zorya-classifier 0.1.0 → 0.2.1

package/docs/evidence/v02/categorical-heldout-v4/report.json

@@ -0,0 +1,258 @@
+{
+  "schema": "categorical-heldout-v4-independent-failure-v1",
+  "status": "failed-quality-qualification",
+  "releaseQualified": false,
+  "measuredAtUtc": "2026-10-20",
+  "producerSourceCommit": "c107ddfc8a129070e4bf7232f2c713b93fe650d4",
+  "runtimeSourceSnapshot": "0c365cb13173e290365a5aa2a1aec122c42b0b4d",
+  "freezeSha256": "97aa1173d0569c02522789a00ffd469e1a49792e564c91b104ba719ff6c176c1",
+  "methodSelectionSha256": "fabb05c2c472c30a8cb20a5f9b6d697b932d94aedc43976359746463ef8708b9",
+  "candidate": {
+    "runId": "v02-categorical-heldout-native-luna-categorical-presence-sam-heldout-target-plus-neutral-v1-2026-10-20T183257519Z-d6cfb60e",
+    "profileId": "native-luna-categorical-presence-sam-heldout-target-plus-neutral-v1",
+    "cases": 80,
+    "positives": 60,
+    "negatives": 20,
+    "runManifestSha256": "8ba9597bea273c93a8844d875ce95dd0d7ba9c74d35aa8949680acaed32b9503",
+    "rawSummarySha256": "1fb307aef68c93231b3d7fb10b415b4e4aa7d85b5a37ca3fa941bd0beeac28e4",
+    "producerSourceSha256": "0a93a965faf8f6ffbae630d03d00d0e68e2fdc69ac31e3d2ba171253e5515ac4",
+    "classification": {
+      "correct": 69,
+      "truePositives": 59,
+      "falsePositives": 10,
+      "falseNegatives": 1,
+      "trueNegatives": 10
+    },
+    "wholeObject": {
+      "matched": 1,
+      "positiveDenominator": 60,
+      "falseLocations": 0,
+      "negativeDenominator": 20,
+      "located": 4,
+      "unknown": 61,
+      "notFound": 8,
+      "failed": 7
+    },
+    "strata": [
+      {
+        "group": "photo",
+        "positiveMatches": 1,
+        "positiveCases": 30,
+        "falseLocations": 0,
+        "negativeCases": 10
+      },
+      {
+        "group": "ui",
+        "positiveMatches": 0,
+        "positiveCases": 15,
+        "falseLocations": 0,
+        "negativeCases": 5
+      },
+      {
+        "group": "map",
+        "positiveMatches": 0,
+        "positiveCases": 15,
+        "falseLocations": 0,
+        "negativeCases": 5
+      }
+    ],
+    "budget": {
+      "admittedPhysicalCalls": 356,
+      "knownCostUsd": 0.1608768999999998,
+      "unknownCostCalls": 7,
+      "inFlightCalls": 0,
+      "accountedExposureUsd": 0.5108768999999999,
+      "maxPhysicalCalls": 600,
+      "maxCostUsd": 10
+    },
+    "http": {
+      "physicalCalls": 356,
+      "status200": 349,
+      "status502": 7,
+      "negativeCase502": 7,
+      "retries": 0,
+      "fallback": false
+    },
+    "sourceUnchanged": true
+  },
+  "noImageControl": {
+    "runId": "v02-deepseek-no-image-heldout-photo-2026-10-20T184450822Z-eb3ceccc",
+    "cases": 40,
+    "positiveMatches": 0,
+    "positiveCases": 30,
+    "classificationCorrect": 10,
+    "runManifestSha256": "da21c46353240df05c35a04b63bc18224f7489985ca0709c9e17d8d8a1c9b553",
+    "rawSummarySha256": "1180d62c4a43e1cec45fadb6b33bb37e5f4b311dc28fc78ec08e7cbdd890915d",
+    "producerSourceSha256": "69c85bacace07e09356186d1a72252402bf10c28e114697f35a121be97887619",
+    "budget": {
+      "admittedPhysicalCalls": 100,
+      "knownCostUsd": 0.05400330000000001,
+      "unknownCostCalls": 0,
+      "inFlightCalls": 0,
+      "accountedExposureUsd": 0.05400330000000001,
+      "maxPhysicalCalls": 300,
+      "maxCostUsd": 5
+    },
+    "http": {
+      "physicalCalls": 100,
+      "status200": 100,
+      "retries": 0,
+      "fallback": false
+    },
+    "actualFetchAudit": {
+      "fetchesChecked": 100,
+      "fetchAttempts": 100,
+      "zeroImagePartsConfirmed": 100,
+      "zeroImageDerivedStateConfirmed": 100,
+      "imagePartViolations": 0,
+      "unverifiableBodies": 0,
+      "complete": true
+    }
+  },
+  "pairedPhotoQualification": {
+    "methodSelectionSha256": "fabb05c2c472c30a8cb20a5f9b6d697b932d94aedc43976359746463ef8708b9",
+    "photoCases": 40,
+    "positivePairs": 30,
+    "allGroupsNegativeCases": 20,
+    "candidateFalseLocationsAllGroups": [],
+    "maximumAllowedFalseLocationsAllGroups": 1,
+    "allGroupsNegativeGatePasses": true,
+    "bootstrap": {
+      "pairs": 30,
+      "iterations": 10000,
+      "seed": "0x5a4f5259",
+      "method": "paired-percentile-nearest-rank",
+      "difference": 0.03333333333333333,
+      "confidence": 0.95,
+      "lower": 0,
+      "upper": 0.1,
+      "passes": false
+    },
+    "denominator": "all 30 positive photo pairs; transport failures and unknowns count as failures"
+  },
+  "positiveFunnel": {
+    "cases": 60,
+    "initialPresenceAbsent": 1,
+    "finalSeedAbsentWithCellContradiction": 7,
+    "finalSeedAbsentWithNoCell": 6,
+    "finalSeedPresentWithNoCell": 9,
+    "reachedNativeProposals": 37,
+    "noMaskSelected": 15,
+    "selectedMaskCompletenessRejected": 18,
+    "locatedBoxFailed": 3,
+    "locatedBoxPassed": 1,
+    "reconciliation": "1+7+6+9+15+18+3+1 = 60; transport/parse/timeouts caused none of these positive outcomes",
+    "use": "diagnostic only; not a new method selection or revised qualification"
+  },
+  "maskCoverage": {
+    "initialEvaluatorReferenceInstances": 0,
+    "actualOfficialReferenceInstances": 32,
+    "actualOfficialAnnotatedCases": 21,
+    "exactSourceCanvasInstances": 2,
+    "differentFullFrameCanvasInstances": 30,
+    "strictJoinStatus": "rejected-different-source-and-annotation-dimensions",
+    "qualifiedMaskAccuracy": false,
+    "explanation": "Canonical photo truth and official mask index are separate. Initial evaluation did not join references. Strict join rejects 30 differently sized full-frame masks. Raw outputs stay unchanged; any normalized-canvas calculation is separate exploratory evidence and cannot change whole-box or paired results.",
+    "exploratoryEvaluation": "mask-normalized-exploratory.json",
+    "exploratoryScoredReferenceInstances": 7,
+    "exploratoryUnavailableReferenceInstances": 25,
+    "exploratoryPassingReferenceInstances": 3
+  },
+  "costAccounting": {
+    "physicalCalls": 456,
+    "knownCostUsd": 0.2148801999999998,
+    "unknownCostCalls": 7,
+    "unknownCostReserveUsdPerCall": 0.05,
+    "accountedExposureUsd": 0.5648801999999998,
+    "knownCostIsInvoice": false
+  },
+  "latency": {
+    "rankingEligible": false,
+    "method": "nearest rank, ceil(p*n), actual completed operation/physical fetch duration; shared host and native stress are unranked",
+    "candidate": {
+      "physicalFetch": {
+        "observations": 356,
+        "p50Ms": 1748.815476000018,
+        "p95Ms": 10264.485994999995
+      },
+      "classification": {
+        "observations": 80,
+        "p50Ms": 980.1109229999965,
+        "p95Ms": 4537.356227999997
+      },
+      "localization": {
+        "observations": 80,
+        "p50Ms": 6745.572493999964,
+        "p95Ms": 41874.570619999984
+      }
+    },
+    "noImageControl": {
+      "physicalFetch": {
+        "observations": 100,
+        "p50Ms": 517.789005000006,
+        "p95Ms": 1536.4638100000002
+      },
+      "classification": {
+        "observations": 40,
+        "p50Ms": 605.3571630000006,
+        "p95Ms": 1721.7188230000029
+      },
+      "localization": {
+        "observations": 40,
+        "p50Ms": 926.6581420000002,
+        "p95Ms": 2654.099829999999
+      }
+    }
+  },
+  "limitations": [
+    "Source selection was frozen before new pixels/per-case truth were opened; this 80-case corpus is now exposed regression data.",
+    "UI is one controlled Mousepad family; map is one controlled Minetest minimap family, not arbitrary real games.",
+    "Candidate and no-image control use different model families: this compares complete recipes, not a causal image-only ablation.",
+    "Seven physical 502 receipts have unknown usage cost; reservations remain, not zero charges.",
+    "All mandatory release gates are retained. A new development method and a fresh source-group-disjoint blind cohort are required."
+  ],
+  "artifacts": [
+    {
+      "path": "candidate-evaluation.json",
+      "sha256": "f5ac16c65b63115b22fe753a69a1b340a108e521df49b37aa678f8e2c1db4484",
+      "bytes": 79628,
+      "preservation": "byte-exact original evaluator output; initial evaluation has no joined mask references"
+    },
+    {
+      "path": "candidate-provenance.json",
+      "sha256": "17d1d5970d1f4b9e3785281639c043172119f7dd7dc1cd3827492313d0ed5e62",
+      "bytes": 9491,
+      "preservation": "byte-exact original evaluator output; initial evaluation has no joined mask references"
+    },
+    {
+      "path": "control-evaluation.json",
+      "sha256": "8043f1124caf3224e5bd20b46e337bdd3b014a28eb2b842b64488042cc652576",
+      "bytes": 25062,
+      "preservation": "byte-exact original evaluator output; initial evaluation has no joined mask references"
+    },
+    {
+      "path": "control-provenance.json",
+      "sha256": "3642407e848cfabc6f0fbad0fe5d606a52fec14d34482efbe43df8413fe24162",
+      "bytes": 1410,
+      "preservation": "byte-exact original evaluator output; initial evaluation has no joined mask references"
+    },
+    {
+      "path": "paired-qualification.json",
+      "sha256": "3abe7c561901f5ea2eb34f802db5da872dd7264615dcfcc35db4b4952d7078ee",
+      "bytes": 654,
+      "preservation": "byte-exact original evaluator output; initial evaluation has no joined mask references"
+    },
+    {
+      "path": "candidate-physical-receipts.json",
+      "sha256": "d5eec9791bca9db46257d088d64fd96f2e41282b9f359a0880d8c08d069632be",
+      "bytes": 513037,
+      "preservation": "metadata-only JSON derivative of original physicalReceipts array; no request bodies"
+    },
+    {
+      "path": "control-physical-receipts.json",
+      "sha256": "97e7aa487ef57f27a09902b2c82e8da2c8b5fa61d1ac562d61ba4ccdf4c54796",
+      "bytes": 145621,
+      "preservation": "metadata-only JSON derivative of original physicalReceipts array; no request bodies"
+    }
+  ]
+}

package/docs/evidence/v02/categorical-presence-dev6.json

@@ -0,0 +1,144 @@
+{
+  "schemaVersion": 1,
+  "title": "Categorical presence split-port DEV6 pilot independent evaluation",
+  "verdict": "FAIL: ineligible for full DEV40; the predeclared pilot requires at least one of two positive whole-object locations and zero false locations among four negatives. This run located zero positives.",
+  "purpose": "six exposed DEV cases to evaluate caller-owned categorical presence over native Choice; not a release qualification run",
+  "phase": "dev-pilot",
+  "profileId": "split-sol-spatial-native-luna-categorical-presence-sam-v1",
+  "runId": "v02-categorical-presence-dev6-2026-10-20T161832712Z-98826bc1",
+  "pins": {
+    "summaryCanonicalSha256": "a5a0f03b8fb889868505ea5a94da75db06c43f4b71ed80bbaec900a2fce03bc2",
+    "summaryPrettyFileSha256": "33deeff7037e5cc43d0dce5148ecbd2f65241a5e66ac4aefe0a1e2f360045022",
+    "runManifestSha256": "5d9b535365ac30a9057f4fba4ab69164eb0c6f5eeb77438aba1d224cc9a5213f",
+    "completionMarkerSha256": "3a13028bdcf6ddefa03ca518d07299bc86b5b34b7496ad71d57c491488cb8416",
+    "selectedInputsSha256": "98c9cae05ce1e85cd3dba6fb93b86406caf258600ac7bef8d64c7edd4439bc12",
+    "parentTwentyPhotoInputsSha256": "6c9ed29d8586c7887ae0861fc853baf203cb43790df4959390da4b5cc9ac679a",
+    "devTruthSha256": "793e26845deaa2cb56748ef270b8cb02877218d680068573f02e5af92b26af6a",
+    "variantSha256": "069a6c1bd8b32cd86e6af5838282817626b69c9790f0bb8dd9c4868450d86e6c",
+    "producerSourceSha256": "23ba879559c3fb69fd704589c89b98e2e601abc2cacd0d811a0b94e021613221"
+  },
+  "completeness": {
+    "planned": 6,
+    "completed": 6,
+    "missing": 0,
+    "status": "complete"
+  },
+  "classification": {
+    "correct": 4,
+    "total": 6,
+    "accuracy": 0.6666666666666666,
+    "positiveCorrect": 2,
+    "positiveTotal": 2,
+    "negativeCorrect": 2,
+    "negativeTotal": 4,
+    "falsePresent": 2
+  },
+  "wholeObjectLocalization": {
+    "locatedPositive": 0,
+    "positiveTotal": 2,
+    "falseLocatedNegative": 0,
+    "negativeTotal": 4,
+    "unknownTotal": 6,
+    "reasonCounts": {
+      "extent-uncertain": 2,
+      "uncertain-presence": 1,
+      "candidate-union-empty": 2,
+      "contradictory-evidence": 1
+    },
+    "bboxMetric": "not available: no row returned a location"
+  },
+  "extraction": {
+    "extracted": 0,
+    "unknown": 6,
+    "maskArtifactCases": 0,
+    "maskMetric": "not available: no masks returned"
+  },
+  "budget": {
+    "physicalCalls": 39,
+    "knownCostUsd": 0.23919200000000004,
+    "unknownCostCalls": 0,
+    "inFlightCalls": 0,
+    "latencyRankingEligible": false,
+    "latencyEnvironment": "concurrent-development-screen-unranked",
+    "receiptsReconciled": true
+  },
+  "cases": [
+    {
+      "caseId": "photo-dev-present-01",
+      "imageSha256": "71e781adb0c2d47d97e6fd5868e6dbe4de8c1375536e1753f897931ff90575dc",
+      "expectedPresence": "present",
+      "classification": "present",
+      "classificationCorrect": true,
+      "locateStatus": "unknown",
+      "locateReason": "extent-uncertain",
+      "wholeObjectBox": null,
+      "extractStatus": "unknown",
+      "maskArtifactPresent": false
+    },
+    {
+      "caseId": "photo-dev-present-02",
+      "imageSha256": "a27d93b12fadd550899a1d1d8ce70d34ba3c692affebf09134eda838910fcab5",
+      "expectedPresence": "present",
+      "classification": "present",
+      "classificationCorrect": true,
+      "locateStatus": "unknown",
+      "locateReason": "uncertain-presence",
+      "wholeObjectBox": null,
+      "extractStatus": "unknown",
+      "maskArtifactPresent": false
+    },
+    {
+      "caseId": "photo-dev-absent-01",
+      "imageSha256": "604b8fa9c30881acee7c67d078bb7d08f846aeac6152dc1bf23a41ddefe45988",
+      "expectedPresence": "absent",
+      "classification": "absent",
+      "classificationCorrect": true,
+      "locateStatus": "unknown",
+      "locateReason": "candidate-union-empty",
+      "wholeObjectBox": null,
+      "extractStatus": "unknown",
+      "maskArtifactPresent": false
+    },
+    {
+      "caseId": "photo-dev-absent-02",
+      "imageSha256": "8bcbc695bc113f645bb66dd2e03615764cd37d80e5794d8f9ffc79f1e364e077",
+      "expectedPresence": "absent",
+      "classification": "present",
+      "classificationCorrect": false,
+      "locateStatus": "unknown",
+      "locateReason": "extent-uncertain",
+      "wholeObjectBox": null,
+      "extractStatus": "unknown",
+      "maskArtifactPresent": false
+    },
+    {
+      "caseId": "photo-dev-absent-03",
+      "imageSha256": "f8c2cc736fd7941bf2bc51c73147aeec96deba4b22e6948fe088a2d9121c51a6",
+      "expectedPresence": "absent",
+      "classification": "absent",
+      "classificationCorrect": true,
+      "locateStatus": "unknown",
+      "locateReason": "candidate-union-empty",
+      "wholeObjectBox": null,
+      "extractStatus": "unknown",
+      "maskArtifactPresent": false
+    },
+    {
+      "caseId": "photo-dev-absent-04",
+      "imageSha256": "085ac7dc3243e28816e69f5e1ad56b9e8858c7135fac366f3172ad862e28eeae",
+      "expectedPresence": "absent",
+      "classification": "present",
+      "classificationCorrect": false,
+      "locateStatus": "unknown",
+      "locateReason": "contradictory-evidence",
+      "wholeObjectBox": null,
+      "extractStatus": "unknown",
+      "maskArtifactPresent": false
+    }
+  ],
+  "limits": [
+    "Six exposed DEV cases only; this is a pilot and not a heldout evaluation.",
+    "No target boxes or masks were produced, so localization and extraction quality cannot be inferred from the classification answers.",
+    "The source-truth labels and original producer artifacts were left unchanged."
+  ]
+}

package/docs/evidence/v02/categorical-presence-dev6.md

@@ -0,0 +1,18 @@
+# Categorical-presence split-port DEV6 pilot
+
+The run completed all six exposed development cases, but it fails the predeclared pilot gate for moving to a full DEV40 run. The gate required at least one correctly located positive and zero false locations among four negatives. There were **0/2 positive locations**, so the candidate is not eligible for full DEV40.
+
+The independent classification score was **4/6** (2/2 positives and 2/4 negatives); two negative cases were classified present. All six whole-object location results were `unknown`; no box or mask was returned. Thus the run provides no measured box or cutout accuracy.
+
+The run used 39 physical calls at a reconciled known cost of **$0.239192**, with no unknown-cost or in-flight calls. Latency was marked unranked. This was a six-case DEV pilot, not a release or heldout result.
+
+| Case                 | Truth   | Classification    | Locate  | Reason                 |
+| -------------------- | ------- | ----------------- | ------- | ---------------------- |
+| photo-dev-present-01 | present | present (correct) | unknown | extent-uncertain       |
+| photo-dev-present-02 | present | present (correct) | unknown | uncertain-presence     |
+| photo-dev-absent-01  | absent  | absent (correct)  | unknown | candidate-union-empty  |
+| photo-dev-absent-02  | absent  | present (wrong)   | unknown | extent-uncertain       |
+| photo-dev-absent-03  | absent  | absent (correct)  | unknown | candidate-union-empty  |
+| photo-dev-absent-04  | absent  | present (wrong)   | unknown | contradictory-evidence |
+
+Raw producer artifacts are unchanged. The JSON evidence includes both the canonical summary digest sealed by the run marker and the pretty-file digest, plus the run manifest, completion marker, selected-input, parent-cohort and DEV-truth pins. No producer truth, queries or outputs were rewritten.

package/docs/evidence/v02/categorical-vision-method-selection.json

@@ -0,0 +1,90 @@
+{
+  "schema": "zorya-v02-method-selection-v1",
+  "recordedAt": "2026-10-20T17:09:43.940667+00:00",
+  "selectedDevProfileId": "native-luna-categorical-presence-sam-target-plus-neutral-v1",
+  "targetPolicy": "target-plus-neutral-instruction-v1",
+  "producerRevision": "e4f55baa2c3c8a9ccc711a8aafc56d0c69f4af82",
+  "producerSourceSha256": "7ec64fb859b8e322bcd5e5b8956c4b7c0be5e2ceec36d19fc08906e3912ca46d",
+  "decisionRecipeSourceSha256": {
+    "./categorical-presence-port.ts": "0d85dcb5acb12d10c4fab888f7a0b81e16c594c940886794fecc2b9367c26512",
+    "./native-categorical-presence-port.ts": "0bf4fa6a16c52083f1a97a696eeead26a6710d82bbacb27af4380ae01143eb3a",
+    "src/features/prompts.ts": "b808ed8f42e8c92daf08060d2805da618f2b140f99b6a9b07df49622d57b62aa",
+    "src/processes/locate.ts": "a5a1320c2dff4060a3c5ab5e32cb5f0af2d94fcad458caba3818b6a0a28dab3d",
+    "src/processes/proposal-location.ts": "0670d56e30b2f42e499d3311149da8bb9b1af83368c0908ae1538bc6d04877a3",
+    "src/infrastructure/segmentation/proposal-presentation.ts": "c1e5ab928a97209e090b076db56d7fac69d5fa743dacbfd4288307e576990c25"
+  },
+  "route": {
+    "model": "openai/gpt-6-luna-decisions",
+    "provider": "openai",
+    "providerName": "OpenAI",
+    "endpoint": "/api/alpha/decisions",
+    "fallback": false
+  },
+  "method": "pixel-mask-proposals-v1",
+  "native": {
+    "method": "mobile-sam-nine-point-proposals-v1",
+    "executionProvider": "cpu",
+    "threads": 2,
+    "maxActive": 1,
+    "maxWaiting": 2,
+    "encoderSha256": "580f5fb648ea1062c0aabc26217aed56921985f03f0cbbd852bba81d760cc749",
+    "decoderSha256": "8976b90a87ba50a6a72217a5ff994f7d25ce16f2229fcc1ed259e1294c622ffe"
+  },
+  "policy": {
+    "timeoutMs": 120000,
+    "maxInFlightCases": 2,
+    "boxTolerancePx": 8,
+    "maxExtentPasses": 3,
+    "retries": 0,
+    "fallback": false,
+    "categoricalPresence": "closed-choice-present-absent-unknown-v1",
+    "ranking": "existing validated maximum; stable supplied-order ties; raw scores retained"
+  },
+  "development": {
+    "profilesCompared": 2,
+    "photoMatches": {
+      "preserveInstruction": 3,
+      "targetOnly": 0
+    },
+    "photoPositiveDenominator": 16,
+    "classification": {
+      "preserveInstruction": 35,
+      "targetOnly": 35
+    },
+    "classificationDenominator": 40,
+    "falseLocations": {
+      "preserveInstruction": 0,
+      "targetOnly": 0
+    },
+    "negativeDenominator": 10,
+    "selectionRule": "Both passed the DEV negative condition; preserving the neutral instruction wins on photo whole-object matches before any latency/cost tie-break. All failures and unknowns count.",
+    "latencyUsedForRanking": false,
+    "pairedPreregistrationSha256": "c1f2000d59a0724acb41c5c46e647e8c964295e522e8722c2d2512723e02f69a",
+    "evaluationSha256": {
+      "preserveInstruction": "0779badea569409fad7f1938f1007bdea393a511a95b15661c0e310ca7af6e28",
+      "targetOnly": "dafce762ced70a0b877620a6ef1d473457aa03e099147e181539db77fd852c5b"
+    }
+  },
+  "heldoutExposureBeforeSelection": {
+    "newCasePixelsViewedByRootOrProducer": false,
+    "newPerCaseTruthViewedByRootOrProducer": false,
+    "curatorHasLabels": true,
+    "genericOracleSourceReviewed": true
+  },
+  "requiredBeforeBlindDispatch": [
+    "Reviewed complete executable producer/control source and source freeze",
+    "Canonical source-disjoint80input manifests and independent oracle integrity",
+    "Hosted final source checks",
+    "Completed monotonic native resource acceptance"
+  ],
+  "allowsHeldoutDispatch": false,
+  "heldoutQualified": false,
+  "releaseQualified": false,
+  "libraryModelDefault": null,
+  "limits": [
+    "Development selection is not a blind qualification or a universal quality claim.",
+    "This recipe is caller-owned and never installed as a library model/provider default.",
+    "Selected recipe settings and decision/source transformations must not change for blind qualification.",
+    "The later heldout run freeze must bind this selection digest, canonical input pins and its actual producer source closure."
+  ]
+}

package/docs/evidence/v02/checked-region-component-v1/README.md

@@ -0,0 +1,7 @@
+# Exposed checked-region component proof
+
+This metadata-only receipt records three paid DEV component checks on two already exposed photos. The first box for the Car instance was judged `needs-expansion` (0.00 / 1.00 / 0.00); the Knife box was judged `complete` (0.66 / 0.31 / 0.03); and a larger Car box from another same-seed MobileSAM proposal was judged `complete` (0.76 / 0.20 / 0.04). Together the raw receipts show 3 physical calls, $0.000292 known cost, 0 unknown-cost calls, and no in-flight calls.
+
+`evidence.json` pins both raw run manifests, summaries, completion markers, prompt/criteria, route, request hashes, saved-presentation hashes, and provider receipts. The saved images and raw call artifacts remain in the ignored `.run` area and are not copied here.
+
+These are component-level judgments on exposed DEV presentations, not an end-to-end accuracy result, a segmentation-mask evaluation, or release qualification. The calls used the legacy probe prompt and saved fixed-style presentations. The current production box checker uses separate closed-choice wording and renders the full source with the mask-derived box, seed-center cross, and hard-ROI outline. The historical calls did not exercise that production renderer or prompt. No heldout cases were accessed for this probe.