reports/suite_v2/scores/ref_hpf_080000.json", "reports/suite_v2/scores/ref_hpf_100000.json", "reports/suite_v2/scores/ref_hpf_120000.json", "reports/README.md", "reports/sean_part_progress_report.md" ] }, "bicubic_lr_coreg_control": { "label": "bicubic of the real lr_coreg (triad sanity control)", "kind": "baseline_control", "status": "baseline_control", "plane": "2.5m", "date": "2026-07-05", "opensr_correctness_triad": { "hallucination": 0.13400953883926073, "omission": 0.7710708019634088, "improvement": 0.09491965966299176, "n_tiles": 24, "caveat": "opensr-test v1.3.3 correctness triad, 24 tiles, real-LR arms only; absolutes carry the HR-date confound \u2014 read within-group (reports/phase0_baseline_table.md)", "source": "reports/suite_v2/opensr_triad.json" }, "sources": [ "reports/suite_v2/opensr_triad.json" ] }, "cascade_alpha025": { "label": "Full cascade 10 m \u2192 0.5 m: SEAN v5 \u2192 PiD Combined LoRA @ \u0394W\u00d70.25", "kind": "internal", "plane": "cascade_10m_to_0.5m", "status": "ratified \u2014 USER visual verdict 2026-07-02: alpha=0.25, decode quality approved (configs/base.yaml stage_b.lora_scale)", "date": "2026-07-02", "metrics": { "lpips_arm3_sean_synth_input": { "value": 0.4313, "source": "reports/stageb_diag_2026-07-02.md" }, "lpips_bicubic_floor": { "value": 0.472, "source": "reports/stageb_diag_2026-07-02.md", "note": "first configuration to beat the bicubic floor end-to-end" }, "lpips_true_hr25_input": { "value": 0.2794, "source": "reports/stageb_diag_2026-07-02.md" }, "hf_ratio_vs_gt": { "value": 2.0, "source": "reports/stageb_diag_2026-07-02.md" }, "input_recon_corr": { "value": 0.941, "source": "reports/stageb_diag_2026-07-02.md" }, "hallucination_flags": { "value": "0/18", "source": "reports/stageb_diag_2026-07-02.md" } }, "checkpoints": { "stage_a": "outputs/sean/v5/sean_best.pt (G_ema, step 48700)", "stage_b": "outputs/octave_pid/combined/octave_lora.pt @ --lora_scale 0.25" }, "sources": [ "reports/stageb_diag_2026-07-02.md", "reports/README.md", "configs/base.yaml" ] } }, "user_2afc_selftest": { "what": "USER 2AFC self-test \u2014 time-limited native-res forced choice, real vs model output. accuracy 0.5 = indistinguishable (the project's actual success criterion, docs/plans/plane25_indistinguishability.md); accuracy 1.0 = always distinguishable.", "sessions": [ { "session": 1, "arm": "v5_best", "graded": "2026-07-06 09:20", "n_answered": 20, "accuracy": 0.9, "per_stratum": { "residential": 0.875, "industrial_port": 1.0, "ag_row": 0.875, "forest": 0.75, "water_coast": 1.0 }, "calibration_fpr": 0.0, "source": "reports/suite_v2/selftest_log.jsonl" } ] }, "loop_0p5m": { "what": "Autonomous Stage-B ratchet loop at the 0.5 m plane (branch loop/stageb-ratchet). KID/FDD here are computed on the loop's frozen 18-tile 0.5 m fixture vs real-DOF banks (loop/CALIBRATION.md) \u2014 a DIFFERENT scale and plane than suite_v2 KID-HP \u00d710\u00b3 at 2.5 m; the two are NOT comparable.", "champion_lineage": [ { "iter": 0, "champion": "alpha=0.25 sigma=0.10 (user-ratified bootstrap)", "kid": 2.25438 }, { "iter": 2, "champion": "alpha=0.15 sigma=0.10", "kid": 1.92796 }, { "iter": 5, "champion": "NAIP domain-adapt LoRA r64 @ step 34k (alpha=0.15 sigma=0.10)", "kid": 1.8001, "caveat": "hf_ratio 2.021 flagged (led to the USER-authorized hf_ratio\u22641.5 guard, commit 41b38ac); all champion panels PENDING HUMAN REVIEW \u2014 judge sign-off gates autonomous nights (reports/README.md 2026-07-04)" } ], "iterations": [ { "iter": "0", "utc": "2026-07-04T09:15:18+00:00", "git_commit": "63421f9", "regime": "infer", "change": "bootstrap champion (ratified a025 s010)", "kid": "2.25438", "kid_std": "0.0", "fdd": "530.1", "psd_full": "0.2302", "psd_top": "0.2981", "roundtrip": "0.9112", "b32": "1.073", "hf_ratio": "0.875", "ab_win": null, "judge_acc": null, "control_acc": null, "guards": "pass", "verdict": "accept", "notes": null }, { "iter": "1", "utc": "2026-07-04T09:23:28+00:00", "git_commit": "63421f9", "regime": "infer", "change": "calib i1: sigma 0.15", "kid": "2.15364", "kid_std": "0.0", "fdd": "520.5", "psd_full": "0.2112", "psd_top": "0.2739", "roundtrip": "0.8816", "b32": "1.075", "hf_ratio": "0.942", "ab_win": "0.5", "judge_acc": "0.812", "control_acc": "0.958", "guards": "pass", "verdict": "reject:judge", "notes": null }, { "iter": "2", "utc": "2026-07-04T09:30:02+00:00", "git_commit": "63421f9", "regime": "infer", "change": "calib i2: alpha 0.15", "kid": "1.92796", "kid_std": "0.0", "fdd": "511.2", "psd_full": "0.0535", "psd_top": "0.0431", "roundtrip": "0.9008", "b32": "1.063", "hf_ratio": "1.3", "ab_win": "0.625", "judge_acc": "0.5", "control_acc": "0.958", "guards": "pass", "verdict": "accept", "notes": null }, { "iter": "3", "utc": "2026-07-04T09:37:02+00:00", "git_commit": "ef8541e", "regime": "infer", "change": "calib i3: alpha 0.40", "kid": "2.66562", "kid_std": "0.0", "fdd": "563.0", "psd_full": "0.4944", "psd_top": "0.5921", "roundtrip": "0.9185", "b32": "1.106", "hf_ratio": "0.828", "ab_win": "0.125", "judge_acc": "1.0", "control_acc": "0.958", "guards": "pass", "verdict": "reject:realism", "notes": null }, { "iter": "4", "utc": "2026-07-04T13:50:21+00:00", "git_commit": "85ab8d0", "regime": "infer", "change": "t1 prime: NAIP r64 lq_degrade3 lpips0.2 (cache build + first capped arm)", "kid": "1.8001", "kid_std": "0.0", "fdd": "501.9", "psd_full": "0.1791", "psd_top": "0.2227", "roundtrip": "0.8892", "b32": "1.056", "hf_ratio": "2.021", "ab_win": "0.5", "judge_acc": "1.0", "control_acc": "0.875", "guards": "pass", "verdict": "reject:judge", "notes": null }, { "iter": "5", "utc": "2026-07-04T14:02:27+00:00", "git_commit": "2aff4fb", "regime": "infer", "change": "gate retry: t1 naip_r64@34k decoder (iter4 judge control degenerate)", "kid": "1.8001", "kid_std": "0.0", "fdd": "501.9", "psd_full": "0.1791", "psd_top": "0.2227", "roundtrip": "0.8892", "b32": "1.056", "hf_ratio": "2.021", "ab_win": "0.625", "judge_acc": "0.438", "control_acc": "1.0", "guards": "pass", "verdict": "accept", "notes": null }, { "iter": "6", "utc": "2026-07-04T15:49:08+00:00", "git_commit": "95e289c", "regime": "infer", "change": "i1: alpha 0.40 on naip_r64@34k champion", "kid": "2.31611", "kid_std": "0.0", "fdd": "546.0", "psd_full": "0.2518", "psd_top": "0.2984", "roundtrip": "0.9106", "b32": "1.085", "hf_ratio": "1.001", "ab_win": "0.438", "judge_acc": "0.875", "control_acc": "0.708", "guards": "pass", "verdict": "reject:realism", "notes": null }, { "iter": "7", "utc": "2026-07-04T15:56:41+00:00", "git_commit": "1ff7969", "regime": "infer", "change": "i1: alpha 0.25 on naip_r64@34k champion", "kid": "1.97454", "kid_std": "0.0", "fdd": "513.3", "psd_full": "0.0471", "psd_top": "0.0405", "roundtrip": "0.8995", "b32": "1.065", "hf_ratio": "1.385", "ab_win": "0.375", "judge_acc": "0.438", "control_acc": "0.958", "guards": "pass", "verdict": "reject:realism", "notes": null }, { "iter": "8", "utc": "2026-07-04T16:04:04+00:00", "git_commit": "6436d23", "regime": "infer", "change": "i1: alpha 0.10 on naip_r64@34k champion", "kid": "1.75141", "kid_std": "0.0", "fdd": "501.5", "psd_full": "0.2573", "psd_top": "0.3313", "roundtrip": "0.8832", "b32": "1.053", "hf_ratio": "2.451", "ab_win": "0.25", "judge_acc": "0.688", "control_acc": "0.958", "guards": "pass", "verdict": "reject:judge", "notes": null }, { "iter": "9", "utc": "2026-07-04T23:23:47+00:00", "git_commit": "03efe1e", "regime": "infer", "change": "t2: lpips_weight 0.0 (own out dir naip_r64_lpips0)", "kid": "1.887", "kid_std": "0.0", "fdd": "510.6", "psd_full": "0.1391", "psd_top": "0.1605", "roundtrip": "0.8886", "b32": "1.056", "hf_ratio": "1.819", "ab_win": null, "judge_acc": null, "control_acc": null, "guards": "VETO", "verdict": "reject:guard", "notes": null } ], "source": "loop/LOG.tsv" } }