feat(lab): publish M4.8T quality evidence

2026-08-25 23:44:27 +03:00
parent 9a956c318a
commit 3ff3b9b587
20 changed files with 1419 additions and 4 deletions
@@ -45,6 +45,7 @@ import { M47ReferenceGraphResultView } from "./M47ReferenceGraphResult";
 import { M48ObjectCentricQualityResultView } from "./M48ObjectCentricQualityResult";
 import { M48SmallStaticPassageRegressionResultView } from "./M48SmallStaticPassageRegressionResult";
 import { M48SFixedClassDetectorResultView } from "./M48SFixedClassDetectorResult";
+import { M48TRiskQualityResultView } from "./M48TRiskQualityResult";

 export { isAdvancedLaboratoryWorkId };
 export type { AdvancedLaboratoryWorkId };
@@ -96,6 +97,9 @@ export function AdvancedLaboratoryResult({
  if (workId === "m48s-fixed-class-detector" && results.m48s) {
    return <M48SFixedClassDetectorResultView rigLabel={rigLabel} result={results.m48s} />;
  }
+  if (workId === "m48t-risk-quality-temporal" && results.m48t) {
+    return <M48TRiskQualityResultView rigLabel={rigLabel} result={results.m48t} />;
+  }
  if (workId === "m47-reference-graph-shadow" && results.m47Graph) {
    return <M47ReferenceGraphResultView rigLabel={rigLabel} result={results.m47Graph} />;
  }
@@ -0,0 +1,79 @@
+import {
+  LaboratoryEvidence,
+  LaboratoryResultSummary,
+  LaboratorySummary,
+  LaboratoryWorkTemplate,
+} from "../../components/laboratory/LaboratoryPresentation";
+import type { M48TRiskQualityResult } from "../../core/laboratory/m48tRiskQuality";
+import { M48TRiskQualityVisual } from "./M48TRiskQualityVisual";
+
+function percent(value: number, digits = 1): string {
+  return `${(value * 100).toLocaleString("ru-RU", { maximumFractionDigits: digits })}%`;
+}
+
+function decimal(value: number, digits = 1): string {
+  return value.toLocaleString("ru-RU", { maximumFractionDigits: digits });
+}
+
+export function M48TRiskQualityResultView({
+  rigLabel,
+  result,
+}: {
+  rigLabel: string;
+  result: M48TRiskQualityResult;
+}) {
+  return (
+    <LaboratoryWorkTemplate
+      summary={(
+        <LaboratorySummary
+          title="M4.8T · semantic quality + temporal identity"
+          description="RF-DETR-L проверен на полном независимом COCO val2017 по заранее записанным risk-family гейтам. Отдельно тот же advisory-класс стабилизирован на geometry-owned ID полного RAVNOVES00 replay; семантика не участвует в association или occupancy."
+          status="Quality gate failed · temporal invariant passed"
+          statusTone="warning"
+          facts={[
+            { label: "Quality truth", value: `COCO val2017 · ${result.source.quality.truthInstances.toLocaleString("ru-RU")} risk instances` },
+            { label: "Candidate", value: `${result.candidate.modelId} · threshold ${decimal(result.candidate.minimumScore, 2)}` },
+            { label: "Temporal source", value: `${rigLabel} · ${result.source.temporal.frames.toLocaleString("ru-RU")} world states` },
+            { label: "Authority", value: "SHADOW ONLY · geometry-owned occupancy · production NO" },
+          ]}
+          brief={{
+            question: "Достаточно ли качественна текущая risk-семантика RF-DETR и можно ли убрать кадровое мерцание класса без влияния на геометрию?",
+            approach: `Все ${result.source.quality.riskImages.toLocaleString("ru-RU")} COCO-изображений с risk-классами оценены при неизменном score ${decimal(result.candidate.minimumScore, 2)}. Затем bounded history 5 / confirm 2 / switch 3 применена только к advisory-классу уже существующих component ID.`,
+            principalResult: `Recall прошёл: ${percent(result.quality.microRecall)} overall и ${percent(result.quality.mediumLargeRecall)} medium+large. Precision ${percent(result.quality.microPrecision)} и vehicle recall ${percent(result.quality.families.vehicle)} не прошли гейты. Temporal shadow сократил class switches ${result.temporal.rawClassSwitches} → ${result.temporal.stableClassSwitches} и family switches ${result.temporal.rawFamilySwitches} → ${result.temporal.stableFamilySwitches}.`,
+            limitation: "COCO не является truth городского маршрута RAVNOVES00, а temporal replay не имеет независимой track/class truth. Batch timing не принимается как realtime-гейт.",
+          }}
+          method={result.method}
+        />
+      )}
+      evidence={(
+        <LaboratoryEvidence
+          eyebrow="M4.8T VISUAL EVIDENCE · INDEPENDENT COCO TRUTH"
+          title="16 hash-bound review cases: human truth против RF-DETR"
+          kind="diagnostic-model"
+          resizable
+        >
+          <M48TRiskQualityVisual result={result} />
+        </LaboratoryEvidence>
+      )}
+      result={(
+        <LaboratoryResultSummary
+          title="Temporal anti-flicker принят как инвариант; detector quality не принят"
+          status="2/10 predeclared quality checks failed"
+          statusTone="warning"
+          metrics={[
+            { label: "Precision / recall", value: `${percent(result.quality.microPrecision)} / ${percent(result.quality.microRecall)}`, hint: "gates ≥ 80% / ≥ 75%" },
+            { label: "Vehicle family", value: percent(result.quality.families.vehicle), hint: "gate ≥ 85% · failed" },
+            { label: "TP / FP / FN", value: `${result.counts.truePositive.toLocaleString("ru-RU")} / ${result.counts.falsePositive.toLocaleString("ru-RU")} / ${result.counts.falseNegative.toLocaleString("ru-RU")}`, hint: `${result.counts.predictions.toLocaleString("ru-RU")} predictions` },
+            { label: "Class / family switches", value: `${result.temporal.rawClassSwitches}→${result.temporal.stableClassSwitches} / ${result.temporal.rawFamilySwitches}→${result.temporal.stableFamilySwitches}`, hint: `peak active ${result.temporal.peakActiveComponents}/${result.temporal.maximumActiveComponents}` },
+            { label: "Batch throughput", value: `${decimal(result.execution.effectiveImagesPerSecond, 2)} image/s`, hint: `image p95 ${decimal(result.execution.imageP95Ms, 2)} ms · non-admission` },
+          ]}
+          conclusion={{
+            proved: `На независимой COCO truth текущий RF-DETR сохраняет высокий recall: ${percent(result.quality.microRecall)} overall и ${percent(result.quality.mediumLargeRecall)} medium+large. Bounded temporal state подавил или отложил ${result.temporal.suppressedClassSwitches} class-switch и ${result.temporal.suppressedFamilySwitches} family-switch, сохранив association и occupancy class-independent; active state ${result.temporal.peakActiveComponents}/${result.temporal.maximumActiveComponents}, eviction 0.`,
+            notProved: `Semantic candidate не принят: precision ${percent(result.quality.microPrecision)} при gate 80%, vehicle recall ${percent(result.quality.families.vehicle)} при gate 85%. Не доказаны RAVNOVES class truth, physical track identity, child/adult, unknown moving hazards, risk policy и realtime admission этой batch-командой.`,
+            decision: "Не менять зафиксированные гейты и не добавлять второй detector. RF-DETR остаётся shadow-кандидатом; temporal stabilizer допустим только как advisory-слой поверх geometry-owned ID.",
+          }}
+        />
+      )}
+    />
+  );
+}
@@ -0,0 +1,63 @@
+import { useState } from "react";
+import { Icon, IconButton } from "@nodedc/ui-react";
+
+import { LaboratoryEvidenceViewer } from "../../components/laboratory/LaboratoryEvidenceViewer";
+import type { M48TRiskQualityResult } from "../../core/laboratory/m48tRiskQuality";
+
+export function M48TRiskQualityVisual({ result }: { result: M48TRiskQualityResult }) {
+  const [index, setIndex] = useState(0);
+  const [expanded, setExpanded] = useState(false);
+  const item = result.review.cases[index] ?? null;
+  const navigate = (offset: -1 | 1) => {
+    setIndex((current) => (
+      current + offset + result.review.cases.length
+    ) % result.review.cases.length);
+  };
+
+  return (
+    <div className="l3-visual-audit">
+      <LaboratoryEvidenceViewer
+        label="M4.8T independent COCO quality review"
+        mode="quality"
+        modes={[{ value: "quality", label: "COCO TRUTH" }]}
+        expanded={expanded}
+        onModeChange={() => undefined}
+        onExpandedChange={setExpanded}
+        actions={(
+          <div className="l3-visual-audit__pagination">
+            <IconButton label="Предыдущий M4.8T review case" onClick={() => navigate(-1)}>
+              <Icon name="chevron-left" size={16} />
+            </IconButton>
+            <IconButton label="Следующий M4.8T review case" onClick={() => navigate(1)}>
+              <Icon name="chevron-right" size={16} />
+            </IconButton>
+          </div>
+        )}
+        overlay={item ? (
+          <div className="l3-visual-audit__overlay">
+            <div>
+              <span>COCO val2017 · independent human truth</span>
+              <strong>case {index + 1}/{result.review.cases.length} · image {item.imageId}</strong>
+              <small>Зелёный — truth · жёлтый — RF-DETR prediction · score ≥ 0,25</small>
+            </div>
+          </div>
+        ) : null}
+      >
+        {item ? (
+          <div className="l32-camera-scene">
+            <img
+              src={item.imageUrl}
+              alt={`M4.8T COCO review image ${item.imageId}: truth and RF-DETR boxes`}
+              draggable={false}
+            />
+          </div>
+        ) : (
+          <div className="l3-visual-audit__state" role="alert">
+            <Icon name="alert" size={18} />
+            M4.8T visual evidence недоступно.
+          </div>
+        )}
+      </LaboratoryEvidenceViewer>
+    </div>
+  );
+}
@@ -84,6 +84,13 @@ const KNOWN_WORKS: Readonly<Record<Exclude<LaboratoryWorkId, `session:${string}`
    experimentName: "RAVNOVES00 fixed-class risk detector",
    variantName: "M4.8S · RF-DETR-L TensorRT/Triton shadow",
  },
+  "m48t-risk-quality-temporal": {
+    profileId: "rig-ravnoves-perception-gate-v1",
+    profileName: (rigLabel) => `${rig(rigLabel)} RIGHT · COCO quality + bounded temporal identity`,
+    experimentId: "m48t-risk-quality-temporal",
+    experimentName: "RF-DETR independent semantic quality and temporal identity",
+    variantName: "M4.8T · COCO val2017 truth + RAVNOVES00 temporal shadow",
+  },
  "m47-reference-graph-shadow": {
    profileId: "rig-dual-evidence-virtual-corridor-v1",
    profileName: (rigLabel) => `${rig(rigLabel)} RIGHT · Camera + LiDAR dual evidence`,
@@ -22,6 +22,7 @@ function mergeResults(
    m48: next.m48 ?? current.m48,
    m48SmallStatic: next.m48SmallStatic ?? current.m48SmallStatic,
    m48s: next.m48s ?? current.m48s,
+    m48t: next.m48t ?? current.m48t,
    m4Threat: next.m4Threat ?? current.m4Threat,
    l3: next.l3 ?? current.l3,
    l31: next.l31 ?? current.l31,