zorya-classifier 0.1.0 → 0.2.0

package/docs/evidence/v02/region-seeded-dev-v2/evaluation.json

@@ -0,0 +1,232 @@
+{
+  "schema": "region-seeded-dev-v2-public-evidence",
+  "status": "development-admission-passed",
+  "releaseQualified": false,
+  "freeze": {
+    "file": "region-seeded-development-freeze-20261021-v2.json",
+    "sha256": "4f8846aa3ca82462d5fff83ac7dbab8925a090bc85b90d45a6eb9b3433f270dc",
+    "producerSourceSha256": "99beadec729003691c6861166b4738db2cc1ee92699040c98254468b24c998e9",
+    "requireClassifiedPresence": true,
+    "gates": {
+      "completeCases": 40,
+      "positives": 30,
+      "verifiedNegatives": 10,
+      "maximumFalseSeedLocations": 0,
+      "maximumFalseExtractions": 0,
+      "minimumPositiveSeedHits": 1,
+      "minimumMaskDerivedBoxMatches": 1,
+      "minimumExactCanvasMaskIoU08": 1
+    }
+  },
+  "inputs": [
+    {
+      "stratum": "photo",
+      "cases": 20,
+      "inputSha256": "6c9ed29d8586c7887ae0861fc853baf203cb43790df4959390da4b5cc9ac679a",
+      "truthSha256": "793e26845deaa2cb56748ef270b8cb02877218d680068573f02e5af92b26af6a"
+    },
+    {
+      "stratum": "controlled-ui",
+      "cases": 10,
+      "inputSha256": "6013a54bebba277f6457765cd8ef78e12e2dd5bab4453a752967233ebb429bfc",
+      "truthSha256": "c7944190e1cadab695f2688351df10a14132986040dafb72794bb1df4571abfe"
+    },
+    {
+      "stratum": "controlled-map",
+      "cases": 10,
+      "inputSha256": "b04ba6cac5b39c927281f61e4af7dcc7a5c5c2272a637aa9676ab3009fbd5d15",
+      "truthSha256": "874971d3e9344f7c54207064115257e5c802260d2a996a53f96bb6a68b2afad3"
+    }
+  ],
+  "candidate": {
+    "profileId": "native-luna-presence-first-region-seeded-sam-dev-v2",
+    "variant": "region-seeded-sam",
+    "runId": "v02-region-seeded-dev-2026-10-20T221925301Z-95745556",
+    "manifestRawSha256": "e5cb1bbee2a72a32b2a2d65cbca75bbcb0fb6309220399662f6e288dd07ff51e",
+    "summaryRawSha256": "20741658e039172934e0edd3b555e49d8e30367a68855107581bb1478af5e131",
+    "summaryCanonicalSha256InCompletionSeal": "d9d9a1e4e956fcae3b08088b47cd9373c530755da033d14ee06e4750eaa15ca1",
+    "completionRawSha256": "243812213d44cc348c706ba42a73f84a5adc11a37f9255867d472ce8b4caada5",
+    "variantSha256": "4ccbd0960d0984033dca1b49b84108abd5a81b6e55e02e9aef3e02018cead4f7",
+    "sourceBeforeAndAfterSha256": "99beadec729003691c6861166b4738db2cc1ee92699040c98254468b24c998e9",
+    "route": "openai/gpt-6-luna-decisions",
+    "nativeAssetManifestMethod": "mobile-sam-seed-point-v1",
+    "legacyRunnerRegionSeededAssetLabel": "mobile-sam-seed-region-v1",
+    "actualSegmentationResultMethod": "mobile-sam-seed-point-v1",
+    "nativeBackend": "MobileSAM CPU",
+    "nativeEncoderSha256": "580f5fb648ea1062c0aabc26217aed56921985f03f0cbbd852bba81d760cc749",
+    "nativeDecoderSha256": "8976b90a87ba50a6a72217a5ff994f7d25ce16f2229fcc1ed259e1294c622ffe",
+    "denominator": 40,
+    "classification": {
+      "correct": 35,
+      "denominator": 40,
+      "accuracy": 0.875,
+      "unanswered": 0
+    },
+    "strata": {
+      "photo": {
+        "cases": 20,
+        "positive": 16,
+        "verifiedNegative": 4,
+        "classificationCorrect": 17,
+        "extractionStatuses": {
+          "extracted": 10,
+          "failed": 6,
+          "unknown": 3,
+          "not-found": 1
+        },
+        "seedLocated": 10,
+        "positiveSeedBoxIntersections": 10,
+        "falseSeedLocations": 0,
+        "maskDerivedBoxMatches": 5,
+        "falseExtractions": 0
+      },
+      "controlledUi": {
+        "cases": 10,
+        "positive": 7,
+        "verifiedNegative": 3,
+        "classificationCorrect": 8,
+        "extractionStatuses": {
+          "extracted": 5,
+          "unknown": 4,
+          "not-found": 1
+        },
+        "seedLocated": 5,
+        "positiveSeedBoxIntersections": 1,
+        "falseSeedLocations": 0,
+        "maskDerivedBoxMatches": 0,
+        "falseExtractions": 0
+      },
+      "controlledMap": {
+        "cases": 10,
+        "positive": 7,
+        "verifiedNegative": 3,
+        "classificationCorrect": 10,
+        "extractionStatuses": {
+          "extracted": 7,
+          "not-found": 3
+        },
+        "seedLocated": 7,
+        "positiveSeedBoxIntersections": 7,
+        "falseSeedLocations": 0,
+        "maskDerivedBoxMatches": 0,
+        "falseExtractions": 0
+      }
+    },
+    "seedRegionCoarseProxy": {
+      "located": 22,
+      "locatedOnPositiveCases": 22,
+      "locatedOnVerifiedNegatives": 0,
+      "positiveBoxIntersections": 18,
+      "positiveDenominator": 30,
+      "falseSeedLocations": 0,
+      "falseExtractions": 0,
+      "meaning": "Any positive-area seed-cell/reference-box overlap; a coarse search proxy, never whole-object localization."
+    },
+    "maskDerivedWholeObjectBoxes": {
+      "positiveMatches": 5,
+      "positiveDenominator": 30,
+      "criterion": "IoU >= 0.5 and reference-box coverage >= 0.9",
+      "falseLocations": 0,
+      "negativeDenominator": 10,
+      "unknown": 7,
+      "failed": 6
+    },
+    "independentExactCanvasMasks": {
+      "references": 7,
+      "scored": 7,
+      "iouAtLeast0_8": 1,
+      "unavailable": 0,
+      "unannotatedCases": 35,
+      "perReferenceIoU": [
+        0.7589947292, 0, 0.0002087876, 0.9433618606, 0.5063549934, 0.6049145284, 0.3675930123
+      ]
+    },
+    "physical": {
+      "calls": 134,
+      "knownCostUsd": 0.0608289,
+      "unknownCostCalls": 6,
+      "accountedExposureUsd": 0.3608289,
+      "httpStatusCounts": {
+        "200": 128,
+        "502_PROVIDER": 6
+      },
+      "inFlightAtSeal": 0
+    },
+    "independentEvaluation": {
+      "evaluatorCommit": "afb2bce3befb8002050d3fca5660cdc6464f5484",
+      "evaluatorSha256": "208e94d58622f59f3299e79ff3b0ce0e468c118f468060623e0e8b93380c1579",
+      "outputSha256": "64ab30c6085b6b0e404416f18bafd9246358a8de0ccb142350df04f17efd7507"
+    }
+  },
+  "noImageControl": {
+    "profileId": "native-luna-presence-first-region-no-image-control-dev-v2",
+    "variant": "no-image-control",
+    "runId": "v02-region-seeded-dev-2026-10-20T222226064Z-b84bf16f",
+    "manifestRawSha256": "9e0ab7e6662bafbd21f16855398bf254906e544973181adf6def2f6c3c4848bb",
+    "summaryRawSha256": "2e34686c6ba5e85666926c6169ebc22adb2d353e37550ffc35418dc57c18f6f1",
+    "summaryCanonicalSha256InCompletionSeal": "aeb67b5b685f3465616fabadf9479d70fe0d1aa559831605f69a0df9287ed2bc",
+    "completionRawSha256": "2a974aa6d28588195522b8d0eadf482534a9485bf637ae9d49fabc678482d778",
+    "variantSha256": "c2f46326480853101fef59374cfd9363b45d44666f8c455cbe36ab9d206aa953",
+    "denominator": 40,
+    "classification": {
+      "correct": 10,
+      "answered": 40,
+      "unanswered": 0
+    },
+    "imageFetchAudit": {
+      "checked": 40,
+      "imageParts": 0,
+      "verifiedImageFree": true
+    },
+    "seedLocations": 0,
+    "physical": {
+      "calls": 40,
+      "knownCostUsd": 0.0008353,
+      "unknownCostCalls": 0,
+      "accountedExposureUsd": 0.0008353,
+      "httpStatusCounts": {
+        "200": 40
+      },
+      "inFlightAtSeal": 0
+    },
+    "independentEvaluation": {
+      "evaluatorCommit": "afb2bce3befb8002050d3fca5660cdc6464f5484",
+      "evaluatorSha256": "208e94d58622f59f3299e79ff3b0ce0e468c118f468060623e0e8b93380c1579",
+      "outputSha256": "6303a8a818401235e6edba5be2c88d7fbe9f5921080cf873bb677775207a4fda"
+    }
+  },
+  "pairedPhotoSeedDiagnostic": {
+    "candidateProfile": "native-luna-presence-first-region-seeded-sam-dev-v2",
+    "controlProfile": "native-luna-presence-first-region-no-image-control-dev-v2",
+    "population": "16 positive photo cases",
+    "metric": "coarse seed-region/reference-box intersection; not whole-object localization",
+    "bootstrap": {
+      "iterations": 10000,
+      "seed": "0x5a4f5259",
+      "confidence": 0.95,
+      "difference": 0.625,
+      "lower": 0.375,
+      "upper": 0.875
+    },
+    "exactReferenceMaskForegroundSubset": {
+      "pairs": 5,
+      "difference": 0.8,
+      "lower": 0.4,
+      "upper": 1
+    },
+    "outputSha256": "ff3f7e7b3b6b343594c97ff6146071daf1100134461d4034499460c4cb4121ea"
+  },
+  "devAdmission": {
+    "passed": true,
+    "conditions": {
+      "complete40CaseRun": true,
+      "zeroFalseSeedLocations": true,
+      "zeroFalseExtractions": true,
+      "positiveSeedHitPresent": true,
+      "maskDerivedBoxMatchPresent": true,
+      "exactCanvasMaskIoU08Present": true
+    },
+    "releaseOrHeldoutQualification": false,
+    "note": "The DEV admission result permits the frozen recipe to be considered for a separate blind run; it does not establish heldout performance or release qualification."
+  }
+}

package/docs/evidence/v02/region-seeded-dev-v3-candidate-only/README.md

@@ -0,0 +1,9 @@
+# Checked-region DEV v3 candidate-only evaluation
+
+This report is an independent evaluation of the completed 40-case DEV candidate run against the frozen DEV truth. It is candidate-only evidence, not a paired gate verdict or release qualification. The control run is retained but was not scored against the original pair because its run identity does not match the pre-dispatch variant digest; see `control-identity-failure.json`.
+
+The evaluation preserves the full 40-case denominator: 30 positive and 10 verified-negative cases across the 20-photo, 10-controlled-UI, and 10-controlled-map strata. Classification, seed-cell overlap, mask-derived whole-object boxes, visible-box checks, and reference-mask IoU are reported separately. A seed-cell intersection is only a coarse search proxy and never counts as whole-object localization. The UI and map cases are controlled captures from one app/game family each, not natural user captures or evidence of broad application/game diversity.
+
+The candidate used the frozen OpenAI GPT-6 Luna route with no fallback, the existing presence-first gate, full-source MobileSAM seed-region segmentation, and up to two closed visible-box containment checks. The report retains unknown, not-found, and failed outcomes in every denominator. The 7 available reference-mask instances are scored independently: 4 were comparable on the exact canvas, 1 met IoU 0.8, and 3 were unavailable. This does not imply universal cutout accuracy.
+
+The report records hashes for the freeze, producer source closure, all input and truth files, candidate raw manifest/summary/seal, and the independent evaluator output. It contains aggregate metrics only; it does not publish case-level outcomes, image paths, image pixels, masks, or provider credentials.

package/docs/evidence/v02/region-seeded-dev-v3-candidate-only/control-identity-failure.json

@@ -0,0 +1,22 @@
+{
+  "schema": "region-seeded-dev-v3-control-identity-failure-v1",
+  "status": "preserved and unscored against original frozen pair",
+  "originalFreezeSha256": "d980af0ee5c3ae38f2e397db8e195686663d2b2862afad53c9e6f62b68b113a2",
+  "frozenControlVariantSha256": "5a64ebc3f0261dd7ece37d8428ccae77c9e0c47019216d3cf2876f5b30983511",
+  "rawControl": {
+    "profileId": "native-luna-presence-first-region-no-image-control-dev-v3",
+    "variant": "no-image-control",
+    "manifestSha256": "a0e756cf750ec9d8b8d46aa42d64d7346492f28723d6c729a30077fac58010d9",
+    "summarySha256": "5ea51a5795f645ad39324281bc5e9e2fec6bef501549dce31f43408451bd9733",
+    "completionSealSha256": "6e7a3786da5548d98d6a69bf5772a86e4741333aa0d20883a1c3eb98f66eec5c",
+    "actualVariantSha256": "b2d85c509ff454ee62299d46a9f7dc7cb3009ca0aaf0ef015344c8a6dc45ed35",
+    "sourceUnchanged": true,
+    "plannedAndCompletedCases": 40
+  },
+  "identityFormula": "hashCanonical({profileId, phase, variant, requireClassifiedPresence:true, checkedRegion:true, route, policy, promptPolicy, manifests, taskHashes, source, nativeAssets:null})",
+  "diagnosis": "The frozen control digest is reproduced only if the candidate run's promptPolicy is substituted in place of the control promptPolicy. The control producer output has the correct no-image extraction policy and is internally consistent: its actual promptPolicy says it runs only the original region search without segmentation or visible-box checks, and reconstructing the identity from that raw manifest yields actualVariantSha256. The original frozen digest therefore does not identify the control recipe that was dispatched.",
+  "pairedEvaluation": "not produced; evaluator rejected the control variant digest mismatch before scoring control outcomes or computing a paired statistic",
+  "preservedFailures": true,
+  "rawArtifactsModified": false,
+  "candidateRerun": false
+}

package/docs/evidence/v02/region-seeded-dev-v3-candidate-only/evaluation.json

@@ -0,0 +1,160 @@
+{
+  "schema": "region-seeded-dev-v3-candidate-only-evidence-v1",
+  "status": "candidate evaluated; original candidate/control pair incomplete due control identity mismatch",
+  "qualification": "no paired gate verdict; not release qualification",
+  "freeze": {
+    "sourceCommit": "d4b84b9fa025e5a5238968f99fb685633c837c7d",
+    "producerSourceSha256": "be43090b13951e9a6ffa65daffb9413a6a539d8dae335925dc885733a467fb20",
+    "profileId": "native-luna-presence-first-checked-region-sam-dev-v3",
+    "variantSha256": "f4ed437c17160ab1a694961e9b38b63ad76a290a21dbe82aa79d51e169c19697",
+    "route": "openai/gpt-6-luna-decisions",
+    "phaseBudget": {
+      "maxPhysicalCalls": 250,
+      "maxCostUsd": 5,
+      "unknownCostReserveUsd": 0.05
+    },
+    "policy": {
+      "timeoutMs": 120000,
+      "maxInFlightCases": 2,
+      "nativeMaxActive": 1,
+      "nativeMaxWaiting": 2,
+      "nativeThreads": 2,
+      "maxPhysicalCallsPerCase": 6,
+      "maxLogicalCallsPerCase": 6,
+      "maxBoxCheckCalls": 2,
+      "retries": 0,
+      "fallback": false,
+      "latencyRankingEligible": false,
+      "requireClassifiedPresence": true
+    }
+  },
+  "inputs": [
+    {
+      "cases": 20,
+      "manifestSha256": "6c9ed29d8586c7887ae0861fc853baf203cb43790df4959390da4b5cc9ac679a",
+      "truthSha256": "793e26845deaa2cb56748ef270b8cb02877218d680068573f02e5af92b26af6a"
+    },
+    {
+      "cases": 10,
+      "manifestSha256": "6013a54bebba277f6457765cd8ef78e12e2dd5bab4453a752967233ebb429bfc",
+      "truthSha256": "c7944190e1cadab695f2688351df10a14132986040dafb72794bb1df4571abfe"
+    },
+    {
+      "cases": 10,
+      "manifestSha256": "b04ba6cac5b39c927281f61e4af7dcc7a5c5c2272a637aa9676ab3009fbd5d15",
+      "truthSha256": "874971d3e9344f7c54207064115257e5c802260d2a996a53f96bb6a68b2afad3"
+    }
+  ],
+  "candidateRun": {
+    "runId": "v02-region-seeded-dev-2026-10-20T235301825Z-10b11642",
+    "manifestSha256": "e18ca240cb35d7f3081d5ab8bf19f93b451e638e5e7887975cce4715c243a819",
+    "summarySha256": "fea18ed9b5dce920495bce1d9cb9e190675ec5630c07b0ae07f73623d731be39",
+    "completionSealSha256": "c65fc9001acdbdedd2e823764012b0ad6b6dbfc9ef820f6aceb073f95fa0f25b",
+    "independentEvaluationSha256": "2ac4d19caaf58ab7ce9c666b7b4a166ec3cc7008e48f869c8849a9891eae0a89",
+    "sourceUnchanged": true,
+    "complete": true,
+    "denominator": 40
+  },
+  "aggregate": {
+    "classification": {
+      "answered": 40,
+      "correct": 35,
+      "accuracy": 0.875
+    },
+    "wholeObject": {
+      "denominator": 40,
+      "positives": 30,
+      "negatives": 10,
+      "matchedPositiveBoxes": 6,
+      "falseLocations": 0,
+      "unknown": 13,
+      "failed": 6,
+      "notFound": 5
+    },
+    "seedSearchProxy": {
+      "located": 22,
+      "positiveBoxIntersections": 18,
+      "falseSeedLocations": 0,
+      "falsePositiveBoxIntersections": 0,
+      "interpretation": "coarse search proxy only; not whole-object success"
+    },
+    "boxChecks": {
+      "caseDenominator": 40,
+      "casesWithAttempt": 22,
+      "casesWithoutAttempt": 18,
+      "successfulAttempts": 28,
+      "failedAttempts": 0,
+      "verdicts": {
+        "complete": 16,
+        "needsExpansion": 12,
+        "unknown": 0
+      },
+      "expansionOutcomes": {
+        "notNeeded": 13,
+        "noEnclosingMask": 3,
+        "expanded": 6,
+        "unsupported": 0,
+        "rejectedMask": 0
+      }
+    },
+    "maskReferences": {
+      "instances": 7,
+      "scoredExactCanvas": 4,
+      "iouAtLeast0_8": 1,
+      "unavailable": 3,
+      "casesWithoutReferenceMasks": 35
+    },
+    "physicalCalls": {
+      "count": 162,
+      "knownCostUsd": 0.0645113,
+      "unknownCostCalls": 6,
+      "accountedExposureUsd": 0.3645113,
+      "inFlightAtSeal": 0
+    },
+    "strata": {
+      "photo": {
+        "cases": 20,
+        "positive": 16,
+        "negative": 4,
+        "classificationCorrect": 17,
+        "extractionStatuses": {
+          "extracted": 6,
+          "failed": 6,
+          "unknown": 7,
+          "notFound": 1
+        },
+        "seedLocations": 10,
+        "seedPositiveIntersections": 10,
+        "maskDerivedBoxesMatched": 6
+      },
+      "ui": {
+        "cases": 10,
+        "positive": 7,
+        "negative": 3,
+        "classificationCorrect": 8,
+        "extractionStatuses": {
+          "extracted": 5,
+          "unknown": 4,
+          "notFound": 1
+        },
+        "seedLocations": 5,
+        "seedPositiveIntersections": 1,
+        "maskDerivedBoxesMatched": 0
+      },
+      "map": {
+        "cases": 10,
+        "positive": 7,
+        "negative": 3,
+        "classificationCorrect": 10,
+        "extractionStatuses": {
+          "extracted": 5,
+          "unknown": 2,
+          "notFound": 3
+        },
+        "seedLocations": 7,
+        "seedPositiveIntersections": 7,
+        "maskDerivedBoxesMatched": 0
+      }
+    }
+  }
+}

package/docs/evidence/v02/region-seeded-dev-v3-control-r2/README.md

@@ -0,0 +1,9 @@
+# Checked-region DEV v3 control R2 and paired seed comparison
+
+This report records the strict evaluation of the prospective no-image control R2 and its predeclared seed-search comparison with the unchanged, independently evaluated candidate. The original control run is preserved as an identity failure and is not scored or used in this pair; see the candidate-only evidence directory.
+
+Both scored runs use the same frozen 40-case DEV corpus and truth files. The candidate remains a 40-case model-assisted segmentation run; the control is a no-image classification and region-search control, with no segmentation. Candidate whole-object boxes and reference masks are reported separately from the control’s seed-search results. The paired comparison uses only the same-semantic seed-search measures. It never compares a candidate mask box with a control seed cell, and a seed-cell intersection is not a whole-object success.
+
+The candidate produced 6 matched positive whole-object boxes among 30 positives, with no false locations among 10 verified negatives. Its exact-canvas mask results were 1 pass among 7 reference instances (4 scored, 3 unavailable). Those remain low object and mask scores despite the positive paired seed-search comparison. This report is DEV evidence only; it is not a heldout result or release qualification.
+
+The report contains aggregate counts and immutable artifact digests only. It omits per-case outcomes, images, masks, local filesystem paths, and provider credentials.

package/docs/evidence/v02/region-seeded-dev-v3-control-r2/attempt-accounting.json

@@ -0,0 +1,55 @@
+{
+  "schema": "region-seeded-dev-v3-all-attempt-accounting-v1",
+  "scope": "All three physically dispatched DEV runs, including the unscored original control identity failure. This is not an account invoice.",
+  "attempts": [
+    {
+      "attempt": "candidate",
+      "summarySha256": "fea18ed9b5dce920495bce1d9cb9e190675ec5630c07b0ae07f73623d731be39",
+      "cases": 40,
+      "usedInPairedEvaluation": true,
+      "admittedPhysicalCalls": 162,
+      "knownCostUsd": 0.06451130000000004,
+      "unknownCostCalls": 6,
+      "inFlightCalls": 0,
+      "accountedExposureUsd": 0.3645113000000001,
+      "maxPhysicalCalls": 250,
+      "maxCostUsd": 5
+    },
+    {
+      "attempt": "original-control",
+      "summarySha256": "5ea51a5795f645ad39324281bc5e9e2fec6bef501549dce31f43408451bd9733",
+      "cases": 40,
+      "usedInPairedEvaluation": false,
+      "admittedPhysicalCalls": 40,
+      "knownCostUsd": 0.0008353000000000002,
+      "unknownCostCalls": 0,
+      "inFlightCalls": 0,
+      "accountedExposureUsd": 0.0008353000000000002,
+      "maxPhysicalCalls": 250,
+      "maxCostUsd": 5
+    },
+    {
+      "attempt": "control-r2",
+      "summarySha256": "d5bd5a5c15b39b4432e815bdf5271f8ee2a15355d76519918576586a7c3cc98e",
+      "cases": 40,
+      "usedInPairedEvaluation": true,
+      "admittedPhysicalCalls": 40,
+      "knownCostUsd": 0.0008353000000000002,
+      "unknownCostCalls": 0,
+      "inFlightCalls": 0,
+      "accountedExposureUsd": 0.0008353000000000002,
+      "maxPhysicalCalls": 250,
+      "maxCostUsd": 5
+    }
+  ],
+  "total": {
+    "knownCostUsd": 0.06618190000000004,
+    "accountedExposureUsd": 0.3661819000000001,
+    "admittedPhysicalCalls": 242,
+    "unknownCostCalls": 6,
+    "inFlightCalls": 0
+  },
+  "unknownCostReserveUsd": 0.05,
+  "allFailuresAndChargesPreserved": true,
+  "preparationOnlyHeldoutCalls": 0
+}

package/docs/evidence/v02/region-seeded-dev-v3-control-r2/evaluation.json

@@ -0,0 +1,136 @@
+{
+  "schema": "region-seeded-dev-v3-r2-pair-evidence-v1",
+  "phase": "dev",
+  "status": "candidate and prospective control R2 strictly evaluated; original control attempt preserved unscored",
+  "qualification": "DEV evidence only; not heldout or release qualification",
+  "freeze": {
+    "candidateProfileId": "native-luna-presence-first-checked-region-sam-dev-v3",
+    "controlProfileId": "native-luna-presence-first-region-no-image-control-dev-v3",
+    "producerSourceSha256": "be43090b13951e9a6ffa65daffb9413a6a539d8dae335925dc885733a467fb20",
+    "controlR2FreezeSha256": "99251dc8bd69734d64625040ecd4595e81107eeeffb2134d31db4d8c71c4a3df",
+    "candidateVariantSha256": "f4ed437c17160ab1a694961e9b38b63ad76a290a21dbe82aa79d51e169c19697",
+    "controlR2VariantSha256": "b2d85c509ff454ee62299d46a9f7dc7cb3009ca0aaf0ef015344c8a6dc45ed35",
+    "inputAndTruthPins": [
+      {
+        "cases": 20,
+        "inputSha256": "6c9ed29d8586c7887ae0861fc853baf203cb43790df4959390da4b5cc9ac679a",
+        "truthSha256": "793e26845deaa2cb56748ef270b8cb02877218d680068573f02e5af92b26af6a"
+      },
+      {
+        "cases": 10,
+        "inputSha256": "6013a54bebba277f6457765cd8ef78e12e2dd5bab4453a752967233ebb429bfc",
+        "truthSha256": "c7944190e1cadab695f2688351df10a14132986040dafb72794bb1df4571abfe"
+      },
+      {
+        "cases": 10,
+        "inputSha256": "b04ba6cac5b39c927281f61e4af7dcc7a5c5c2272a637aa9676ab3009fbd5d15",
+        "truthSha256": "874971d3e9344f7c54207064115257e5c802260d2a996a53f96bb6a68b2afad3"
+      }
+    ]
+  },
+  "candidate": {
+    "runId": "v02-region-seeded-dev-2026-10-20T235301825Z-10b11642",
+    "manifestSha256": "e18ca240cb35d7f3081d5ab8bf19f93b451e638e5e7887975cce4715c243a819",
+    "summarySha256": "fea18ed9b5dce920495bce1d9cb9e190675ec5630c07b0ae07f73623d731be39",
+    "completionSealSha256": "c65fc9001acdbdedd2e823764012b0ad6b6dbfc9ef820f6aceb073f95fa0f25b",
+    "independentEvaluationSha256": "2ac4d19caaf58ab7ce9c666b7b4a166ec3cc7008e48f869c8849a9891eae0a89",
+    "classification": {
+      "correct": 35,
+      "denominator": 40
+    },
+    "wholeObject": {
+      "positiveMatches": 6,
+      "positiveDenominator": 30,
+      "falseLocations": 0,
+      "negativeDenominator": 10,
+      "unknown": 13,
+      "failed": 6,
+      "notFound": 5
+    },
+    "seedSearch": {
+      "located": 22,
+      "positiveBoxIntersections": 18,
+      "falseSeedLocations": 0,
+      "falseExtractions": 0
+    },
+    "boxChecks": {
+      "successfulAttempts": 28,
+      "verdicts": {
+        "complete": 16,
+        "needsExpansion": 12,
+        "unknown": 0
+      },
+      "expansionOutcomes": {
+        "notNeeded": 13,
+        "noEnclosingMask": 3,
+        "expanded": 6
+      }
+    },
+    "masks": {
+      "referenceInstances": 7,
+      "exactCanvasScored": 4,
+      "iouAtLeast0_8": 1,
+      "unavailable": 3
+    },
+    "physical": {
+      "calls": 162,
+      "knownCostUsd": 0.0645113,
+      "unknownCostCalls": 6,
+      "accountedExposureUsd": 0.3645113
+    }
+  },
+  "controlR2": {
+    "runId": "v02-region-seeded-dev-2026-10-21T000450468Z-35df9878",
+    "manifestSha256": "e4053461190962008d47587fc41b7d4e98afbec8da40d085d2e7808b199332c5",
+    "summarySha256": "d5bd5a5c15b39b4432e815bdf5271f8ee2a15355d76519918576586a7c3cc98e",
+    "completionSealSha256": "12b6d4eb0f49e45310c7f497ec066efb51a0cf959e1634a7398c89ca66f74e06",
+    "independentEvaluationSha256": "132184827f6dcd132b269194453977d72fa6345648422299d40f928fe269f394",
+    "sourceUnchanged": true,
+    "complete": true,
+    "denominator": 40,
+    "classification": {
+      "correct": 10,
+      "denominator": 40
+    },
+    "extractionStatuses": {
+      "notFound": 40
+    },
+    "seedLocations": 0,
+    "maskReferencesUnavailableWithoutPrediction": 7,
+    "physical": {
+      "calls": 40,
+      "knownCostUsd": 0.0008353,
+      "unknownCostCalls": 0,
+      "accountedExposureUsd": 0.0008353
+    }
+  },
+  "pairedPhotoSeedSearch": {
+    "positivePhotoPairs": 16,
+    "method": "paired-percentile-nearest-rank-bootstrap",
+    "iterations": 10000,
+    "seed": "0x5a4f5259",
+    "coarseBoxIntersectionSeedProxy": {
+      "candidateMinusControl": 0.625,
+      "lower95": 0.375,
+      "upper95": 0.875,
+      "passesPositiveDifference": true
+    },
+    "exactReferenceMaskForegroundSeedHit": {
+      "availablePairs": 5,
+      "candidateMinusControl": 0.8,
+      "lower95": 0.4,
+      "upper95": 1,
+      "passesPositiveDifference": true
+    },
+    "interpretation": "These paired comparisons measure seed search only. Seed-cell overlap is a coarse proxy; the exact-mask result applies only to the five photo positives with comparable reference masks. Neither is whole-object localization or cutout accuracy."
+  },
+  "originalControlIdentityFailure": {
+    "manifestSha256": "a0e756cf750ec9d8b8d46aa42d64d7346492f28723d6c729a30077fac58010d9",
+    "summarySha256": "5ea51a5795f645ad39324281bc5e9e2fec6bef501549dce31f43408451bd9733",
+    "completionSealSha256": "6e7a3786da5548d98d6a69bf5772a86e4741333aa0d20883a1c3eb98f66eec5c",
+    "expectedVariantSha256": "5a64ebc3f0261dd7ece37d8428ccae77c9e0c47019216d3cf2876f5b30983511",
+    "observedVariantSha256": "b2d85c509ff454ee62299d46a9f7dc7cb3009ca0aaf0ef015344c8a6dc45ed35",
+    "scored": false,
+    "preservedUnchanged": true
+  }
+}