diff --git a/examples/studio_demo/aiguard_portfolio_cases.json b/examples/studio_demo/aiguard_portfolio_cases.json
new file mode 100644
index 0000000..cc6fcc4
--- /dev/null
+++ b/examples/studio_demo/aiguard_portfolio_cases.json
@@ -0,0 +1,985 @@
+{
+ "schema_version": "inferedge-aiguard-portfolio-demo-v1",
+ "source": "InferEdgeAIGuard Phase 6 portfolio demo cases",
+ "scope": "local-first evidence replay",
+ "case_count": 4,
+ "cases": [
+ {
+ "case_id": "normal_pass",
+ "title": "Normal output quality",
+ "category": "normal",
+ "summary": "BBox and score evidence stay within configured thresholds.",
+ "guard_analysis": {
+ "schema_version": "inferedge-aiguard-diagnosis-v1",
+ "source": {
+ "runtime_result_path": "examples/single/fp32_normal.json"
+ },
+ "guard_verdict": "pass",
+ "severity": "low",
+ "confidence": 0.9,
+ "primary_reason": "BBox and score structural evidence is within configured thresholds.",
+ "evidence": [
+ {
+ "type": "bbox_validity",
+ "metric_name": "invalid_bbox_rate",
+ "observed_value": 0.0,
+ "baseline_value": null,
+ "threshold": 0.05,
+ "delta": null,
+ "delta_pct": null,
+ "increase_factor": null,
+ "severity": "low",
+ "status": "passed",
+ "explanation": "invalid_bbox_rate observed value is 0. Threshold is 0.05. This bbox_validity evidence should be reviewed before deployment.",
+ "why_it_matters": "Invalid boxes indicate that detection geometry may be unusable even when latency looks acceptable.",
+ "suspected_causes": [],
+ "recommendation": "BBox validity is within the configured threshold.",
+ "raw_context": {
+ "total_predictions": 3,
+ "invalid_bbox_count": 0,
+ "invalid_bbox_rate": 0.0,
+ "zero_area_count": 0,
+ "out_of_bounds_count": 0,
+ "nan_or_inf_count": 0,
+ "bbox_collapse_count": 0,
+ "bbox_collapse_ratio": 0.0,
+ "boundary_check_enabled": false
+ }
+ },
+ {
+ "type": "bbox_collapse",
+ "metric_name": "bbox_collapse_ratio",
+ "observed_value": 0.0,
+ "baseline_value": null,
+ "threshold": 0.05,
+ "delta": null,
+ "delta_pct": null,
+ "increase_factor": null,
+ "severity": "low",
+ "status": "passed",
+ "explanation": "bbox_collapse_ratio observed value is 0. Threshold is 0.05. This bbox_collapse evidence should be reviewed before deployment.",
+ "why_it_matters": "Near-zero area boxes can indicate decoder mismatch or quantization artifacts that make detections unusable.",
+ "suspected_causes": [],
+ "recommendation": "BBox collapse ratio is within the configured threshold.",
+ "raw_context": {
+ "total_predictions": 3,
+ "invalid_bbox_count": 0,
+ "invalid_bbox_rate": 0.0,
+ "zero_area_count": 0,
+ "out_of_bounds_count": 0,
+ "nan_or_inf_count": 0,
+ "bbox_collapse_count": 0,
+ "bbox_collapse_ratio": 0.0,
+ "boundary_check_enabled": false
+ }
+ },
+ {
+ "type": "score_range_violation",
+ "metric_name": "score_range_violation_count",
+ "observed_value": 0,
+ "baseline_value": null,
+ "threshold": 0,
+ "delta": null,
+ "delta_pct": null,
+ "increase_factor": null,
+ "severity": "low",
+ "status": "passed",
+ "explanation": "score_range_violation_count observed value is 0. Threshold is 0. This score_range_violation evidence should be reviewed before deployment.",
+ "why_it_matters": "Scores outside 0..1 or non-finite scores make ranking and thresholding unreliable.",
+ "suspected_causes": [],
+ "recommendation": "Score range is valid.",
+ "raw_context": {
+ "total_predictions": 3,
+ "min_score": 0.63,
+ "max_score": 0.91,
+ "mean_score": 0.7600000000000001,
+ "std_score": 0.11518101695447332,
+ "low_confidence_ratio": 0.0,
+ "high_confidence_ratio": 0.0,
+ "saturation_ratio": 0.0,
+ "score_range_violation_count": 0,
+ "score_nan_or_inf_count": 0,
+ "low_threshold": 0.01,
+ "high_threshold": 0.99
+ }
+ },
+ {
+ "type": "confidence_saturation",
+ "metric_name": "saturation_ratio",
+ "observed_value": 0.0,
+ "baseline_value": null,
+ "threshold": 0.7,
+ "delta": null,
+ "delta_pct": null,
+ "increase_factor": null,
+ "severity": "low",
+ "status": "passed",
+ "explanation": "saturation_ratio observed value is 0. Threshold is 0.7. This confidence_saturation evidence should be reviewed before deployment.",
+ "why_it_matters": "Confidence saturation can hide ranking quality problems and may indicate quantization or postprocess mistakes.",
+ "suspected_causes": [],
+ "recommendation": "Confidence distribution is within the configured threshold.",
+ "raw_context": {
+ "total_predictions": 3,
+ "min_score": 0.63,
+ "max_score": 0.91,
+ "mean_score": 0.7600000000000001,
+ "std_score": 0.11518101695447332,
+ "low_confidence_ratio": 0.0,
+ "high_confidence_ratio": 0.0,
+ "saturation_ratio": 0.0,
+ "score_range_violation_count": 0,
+ "score_nan_or_inf_count": 0,
+ "low_threshold": 0.01,
+ "high_threshold": 0.99
+ }
+ }
+ ],
+ "suspected_causes": [],
+ "recommendations": [
+ "BBox validity is within the configured threshold.",
+ "BBox collapse ratio is within the configured threshold.",
+ "Score range is valid.",
+ "Confidence distribution is within the configured threshold."
+ ],
+ "thresholds": {
+ "invalid_bbox_rate_review": 0.05,
+ "invalid_bbox_rate_blocked": 0.2,
+ "bbox_collapse_ratio_review": 0.05,
+ "bbox_collapse_ratio_high": 0.1,
+ "score_saturation_low_threshold": 0.01,
+ "score_saturation_high_threshold": 0.99,
+ "saturation_ratio_review": 0.7,
+ "saturation_ratio_high": 0.85
+ },
+ "baseline_summary": {},
+ "candidate_summary": {
+ "model": "yolov8n",
+ "precision": "fp32",
+ "image_id": "sample_001",
+ "bbox": {
+ "total_predictions": 3,
+ "invalid_bbox_count": 0,
+ "invalid_bbox_rate": 0.0,
+ "zero_area_count": 0,
+ "out_of_bounds_count": 0,
+ "nan_or_inf_count": 0,
+ "bbox_collapse_count": 0,
+ "bbox_collapse_ratio": 0.0,
+ "boundary_check_enabled": false
+ },
+ "score": {
+ "total_predictions": 3,
+ "min_score": 0.63,
+ "max_score": 0.91,
+ "mean_score": 0.7600000000000001,
+ "std_score": 0.11518101695447332,
+ "low_confidence_ratio": 0.0,
+ "high_confidence_ratio": 0.0,
+ "saturation_ratio": 0.0,
+ "score_range_violation_count": 0,
+ "score_nan_or_inf_count": 0,
+ "low_threshold": 0.01,
+ "high_threshold": 0.99
+ }
+ },
+ "created_at": "2026-05-02T12:54:53Z"
+ }
+ },
+ {
+ "case_id": "bbox_collapse_blocked",
+ "title": "Latency improvement with bbox collapse",
+ "category": "bbox_quality",
+ "summary": "Candidate latency improves, but bbox collapse increases enough to block deployment review.",
+ "guard_analysis": {
+ "schema_version": "inferedge-aiguard-diagnosis-v1",
+ "source": {
+ "baseline_result_path": "examples/single/fp32_normal.json",
+ "candidate_result_path": "examples/single/int8_bbox_collapse.json"
+ },
+ "guard_verdict": "blocked",
+ "severity": "high",
+ "confidence": 0.92,
+ "primary_reason": "Candidate output quality deviates from the baseline diagnosis profile.",
+ "evidence": [
+ {
+ "type": "baseline_deviation",
+ "metric_name": "invalid_bbox_rate_factor",
+ "observed_value": 500000000.0,
+ "baseline_value": 0.0,
+ "threshold": 5.0,
+ "delta": 0.5,
+ "delta_pct": null,
+ "increase_factor": 500000000.0,
+ "severity": "high",
+ "status": "failed",
+ "explanation": "invalid_bbox_rate changed from 0 in the baseline to 0.5 in the candidate (500000000x). Review threshold is 5x.",
+ "why_it_matters": "A candidate with a much higher invalid bbox rate than the baseline may have unusable detection geometry even if it runs faster.",
+ "suspected_causes": [
+ "Incorrect bbox decoder",
+ "Output tensor layout mismatch",
+ "Preprocessing/postprocess mismatch"
+ ],
+ "recommendation": "Review decoder, output layout, and postprocess settings.",
+ "raw_context": {
+ "baseline_detection_count": 3,
+ "candidate_detection_count": 2,
+ "detection_count_delta": -1,
+ "detection_count_delta_pct": -0.3333333333333333,
+ "detection_count_drop_pct": 0.3333333333333333,
+ "invalid_bbox_rate_factor": 500000000.0,
+ "bbox_collapse_ratio_factor": 500000000.0,
+ "score_saturation_factor": 1.0,
+ "latency_delta_pct": -0.735857362976007,
+ "accuracy_delta_pp": null
+ }
+ },
+ {
+ "type": "baseline_deviation",
+ "metric_name": "bbox_collapse_ratio_factor",
+ "observed_value": 500000000.0,
+ "baseline_value": 0.0,
+ "threshold": 5.0,
+ "delta": 0.5,
+ "delta_pct": null,
+ "increase_factor": 500000000.0,
+ "severity": "high",
+ "status": "failed",
+ "explanation": "bbox_collapse_ratio changed from 0 in the baseline to 0.5 in the candidate (500000000x). Review threshold is 5x.",
+ "why_it_matters": "Near-zero boxes increasing over baseline can indicate decoder or quantization problems that make detections unreliable.",
+ "suspected_causes": [
+ "INT8 quantization artifact",
+ "Incorrect bbox decoder",
+ "Preprocessing/postprocess mismatch"
+ ],
+ "recommendation": "Do not deploy until bbox collapse drift is reviewed.",
+ "raw_context": {
+ "baseline_detection_count": 3,
+ "candidate_detection_count": 2,
+ "detection_count_delta": -1,
+ "detection_count_delta_pct": -0.3333333333333333,
+ "detection_count_drop_pct": 0.3333333333333333,
+ "invalid_bbox_rate_factor": 500000000.0,
+ "bbox_collapse_ratio_factor": 500000000.0,
+ "score_saturation_factor": 1.0,
+ "latency_delta_pct": -0.735857362976007,
+ "accuracy_delta_pp": null
+ }
+ },
+ {
+ "type": "baseline_deviation",
+ "metric_name": "score_saturation_factor",
+ "observed_value": 1.0,
+ "baseline_value": 0.0,
+ "threshold": 5.0,
+ "delta": 0.0,
+ "delta_pct": null,
+ "increase_factor": 1.0,
+ "severity": "low",
+ "status": "passed",
+ "explanation": "saturation_ratio changed from 0 in the baseline to 0 in the candidate (1x). Review threshold is 5x.",
+ "why_it_matters": "Score saturation drift can hide ranking failures and may indicate quantization or score decoder mistakes.",
+ "suspected_causes": [],
+ "recommendation": "Baseline deviation is within threshold.",
+ "raw_context": {
+ "baseline_detection_count": 3,
+ "candidate_detection_count": 2,
+ "detection_count_delta": -1,
+ "detection_count_delta_pct": -0.3333333333333333,
+ "detection_count_drop_pct": 0.3333333333333333,
+ "invalid_bbox_rate_factor": 500000000.0,
+ "bbox_collapse_ratio_factor": 500000000.0,
+ "score_saturation_factor": 1.0,
+ "latency_delta_pct": -0.735857362976007,
+ "accuracy_delta_pp": null
+ }
+ },
+ {
+ "type": "detection_count_drift",
+ "metric_name": "detection_count_drop_pct",
+ "observed_value": 0.3333333333333333,
+ "baseline_value": 3,
+ "threshold": 0.5,
+ "delta": -1,
+ "delta_pct": -0.3333333333333333,
+ "increase_factor": null,
+ "severity": "low",
+ "status": "passed",
+ "explanation": "Candidate detection count changed from 3 to 2 (-0.333 signed delta).",
+ "why_it_matters": "A large detection count drop can mean the candidate became faster by missing objects rather than by improving execution efficiency.",
+ "suspected_causes": [],
+ "recommendation": "Detection count drift is within threshold.",
+ "raw_context": {
+ "baseline_detection_count": 3,
+ "candidate_detection_count": 2,
+ "detection_count_delta": -1,
+ "detection_count_delta_pct": -0.3333333333333333,
+ "detection_count_drop_pct": 0.3333333333333333,
+ "invalid_bbox_rate_factor": 500000000.0,
+ "bbox_collapse_ratio_factor": 500000000.0,
+ "score_saturation_factor": 1.0,
+ "latency_delta_pct": -0.735857362976007,
+ "accuracy_delta_pp": null
+ }
+ },
+ {
+ "type": "latency_quality_tradeoff",
+ "metric_name": "latency_delta_pct",
+ "observed_value": -0.735857362976007,
+ "baseline_value": 45.43,
+ "threshold": 0,
+ "delta": -33.43,
+ "delta_pct": -0.735857362976007,
+ "increase_factor": null,
+ "severity": "high",
+ "status": "failed",
+ "explanation": "Candidate latency improved by 0.736, but baseline comparison evidence indicates output quality risk.",
+ "why_it_matters": "Latency improvement alone is not deployment-ready evidence when bbox, score, or detection count quality regresses.",
+ "suspected_causes": [
+ "Quantization artifact",
+ "Decoder mismatch",
+ "Run configuration mismatch"
+ ],
+ "recommendation": "Treat speedup as review evidence until output quality drift is resolved.",
+ "raw_context": {
+ "baseline_detection_count": 3,
+ "candidate_detection_count": 2,
+ "detection_count_delta": -1,
+ "detection_count_delta_pct": -0.3333333333333333,
+ "detection_count_drop_pct": 0.3333333333333333,
+ "invalid_bbox_rate_factor": 500000000.0,
+ "bbox_collapse_ratio_factor": 500000000.0,
+ "score_saturation_factor": 1.0,
+ "latency_delta_pct": -0.735857362976007,
+ "accuracy_delta_pp": null
+ }
+ }
+ ],
+ "suspected_causes": [
+ "Incorrect bbox decoder",
+ "Output tensor layout mismatch",
+ "Preprocessing/postprocess mismatch",
+ "INT8 quantization artifact",
+ "Quantization artifact",
+ "Decoder mismatch",
+ "Run configuration mismatch"
+ ],
+ "recommendations": [
+ "Review decoder, output layout, and postprocess settings.",
+ "Do not deploy until bbox collapse drift is reviewed.",
+ "Baseline deviation is within threshold.",
+ "Detection count drift is within threshold.",
+ "Treat speedup as review evidence until output quality drift is resolved."
+ ],
+ "thresholds": {
+ "invalid_bbox_rate_review": 0.05,
+ "invalid_bbox_rate_blocked": 0.2,
+ "bbox_collapse_ratio_review": 0.05,
+ "bbox_collapse_ratio_high": 0.1,
+ "score_saturation_low_threshold": 0.01,
+ "score_saturation_high_threshold": 0.99,
+ "saturation_ratio_review": 0.7,
+ "saturation_ratio_high": 0.85,
+ "invalid_bbox_rate_factor_review": 5.0,
+ "invalid_bbox_rate_factor_blocked": 10.0,
+ "bbox_collapse_ratio_factor_review": 5.0,
+ "bbox_collapse_ratio_factor_blocked": 10.0,
+ "score_saturation_factor_review": 5.0,
+ "score_saturation_factor_blocked": 10.0,
+ "detection_count_drop_pct_review": 0.5,
+ "detection_count_drop_pct_blocked": 0.8
+ },
+ "baseline_summary": {
+ "label": "fp32_baseline",
+ "model": "yolov8n",
+ "precision": "fp32",
+ "bbox": {
+ "total_predictions": 3,
+ "invalid_bbox_count": 0,
+ "invalid_bbox_rate": 0.0,
+ "zero_area_count": 0,
+ "out_of_bounds_count": 0,
+ "nan_or_inf_count": 0,
+ "bbox_collapse_count": 0,
+ "bbox_collapse_ratio": 0.0,
+ "boundary_check_enabled": false
+ },
+ "score": {
+ "total_predictions": 3,
+ "min_score": 0.63,
+ "max_score": 0.91,
+ "mean_score": 0.7600000000000001,
+ "std_score": 0.11518101695447332,
+ "low_confidence_ratio": 0.0,
+ "high_confidence_ratio": 0.0,
+ "saturation_ratio": 0.0,
+ "score_range_violation_count": 0,
+ "score_nan_or_inf_count": 0,
+ "low_threshold": 0.01,
+ "high_threshold": 0.99
+ },
+ "latency_ms": 45.43,
+ "accuracy": null
+ },
+ "candidate_summary": {
+ "label": "int8_bbox_collapse",
+ "model": "yolov8n",
+ "precision": "int8",
+ "bbox": {
+ "total_predictions": 2,
+ "invalid_bbox_count": 1,
+ "invalid_bbox_rate": 0.5,
+ "zero_area_count": 1,
+ "out_of_bounds_count": 0,
+ "nan_or_inf_count": 0,
+ "bbox_collapse_count": 1,
+ "bbox_collapse_ratio": 0.5,
+ "boundary_check_enabled": false
+ },
+ "score": {
+ "total_predictions": 2,
+ "min_score": 0.72,
+ "max_score": 0.89,
+ "mean_score": 0.8049999999999999,
+ "std_score": 0.08500000000000002,
+ "low_confidence_ratio": 0.0,
+ "high_confidence_ratio": 0.0,
+ "saturation_ratio": 0.0,
+ "score_range_violation_count": 0,
+ "score_nan_or_inf_count": 0,
+ "low_threshold": 0.01,
+ "high_threshold": 0.99
+ },
+ "latency_ms": 12.0,
+ "accuracy": null,
+ "comparison": {
+ "baseline_detection_count": 3,
+ "candidate_detection_count": 2,
+ "detection_count_delta": -1,
+ "detection_count_delta_pct": -0.3333333333333333,
+ "detection_count_drop_pct": 0.3333333333333333,
+ "invalid_bbox_rate_factor": 500000000.0,
+ "bbox_collapse_ratio_factor": 500000000.0,
+ "score_saturation_factor": 1.0,
+ "latency_delta_pct": -0.735857362976007,
+ "accuracy_delta_pp": null
+ }
+ },
+ "created_at": "2026-05-02T12:54:53Z"
+ }
+ },
+ {
+ "case_id": "score_saturation_blocked",
+ "title": "Confidence score saturation",
+ "category": "confidence_distribution",
+ "summary": "Candidate scores concentrate near 0 or 1, indicating possible quantization or postprocess risk.",
+ "guard_analysis": {
+ "schema_version": "inferedge-aiguard-diagnosis-v1",
+ "source": {
+ "baseline_result_path": "examples/single/fp32_normal.json",
+ "candidate_result_path": "examples/single/int8_conf_saturation.json"
+ },
+ "guard_verdict": "blocked",
+ "severity": "high",
+ "confidence": 0.92,
+ "primary_reason": "Candidate output quality deviates from the baseline diagnosis profile.",
+ "evidence": [
+ {
+ "type": "baseline_deviation",
+ "metric_name": "invalid_bbox_rate_factor",
+ "observed_value": 1.0,
+ "baseline_value": 0.0,
+ "threshold": 5.0,
+ "delta": 0.0,
+ "delta_pct": null,
+ "increase_factor": 1.0,
+ "severity": "low",
+ "status": "passed",
+ "explanation": "invalid_bbox_rate changed from 0 in the baseline to 0 in the candidate (1x). Review threshold is 5x.",
+ "why_it_matters": "A candidate with a much higher invalid bbox rate than the baseline may have unusable detection geometry even if it runs faster.",
+ "suspected_causes": [],
+ "recommendation": "Baseline deviation is within threshold.",
+ "raw_context": {
+ "baseline_detection_count": 3,
+ "candidate_detection_count": 5,
+ "detection_count_delta": 2,
+ "detection_count_delta_pct": 0.6666666666666666,
+ "detection_count_drop_pct": 0.0,
+ "invalid_bbox_rate_factor": 1.0,
+ "bbox_collapse_ratio_factor": 1.0,
+ "score_saturation_factor": 1000000000.0,
+ "latency_delta_pct": -0.6808276469293418,
+ "accuracy_delta_pp": null
+ }
+ },
+ {
+ "type": "baseline_deviation",
+ "metric_name": "bbox_collapse_ratio_factor",
+ "observed_value": 1.0,
+ "baseline_value": 0.0,
+ "threshold": 5.0,
+ "delta": 0.0,
+ "delta_pct": null,
+ "increase_factor": 1.0,
+ "severity": "low",
+ "status": "passed",
+ "explanation": "bbox_collapse_ratio changed from 0 in the baseline to 0 in the candidate (1x). Review threshold is 5x.",
+ "why_it_matters": "Near-zero boxes increasing over baseline can indicate decoder or quantization problems that make detections unreliable.",
+ "suspected_causes": [],
+ "recommendation": "Baseline deviation is within threshold.",
+ "raw_context": {
+ "baseline_detection_count": 3,
+ "candidate_detection_count": 5,
+ "detection_count_delta": 2,
+ "detection_count_delta_pct": 0.6666666666666666,
+ "detection_count_drop_pct": 0.0,
+ "invalid_bbox_rate_factor": 1.0,
+ "bbox_collapse_ratio_factor": 1.0,
+ "score_saturation_factor": 1000000000.0,
+ "latency_delta_pct": -0.6808276469293418,
+ "accuracy_delta_pp": null
+ }
+ },
+ {
+ "type": "baseline_deviation",
+ "metric_name": "score_saturation_factor",
+ "observed_value": 1000000000.0,
+ "baseline_value": 0.0,
+ "threshold": 5.0,
+ "delta": 1.0,
+ "delta_pct": null,
+ "increase_factor": 1000000000.0,
+ "severity": "high",
+ "status": "failed",
+ "explanation": "saturation_ratio changed from 0 in the baseline to 1 in the candidate (1000000000x). Review threshold is 5x.",
+ "why_it_matters": "Score saturation drift can hide ranking failures and may indicate quantization or score decoder mistakes.",
+ "suspected_causes": [
+ "Quantization artifact",
+ "Duplicated sigmoid/postprocess",
+ "Incorrect score decoder"
+ ],
+ "recommendation": "Review score decoder and quantization calibration.",
+ "raw_context": {
+ "baseline_detection_count": 3,
+ "candidate_detection_count": 5,
+ "detection_count_delta": 2,
+ "detection_count_delta_pct": 0.6666666666666666,
+ "detection_count_drop_pct": 0.0,
+ "invalid_bbox_rate_factor": 1.0,
+ "bbox_collapse_ratio_factor": 1.0,
+ "score_saturation_factor": 1000000000.0,
+ "latency_delta_pct": -0.6808276469293418,
+ "accuracy_delta_pp": null
+ }
+ },
+ {
+ "type": "detection_count_drift",
+ "metric_name": "detection_count_drop_pct",
+ "observed_value": 0.0,
+ "baseline_value": 3,
+ "threshold": 0.5,
+ "delta": 2,
+ "delta_pct": 0.6666666666666666,
+ "increase_factor": null,
+ "severity": "low",
+ "status": "passed",
+ "explanation": "Candidate detection count changed from 3 to 5 (0.667 signed delta).",
+ "why_it_matters": "A large detection count drop can mean the candidate became faster by missing objects rather than by improving execution efficiency.",
+ "suspected_causes": [],
+ "recommendation": "Detection count drift is within threshold.",
+ "raw_context": {
+ "baseline_detection_count": 3,
+ "candidate_detection_count": 5,
+ "detection_count_delta": 2,
+ "detection_count_delta_pct": 0.6666666666666666,
+ "detection_count_drop_pct": 0.0,
+ "invalid_bbox_rate_factor": 1.0,
+ "bbox_collapse_ratio_factor": 1.0,
+ "score_saturation_factor": 1000000000.0,
+ "latency_delta_pct": -0.6808276469293418,
+ "accuracy_delta_pp": null
+ }
+ },
+ {
+ "type": "latency_quality_tradeoff",
+ "metric_name": "latency_delta_pct",
+ "observed_value": -0.6808276469293418,
+ "baseline_value": 45.43,
+ "threshold": 0,
+ "delta": -30.93,
+ "delta_pct": -0.6808276469293418,
+ "increase_factor": null,
+ "severity": "high",
+ "status": "failed",
+ "explanation": "Candidate latency improved by 0.681, but baseline comparison evidence indicates output quality risk.",
+ "why_it_matters": "Latency improvement alone is not deployment-ready evidence when bbox, score, or detection count quality regresses.",
+ "suspected_causes": [
+ "Quantization artifact",
+ "Decoder mismatch",
+ "Run configuration mismatch"
+ ],
+ "recommendation": "Treat speedup as review evidence until output quality drift is resolved.",
+ "raw_context": {
+ "baseline_detection_count": 3,
+ "candidate_detection_count": 5,
+ "detection_count_delta": 2,
+ "detection_count_delta_pct": 0.6666666666666666,
+ "detection_count_drop_pct": 0.0,
+ "invalid_bbox_rate_factor": 1.0,
+ "bbox_collapse_ratio_factor": 1.0,
+ "score_saturation_factor": 1000000000.0,
+ "latency_delta_pct": -0.6808276469293418,
+ "accuracy_delta_pp": null
+ }
+ }
+ ],
+ "suspected_causes": [
+ "Quantization artifact",
+ "Duplicated sigmoid/postprocess",
+ "Incorrect score decoder",
+ "Decoder mismatch",
+ "Run configuration mismatch"
+ ],
+ "recommendations": [
+ "Baseline deviation is within threshold.",
+ "Review score decoder and quantization calibration.",
+ "Detection count drift is within threshold.",
+ "Treat speedup as review evidence until output quality drift is resolved."
+ ],
+ "thresholds": {
+ "invalid_bbox_rate_review": 0.05,
+ "invalid_bbox_rate_blocked": 0.2,
+ "bbox_collapse_ratio_review": 0.05,
+ "bbox_collapse_ratio_high": 0.1,
+ "score_saturation_low_threshold": 0.01,
+ "score_saturation_high_threshold": 0.99,
+ "saturation_ratio_review": 0.7,
+ "saturation_ratio_high": 0.85,
+ "invalid_bbox_rate_factor_review": 5.0,
+ "invalid_bbox_rate_factor_blocked": 10.0,
+ "bbox_collapse_ratio_factor_review": 5.0,
+ "bbox_collapse_ratio_factor_blocked": 10.0,
+ "score_saturation_factor_review": 5.0,
+ "score_saturation_factor_blocked": 10.0,
+ "detection_count_drop_pct_review": 0.5,
+ "detection_count_drop_pct_blocked": 0.8
+ },
+ "baseline_summary": {
+ "label": "fp32_baseline",
+ "model": "yolov8n",
+ "precision": "fp32",
+ "bbox": {
+ "total_predictions": 3,
+ "invalid_bbox_count": 0,
+ "invalid_bbox_rate": 0.0,
+ "zero_area_count": 0,
+ "out_of_bounds_count": 0,
+ "nan_or_inf_count": 0,
+ "bbox_collapse_count": 0,
+ "bbox_collapse_ratio": 0.0,
+ "boundary_check_enabled": false
+ },
+ "score": {
+ "total_predictions": 3,
+ "min_score": 0.63,
+ "max_score": 0.91,
+ "mean_score": 0.7600000000000001,
+ "std_score": 0.11518101695447332,
+ "low_confidence_ratio": 0.0,
+ "high_confidence_ratio": 0.0,
+ "saturation_ratio": 0.0,
+ "score_range_violation_count": 0,
+ "score_nan_or_inf_count": 0,
+ "low_threshold": 0.01,
+ "high_threshold": 0.99
+ },
+ "latency_ms": 45.43,
+ "accuracy": null
+ },
+ "candidate_summary": {
+ "label": "int8_score_saturation",
+ "model": "yolov8n",
+ "precision": "int8",
+ "bbox": {
+ "total_predictions": 5,
+ "invalid_bbox_count": 0,
+ "invalid_bbox_rate": 0.0,
+ "zero_area_count": 0,
+ "out_of_bounds_count": 0,
+ "nan_or_inf_count": 0,
+ "bbox_collapse_count": 0,
+ "bbox_collapse_ratio": 0.0,
+ "boundary_check_enabled": false
+ },
+ "score": {
+ "total_predictions": 5,
+ "min_score": 0.002,
+ "max_score": 0.999,
+ "mean_score": 0.5993999999999999,
+ "std_score": 0.48696061442379507,
+ "low_confidence_ratio": 0.4,
+ "high_confidence_ratio": 0.6,
+ "saturation_ratio": 1.0,
+ "score_range_violation_count": 0,
+ "score_nan_or_inf_count": 0,
+ "low_threshold": 0.01,
+ "high_threshold": 0.99
+ },
+ "latency_ms": 14.5,
+ "accuracy": null,
+ "comparison": {
+ "baseline_detection_count": 3,
+ "candidate_detection_count": 5,
+ "detection_count_delta": 2,
+ "detection_count_delta_pct": 0.6666666666666666,
+ "detection_count_drop_pct": 0.0,
+ "invalid_bbox_rate_factor": 1.0,
+ "bbox_collapse_ratio_factor": 1.0,
+ "score_saturation_factor": 1000000000.0,
+ "latency_delta_pct": -0.6808276469293418,
+ "accuracy_delta_pp": null
+ }
+ },
+ "created_at": "2026-05-02T12:54:53Z"
+ }
+ },
+ {
+ "case_id": "temporal_instability_review",
+ "title": "Temporal instability",
+ "category": "temporal_consistency",
+ "summary": "Frame-level detection count variance is high enough to require review before deployment.",
+ "guard_analysis": {
+ "schema_version": "inferedge-aiguard-diagnosis-v1",
+ "source": {
+ "sequence_path": "examples/portfolio_demo/temporal_instability"
+ },
+ "guard_verdict": "review_required",
+ "severity": "medium",
+ "confidence": 0.88,
+ "primary_reason": "Temporal consistency should be reviewed before deployment.",
+ "evidence": [
+ {
+ "type": "temporal_consistency",
+ "metric_name": "frame_to_frame_detection_count_cv",
+ "observed_value": 1.224744871391589,
+ "baseline_value": null,
+ "threshold": 1.0,
+ "delta": null,
+ "delta_pct": null,
+ "increase_factor": null,
+ "severity": "medium",
+ "status": "warning",
+ "explanation": "Frame-to-frame detection count CV is 1.225. Review threshold is 1.",
+ "why_it_matters": "Large detection count variance can indicate unstable output across adjacent frames in the same scene.",
+ "suspected_causes": [
+ "Temporal instability",
+ "Preprocessing/postprocess mismatch",
+ "Confidence threshold instability"
+ ],
+ "recommendation": "Review frame sequence output before deployment.",
+ "raw_context": {
+ "frame_count": 5,
+ "detection_counts": [
+ 0,
+ 10,
+ 0,
+ 10,
+ 0
+ ],
+ "mean_detection_count": 4.0,
+ "std_detection_count": 4.898979485566356,
+ "frame_to_frame_detection_count_cv": 1.224744871391589,
+ "zero_detection_frame_count": 3,
+ "zero_detection_frame_ratio": 0.6,
+ "bbox_center_jump_mean": 0.0,
+ "bbox_center_jump_p95": 0.0,
+ "bbox_center_jump_unit": "pixels",
+ "class_flip_rate": 0.0,
+ "dominant_classes": [
+ null,
+ 0,
+ null,
+ 0,
+ null
+ ]
+ }
+ },
+ {
+ "type": "temporal_consistency",
+ "metric_name": "zero_detection_frame_ratio",
+ "observed_value": 0.6,
+ "baseline_value": null,
+ "threshold": 1.0,
+ "delta": null,
+ "delta_pct": null,
+ "increase_factor": null,
+ "severity": "low",
+ "status": "passed",
+ "explanation": "Zero-detection frame ratio is 0.6. Blocked threshold is 1.",
+ "why_it_matters": "Frequent zero-detection frames can mean objects disappear across a sequence even when individual frames appear valid.",
+ "suspected_causes": [],
+ "recommendation": "Zero-detection frame ratio is within threshold.",
+ "raw_context": {
+ "frame_count": 5,
+ "detection_counts": [
+ 0,
+ 10,
+ 0,
+ 10,
+ 0
+ ],
+ "mean_detection_count": 4.0,
+ "std_detection_count": 4.898979485566356,
+ "frame_to_frame_detection_count_cv": 1.224744871391589,
+ "zero_detection_frame_count": 3,
+ "zero_detection_frame_ratio": 0.6,
+ "bbox_center_jump_mean": 0.0,
+ "bbox_center_jump_p95": 0.0,
+ "bbox_center_jump_unit": "pixels",
+ "class_flip_rate": 0.0,
+ "dominant_classes": [
+ null,
+ 0,
+ null,
+ 0,
+ null
+ ]
+ }
+ },
+ {
+ "type": "temporal_consistency",
+ "metric_name": "bbox_center_jump_p95",
+ "observed_value": 0.0,
+ "baseline_value": null,
+ "threshold": 0.5,
+ "delta": null,
+ "delta_pct": null,
+ "increase_factor": null,
+ "severity": "low",
+ "status": "passed",
+ "explanation": "BBox center jump p95 is 0 (pixels). Review threshold is 0.5.",
+ "why_it_matters": "Large bbox center jumps across adjacent frames can indicate unstable geometry without requiring a full tracker.",
+ "suspected_causes": [],
+ "recommendation": "BBox center jumps are within threshold.",
+ "raw_context": {
+ "frame_count": 5,
+ "detection_counts": [
+ 0,
+ 10,
+ 0,
+ 10,
+ 0
+ ],
+ "mean_detection_count": 4.0,
+ "std_detection_count": 4.898979485566356,
+ "frame_to_frame_detection_count_cv": 1.224744871391589,
+ "zero_detection_frame_count": 3,
+ "zero_detection_frame_ratio": 0.6,
+ "bbox_center_jump_mean": 0.0,
+ "bbox_center_jump_p95": 0.0,
+ "bbox_center_jump_unit": "pixels",
+ "class_flip_rate": 0.0,
+ "dominant_classes": [
+ null,
+ 0,
+ null,
+ 0,
+ null
+ ]
+ }
+ },
+ {
+ "type": "temporal_consistency",
+ "metric_name": "class_flip_rate",
+ "observed_value": 0.0,
+ "baseline_value": null,
+ "threshold": 0.3,
+ "delta": null,
+ "delta_pct": null,
+ "increase_factor": null,
+ "severity": "low",
+ "status": "passed",
+ "explanation": "Dominant class flip rate is 0. Review threshold is 0.3.",
+ "why_it_matters": "Frequent dominant class changes in adjacent non-empty frames can indicate unstable classification in a detection sequence.",
+ "suspected_causes": [],
+ "recommendation": "Dominant class stability is within threshold.",
+ "raw_context": {
+ "frame_count": 5,
+ "detection_counts": [
+ 0,
+ 10,
+ 0,
+ 10,
+ 0
+ ],
+ "mean_detection_count": 4.0,
+ "std_detection_count": 4.898979485566356,
+ "frame_to_frame_detection_count_cv": 1.224744871391589,
+ "zero_detection_frame_count": 3,
+ "zero_detection_frame_ratio": 0.6,
+ "bbox_center_jump_mean": 0.0,
+ "bbox_center_jump_p95": 0.0,
+ "bbox_center_jump_unit": "pixels",
+ "class_flip_rate": 0.0,
+ "dominant_classes": [
+ null,
+ 0,
+ null,
+ 0,
+ null
+ ]
+ }
+ }
+ ],
+ "suspected_causes": [
+ "Temporal instability",
+ "Preprocessing/postprocess mismatch",
+ "Confidence threshold instability"
+ ],
+ "recommendations": [
+ "Review frame sequence output before deployment.",
+ "Zero-detection frame ratio is within threshold.",
+ "BBox center jumps are within threshold.",
+ "Dominant class stability is within threshold."
+ ],
+ "thresholds": {
+ "detection_count_cv_review": 1.0,
+ "zero_detection_frame_ratio_blocked": 1.0,
+ "bbox_center_jump_p95_review": 0.5,
+ "class_flip_rate_review": 0.3
+ },
+ "baseline_summary": {},
+ "candidate_summary": {
+ "sequence_id": "temporal_instability_demo",
+ "frame_count": 5,
+ "temporal": {
+ "frame_count": 5,
+ "detection_counts": [
+ 0,
+ 10,
+ 0,
+ 10,
+ 0
+ ],
+ "mean_detection_count": 4.0,
+ "std_detection_count": 4.898979485566356,
+ "frame_to_frame_detection_count_cv": 1.224744871391589,
+ "zero_detection_frame_count": 3,
+ "zero_detection_frame_ratio": 0.6,
+ "bbox_center_jump_mean": 0.0,
+ "bbox_center_jump_p95": 0.0,
+ "bbox_center_jump_unit": "pixels",
+ "class_flip_rate": 0.0,
+ "dominant_classes": [
+ null,
+ 0,
+ null,
+ 0,
+ null
+ ]
+ }
+ },
+ "created_at": "2026-05-02T12:54:53Z"
+ }
+ }
+ ]
+}
diff --git a/inferedgelab/studio/routes.py b/inferedgelab/studio/routes.py
index 026025d..8897ada 100644
--- a/inferedgelab/studio/routes.py
+++ b/inferedgelab/studio/routes.py
@@ -34,6 +34,7 @@
"contract_shape_mismatch_report.json",
)
LATENCY_REGRESSION_SUMMARY = "latency_regression_summary.json"
+AIGUARD_PORTFOLIO_CASES = "aiguard_portfolio_cases.json"
DEMO_JOB_ID = "demo_yolov8n_trt_vs_onnx"
STATIC_ASSETS = {
"app.js": "application/javascript",
@@ -172,6 +173,7 @@ def studio_demo_evidence(request: Request) -> dict[str, Any]:
results = [_load_demo_result(file_name) for file_name in DEMO_EVIDENCE_FILES]
evaluation_report = _load_demo_evaluation_report()
problem_cases = _load_demo_problem_cases()
+ guard_demo_cases = _load_aiguard_portfolio_cases()
imported_results = _get_imported_results(request)
imported_results.extend(results)
guard_analysis = _build_demo_guard_analysis(results, evaluation_report)
@@ -181,7 +183,7 @@ def studio_demo_evidence(request: Request) -> dict[str, Any]:
results[1],
guard_analysis=guard_analysis,
)
- demo_job = _build_demo_job(results, compare, evaluation_report, problem_cases)
+ demo_job = _build_demo_job(results, compare, evaluation_report, problem_cases, guard_demo_cases)
_get_demo_jobs(request)[DEMO_JOB_ID] = demo_job
return {
"status": "loaded",
@@ -194,6 +196,7 @@ def studio_demo_evidence(request: Request) -> dict[str, Any]:
"compare": compare,
"evaluation_report": evaluation_report,
"problem_cases": problem_cases,
+ "guard_demo_cases": guard_demo_cases,
"guard_analysis": guard_analysis,
"deployment_decision": compare["deployment_decision"],
}
@@ -397,6 +400,28 @@ def _load_latency_regression_summary() -> dict[str, Any]:
}
+def _load_aiguard_portfolio_cases() -> dict[str, Any]:
+ path = DEMO_EVIDENCE_DIR / AIGUARD_PORTFOLIO_CASES
+ try:
+ bundle = json.loads(path.read_text(encoding="utf-8"))
+ except OSError as exc:
+ raise HTTPException(status_code=500, detail=f"AIGuard portfolio cases not found: {AIGUARD_PORTFOLIO_CASES}") from exc
+ except json.JSONDecodeError as exc:
+ raise HTTPException(status_code=500, detail=f"AIGuard portfolio cases are invalid JSON: {AIGUARD_PORTFOLIO_CASES}") from exc
+
+ cases = bundle.get("cases") if isinstance(bundle, dict) else None
+ if not isinstance(cases, list):
+ raise HTTPException(status_code=500, detail=f"AIGuard portfolio cases schema error: {AIGUARD_PORTFOLIO_CASES}")
+
+ return {
+ "schema_version": bundle.get("schema_version"),
+ "source": f"examples/studio_demo/{AIGUARD_PORTFOLIO_CASES}",
+ "scope": bundle.get("scope"),
+ "case_count": bundle.get("case_count", len(cases)),
+ "cases": cases,
+ }
+
+
def _load_problem_report(file_name: str) -> dict[str, Any]:
path = VALIDATION_PROBLEM_DIR / file_name
try:
@@ -429,6 +454,7 @@ def _build_demo_job(
compare: dict[str, Any],
evaluation_report: dict[str, Any],
problem_cases: list[dict[str, Any]],
+ guard_demo_cases: dict[str, Any],
) -> dict[str, Any]:
now = _utc_now_iso()
runtime_result = results[-1] if results else {}
@@ -450,6 +476,7 @@ def _build_demo_job(
"guard_analysis": compare.get("guard_analysis"),
"evaluation_report": evaluation_report,
"problem_cases": problem_cases,
+ "guard_demo_cases": guard_demo_cases,
"summary": compare["judgement"]["summary"],
},
"error": None,
diff --git a/inferedgelab/studio/static/app.js b/inferedgelab/studio/static/app.js
index 0987777..576ead9 100644
--- a/inferedgelab/studio/static/app.js
+++ b/inferedgelab/studio/static/app.js
@@ -31,6 +31,7 @@ let importedResult = null;
let demoEvaluationReport = null;
let demoProblemCases = [];
let activeGuardAnalysis = null;
+let guardDemoCases = null;
const importedResultsByJobId = {};
function createElement(tagName, className, textContent) {
@@ -373,6 +374,7 @@ async function loadDemoEvidence() {
importedResult = results[results.length - 1] || null;
demoEvaluationReport = payload.evaluation_report || null;
demoProblemCases = Array.isArray(payload.problem_cases) ? payload.problem_cases : [];
+ guardDemoCases = payload.guard_demo_cases || null;
compareData = payload.compare || null;
updateGuardEvidence(payload.guard_analysis || payload.compare?.guard_analysis || null);
selectedJobId = payload.job_id || payload.job?.job_id || selectedJobId;
@@ -384,6 +386,7 @@ async function loadDemoEvidence() {
renderImportEvidence({ result: importedResult });
renderDemoEvaluation(demoEvaluationReport);
renderDemoProblemCases(demoProblemCases);
+ renderGuardDemoCases(guardDemoCases);
renderImportedResult();
await loadJobs(selectedJobId);
await loadCompare();
@@ -420,6 +423,46 @@ function renderDemoProblemCases(problemCases = []) {
});
}
+function renderGuardDemoCases(bundle) {
+ const target = document.querySelector("#guard-demo-cases");
+ if (!target) {
+ return;
+ }
+ target.replaceChildren();
+
+ const cases = Array.isArray(bundle?.cases) ? bundle.cases : [];
+ if (!cases.length) {
+ return;
+ }
+
+ const heading = createElement("div", "guard-demo-heading");
+ heading.append(
+ createElement("p", "caption", "Portfolio diagnosis cases"),
+ createElement("h3", "", "AIGuard demo evidence"),
+ createElement(
+ "p",
+ "body-text",
+ "Normal, bbox collapse, score saturation, and temporal instability evidence are replayed from local fixtures.",
+ ),
+ );
+ target.append(heading);
+
+ cases.forEach((item) => {
+ const analysis = item.guard_analysis || {};
+ const verdict = guardVerdict(analysis);
+ const card = createElement("article", `guard-demo-card ${decisionTone(verdict)}`);
+ card.append(
+ createElement("p", "caption", item.category || "diagnosis"),
+ createElement("h4", "", item.title || item.case_id || "AIGuard case"),
+ createElement("p", "body-text", item.summary || analysis.primary_reason || "-"),
+ evidenceItem("guard_verdict", verdict),
+ evidenceItem("severity", analysis.severity || "-"),
+ evidenceItem("primary_metric", primaryGuardMetric(analysis)),
+ );
+ target.append(card);
+ });
+}
+
function problemCaseDetail(problem = {}) {
if (problem.problem_case_type === "runtime_latency" || problem.latency_checks) {
const checks = problem.latency_checks || {};
@@ -523,6 +566,7 @@ function renderRunPanel() {
setState("#demo-state", "idle");
renderDemoEvaluation(null);
renderDemoProblemCases([]);
+ renderGuardDemoCases(null);
}
function resetTransientInputs() {
@@ -1001,6 +1045,12 @@ function guardEvidenceItems(guardAnalysis = {}) {
return [];
}
+function primaryGuardMetric(guardAnalysis = {}) {
+ const evidence = guardEvidenceItems(guardAnalysis);
+ const primary = evidence.find((item) => item.status === "failed" || item.status === "warning") || evidence[0] || {};
+ return primary.metric_name || primary.type || "-";
+}
+
function decisionReason(decision) {
const decisionName = String(decision?.decision || "unknown").toLowerCase();
if (decisionName === "unknown" && !decision?.guard_status) {
diff --git a/inferedgelab/studio/static/index.html b/inferedgelab/studio/static/index.html
index bd07216..e6aae85 100644
--- a/inferedgelab/studio/static/index.html
+++ b/inferedgelab/studio/static/index.html
@@ -137,8 +137,8 @@
}
}
-
-
+
+
@@ -320,6 +320,7 @@ AIGuard Evidence
+
@@ -344,7 +345,7 @@ Future Work
-
-
+
+