Stage 09

Executed

Resource-aware explainability

Experiment EXP-XAI-0001 — explanations generated for the four evidence-selected FIXED_ORIGIN models, loaded from their frozen artifacts with no retraining. This stage prepares evidence for Stages 10-12; it does not itself report fidelity, stability, or latency conclusions (see below — all three remain NOT EXECUTED).

What actually ran

Global importance rows

4,864

Local samples

220

Reduced top-k rows

722

Stage-10 fidelity-prep rows

880

Method applicability — every model × method × scope

No model was forced through a method it does not scientifically support — every NOT_APPLICABLE below carries the actual capability-detection reason.

Model	Method	Scope	Status	Reason
MODEL-C1	`INTRINSIC_COEFFICIENT`	global	Executed
MODEL-C1	`INTRINSIC_IMPURITY`	global	Blocked	capability not detected on fitted estimator
MODEL-C1	`PERMUTATION_IMPORTANCE_MACRO_F1`	global	Executed
MODEL-C1	`SINGLE_FEATURE_ABLATION_LOCAL`	local	Executed
MODEL-C1	`INTRINSIC_COEFFICIENT`	local	Executed
MODEL-C2	`INTRINSIC_IMPURITY`	global	Executed
MODEL-C2	`INTRINSIC_COEFFICIENT`	global	Blocked	capability not detected on fitted estimator
MODEL-C2	`PERMUTATION_IMPORTANCE_MACRO_F1`	global	Executed
MODEL-C2	`SINGLE_FEATURE_ABLATION_LOCAL`	local	Executed
MODEL-C2	`INTRINSIC_COEFFICIENT`	local	Blocked	capability not detected on fitted estimator
MODEL-C3	`INTRINSIC_COEFFICIENT`	global	Blocked	capability not detected on fitted estimator
MODEL-C3	`INTRINSIC_IMPURITY`	global	Blocked	capability not detected on fitted estimator
MODEL-C3	`PERMUTATION_IMPORTANCE_MACRO_F1`	global	Executed
MODEL-C3	`SINGLE_FEATURE_ABLATION_LOCAL`	local	Executed
MODEL-C3	`INTRINSIC_COEFFICIENT`	local	Blocked	capability not detected on fitted estimator
MODEL-C4	`INTRINSIC_COEFFICIENT`	global	Blocked	capability not detected on fitted estimator
MODEL-C4	`INTRINSIC_IMPURITY`	global	Blocked	capability not detected on fitted estimator
MODEL-C4	`PERMUTATION_IMPORTANCE_MACRO_F1`	global	Executed
MODEL-C4	`SINGLE_FEATURE_ABLATION_LOCAL`	local	Executed
MODEL-C4	`INTRINSIC_COEFFICIENT`	local	Blocked	capability not detected on fitted estimator

Top-3 features (Batch 2 / model-level)

Model	Method	Batch	#1	#2	#3
MODEL-C1	`INTRINSIC_COEFFICIENT`	ALL	S9:dR_norm	S10:dR_norm	S2:dR_norm
MODEL-C1	`PERMUTATION_IMPORTANCE_MACRO_F1`	2	S9:dR_norm	S2:dR_norm	S10:dR_norm
MODEL-C2	`INTRINSIC_IMPURITY`	ALL	S2:dR	S1:dR	S1:dR_norm
MODEL-C2	`PERMUTATION_IMPORTANCE_MACRO_F1`	2	S1:EMAd_0.1	S9:EMAd_0.1	S2:EMAd_0.001
MODEL-C3	`PERMUTATION_IMPORTANCE_MACRO_F1`	2	S9:dR_norm	S13:EMAi_0.01	S14:EMAi_0.01
MODEL-C4	`PERMUTATION_IMPORTANCE_MACRO_F1`	2	S2:EMAd_0.1	S13:EMAi_0.1	S4:EMAi_0.1

Rankings differ by model and method — an empirical observation, not a claim about which ranking is more correct (that is Stage 10's question). Full research note: docs/experiments/STAGE09_RESOURCE_AWARE_XAI.md.

Local sample categories

CORRECT REPRESENTATIVE

103

MISCLASSIFIED

NEAR DECISION BOUNDARY

Artifacts

Validation stage status

Stage 09 prepares evidence; it does not itself compute these.

Measurement	Status
Explanation fidelity (Stage 10)	Executed
Explanation stability across drift (Stage 11)	Executed
Explanation latency (Stage 12)	Executed

Stage 10 fidelity

Experiment EXP-XAI-FIDELITY-001 tested whether ranked inputs actually influence frozen model behavior under controlled Batch-1-reference perturbations.

What does fidelity mean here?

Fidelity means that features identified as important change the frozen model's behavior when experimentally perturbed. It is not human interpretability, causal truth, chemical causality, or model accuracy.

Method/scope summaries

Bootstrap intervals

688

Scope	Model	Method	Mean selected − random	N	Unit
GLOBAL	MODEL-C1	`INTRINSIC_COEFFICIENT`	0.0370	45	proportion
GLOBAL	MODEL-C1	`PERMUTATION_IMPORTANCE_MACRO_F1`	0.0411	45	proportion
GLOBAL	MODEL-C2	`INTRINSIC_IMPURITY`	0.1784	45	proportion
GLOBAL	MODEL-C2	`PERMUTATION_IMPORTANCE_MACRO_F1`	0.1050	45	proportion
GLOBAL	MODEL-C3	`PERMUTATION_IMPORTANCE_MACRO_F1`	0.0275	45	proportion
GLOBAL	MODEL-C4	`PERMUTATION_IMPORTANCE_MACRO_F1`	0.0629	45	proportion
SENSOR	MODEL-C1	`INTRINSIC_COEFFICIENT`	0.0082	27	proportion
SENSOR	MODEL-C1	`PERMUTATION_IMPORTANCE_MACRO_F1`	0.0070	27	proportion
SENSOR	MODEL-C2	`INTRINSIC_IMPURITY`	0.2054	27	proportion
SENSOR	MODEL-C2	`PERMUTATION_IMPORTANCE_MACRO_F1`	0.1420	27	proportion
SENSOR	MODEL-C3	`PERMUTATION_IMPORTANCE_MACRO_F1`	-0.0010	27	proportion
SENSOR	MODEL-C4	`PERMUTATION_IMPORTANCE_MACRO_F1`	0.0456	27	proportion
LOCAL	MODEL-C1	`INTRINSIC_COEFFICIENT`	0.1407	216	probability
LOCAL	MODEL-C1	`SINGLE_FEATURE_ABLATION_LOCAL`	0.1266	216	probability
LOCAL	MODEL-C2	`SINGLE_FEATURE_ABLATION_LOCAL`	0.0846	212	probability
LOCAL	MODEL-C3	`SINGLE_FEATURE_ABLATION_LOCAL`	0.0614	216	probability
LOCAL	MODEL-C4	`SINGLE_FEATURE_ABLATION_LOCAL`	0.1081	236	probability

Results are heterogeneous; no universal explanation-fidelity claim is made. Local ablation is cumulative ablation consistency because its evaluation is partly circular.

Stage 11 stability under chronological drift

Experiment EXP-XAI-STABILITY-001 separates explanation change from input change and frozen-model output change.

What does stability mean here?

Stability asks whether similar model situations produce similar explanations, and whether explanation changes are proportionate to actual changes in sensor inputs or model behavior.

Stability does not mean

The explanation is faithful or causal.
The model is accurate.
Sensors are physically unchanged.
The same explanation should persist under real drift.

Method/scope summaries

Bootstrap intervals

106

Scope	Model	Method	Metric	Mean	N
GLOBAL_RANK	MODEL-C1	`PERMUTATION_IMPORTANCE_MACRO_F1`	spearman	0.3694	16
GLOBAL_RANK	MODEL-C2	`PERMUTATION_IMPORTANCE_MACRO_F1`	spearman	0.0970	16
GLOBAL_RANK	MODEL-C3	`PERMUTATION_IMPORTANCE_MACRO_F1`	spearman	0.1005	16
GLOBAL_RANK	MODEL-C4	`PERMUTATION_IMPORTANCE_MACRO_F1`	spearman	0.0593	16
GLOBAL_MAGNITUDE	MODEL-C1	`PERMUTATION_IMPORTANCE_MACRO_F1`	normalized_l1_distance	0.5448	16
GLOBAL_MAGNITUDE	MODEL-C2	`PERMUTATION_IMPORTANCE_MACRO_F1`	normalized_l1_distance	0.5984	16
GLOBAL_MAGNITUDE	MODEL-C3	`PERMUTATION_IMPORTANCE_MACRO_F1`	normalized_l1_distance	0.4679	16
GLOBAL_MAGNITUDE	MODEL-C4	`PERMUTATION_IMPORTANCE_MACRO_F1`	normalized_l1_distance	0.5680	16
SENSOR	MODEL-C1	`PERMUTATION_IMPORTANCE_MACRO_F1`	sensor_importance_l1_distance	0.3803	16
SENSOR	MODEL-C2	`PERMUTATION_IMPORTANCE_MACRO_F1`	sensor_importance_l1_distance	0.2802	16
SENSOR	MODEL-C3	`PERMUTATION_IMPORTANCE_MACRO_F1`	sensor_importance_l1_distance	0.2587	16
SENSOR	MODEL-C4	`PERMUTATION_IMPORTANCE_MACRO_F1`	sensor_importance_l1_distance	0.2702	16
LOCAL_NEIGHBOR	MODEL-C1	`INTRINSIC_COEFFICIENT`	explanation_distance	0.1732	98
LOCAL_NEIGHBOR	MODEL-C1	`SINGLE_FEATURE_ABLATION_LOCAL`	explanation_distance	0.2075	98
LOCAL_NEIGHBOR	MODEL-C2	`SINGLE_FEATURE_ABLATION_LOCAL`	explanation_distance	0.2806	97
LOCAL_NEIGHBOR	MODEL-C3	`SINGLE_FEATURE_ABLATION_LOCAL`	explanation_distance	0.1962	90
LOCAL_NEIGHBOR	MODEL-C4	`SINGLE_FEATURE_ABLATION_LOCAL`	explanation_distance	0.2583	75
CROSS_BATCH	MODEL-C1	`INTRINSIC_COEFFICIENT`	explanation_distance	0.4199	33
CROSS_BATCH	MODEL-C1	`SINGLE_FEATURE_ABLATION_LOCAL`	explanation_distance	0.4640	33
CROSS_BATCH	MODEL-C2	`SINGLE_FEATURE_ABLATION_LOCAL`	explanation_distance	0.5534	26
CROSS_BATCH	MODEL-C3	`SINGLE_FEATURE_ABLATION_LOCAL`	explanation_distance	0.4531	32
CROSS_BATCH	MODEL-C4	`SINGLE_FEATURE_ABLATION_LOCAL`	explanation_distance	0.4739	39

Evidence is method- and context-dependent. Intrinsic global vectors marked ALL were not fabricated into chronological series, and matched cross-batch samples are cross-sectional rather than longitudinal.

Computational cost

Experiment EXP-XAI-LATENCY-001 measured warm, steady-state host computation with raw nanosecond timing and verified single-thread controls.

Host measured

Executed

Median, p95, N, model, method, scope, environment, and artifact are retained below.

MCU not measured

Not executed

Physical nRF52840 latency, energy, Flash, and SRAM remain unmeasured.

Host latency is not an estimate of nRF52840 latency or energy.

Scope	Model	Method / phase	Median (µs)	p95 (µs)	N	Environment	Experiment
GLOBAL	MODEL-C1	`INTRINSIC_COEFFICIENT` ONE_TIME_GLOBAL_EXTRACTION_COST	12.8	13.805	100	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C1	`INTRINSIC_COEFFICIENT` EXPLANATION_COMPUTE	444.3	483.375	570	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C1	`INTRINSIC_COEFFICIENT` EXPLANATION_COMPUTE	443	499.8	450	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C1	`INTRINSIC_COEFFICIENT` EXPLANATION_COMPUTE	445	467.96	600	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
GLOBAL	MODEL-C1	`PERMUTATION_IMPORTANCE_MACRO_F1` GLOBAL_EXPLANATION_TOTAL	1,330,028.7	1,373,144.73	3	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
GLOBAL	MODEL-C1	`PERMUTATION_IMPORTANCE_MACRO_F1` GLOBAL_EXPLANATION_TOTAL	1,315,073.6	1,341,278.72	3	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
GLOBAL	MODEL-C1	`PERMUTATION_IMPORTANCE_MACRO_F1` GLOBAL_EXPLANATION_TOTAL	1,320,504.3	1,334,284.29	3	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C1	`SINGLE_FEATURE_ABLATION_LOCAL_NAIVE_REFERENCE` EXPLANATION_COMPUTE	40,892.8	41,754.34	5	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C1	`SINGLE_FEATURE_ABLATION_LOCAL_NAIVE_REFERENCE` EXPLANATION_COMPUTE	39,936.9	40,232	5	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C1	`SINGLE_FEATURE_ABLATION_LOCAL_NAIVE_REFERENCE` EXPLANATION_COMPUTE	39,340.7	39,478.3	5	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C1	`SINGLE_FEATURE_ABLATION_LOCAL_VECTORIZED` EXPLANATION_COMPUTE	707.9	757.67	570	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C1	`SINGLE_FEATURE_ABLATION_LOCAL_VECTORIZED` EXPLANATION_COMPUTE	707.85	765.015	450	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C1	`SINGLE_FEATURE_ABLATION_LOCAL_VECTORIZED` EXPLANATION_COMPUTE	711.5	802.295	600	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
GLOBAL	MODEL-C2	`INTRINSIC_IMPURITY` ONE_TIME_GLOBAL_EXTRACTION_COST	21,772.7	22,408.54	100	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
GLOBAL	MODEL-C2	`PERMUTATION_IMPORTANCE_MACRO_F1` GLOBAL_EXPLANATION_TOTAL	9,463,815.5	9,467,436.02	3	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
GLOBAL	MODEL-C2	`PERMUTATION_IMPORTANCE_MACRO_F1` GLOBAL_EXPLANATION_TOTAL	9,359,869.2	9,360,621.24	3	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
GLOBAL	MODEL-C2	`PERMUTATION_IMPORTANCE_MACRO_F1` GLOBAL_EXPLANATION_TOTAL	9,677,474.5	9,773,391.1	3	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C2	`SINGLE_FEATURE_ABLATION_LOCAL_NAIVE_REFERENCE` EXPLANATION_COMPUTE	1,394,145.6	1,407,445.8	5	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C2	`SINGLE_FEATURE_ABLATION_LOCAL_NAIVE_REFERENCE` EXPLANATION_COMPUTE	1,397,350.1	1,402,469.36	5	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C2	`SINGLE_FEATURE_ABLATION_LOCAL_NAIVE_REFERENCE` EXPLANATION_COMPUTE	1,416,216.1	1,459,796.38	5	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C2	`SINGLE_FEATURE_ABLATION_LOCAL_VECTORIZED` EXPLANATION_COMPUTE	22,372.35	23,183.51	600	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C2	`SINGLE_FEATURE_ABLATION_LOCAL_VECTORIZED` EXPLANATION_COMPUTE	22,410.7	23,240.1	450	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C2	`SINGLE_FEATURE_ABLATION_LOCAL_VECTORIZED` EXPLANATION_COMPUTE	22,494.55	23,550.32	540	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
GLOBAL	MODEL-C3	`PERMUTATION_IMPORTANCE_MACRO_F1` GLOBAL_EXPLANATION_TOTAL	3,319,257.5	3,335,536.07	3	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
GLOBAL	MODEL-C3	`PERMUTATION_IMPORTANCE_MACRO_F1` GLOBAL_EXPLANATION_TOTAL	3,329,822.6	3,349,876.94	3	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
GLOBAL	MODEL-C3	`PERMUTATION_IMPORTANCE_MACRO_F1` GLOBAL_EXPLANATION_TOTAL	3,261,422.6	3,283,771.31	3	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C3	`SINGLE_FEATURE_ABLATION_LOCAL_NAIVE_REFERENCE` EXPLANATION_COMPUTE	44,858.1	45,158.3	5	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C3	`SINGLE_FEATURE_ABLATION_LOCAL_NAIVE_REFERENCE` EXPLANATION_COMPUTE	43,169.4	44,206.26	5	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C3	`SINGLE_FEATURE_ABLATION_LOCAL_NAIVE_REFERENCE` EXPLANATION_COMPUTE	42,713.1	42,777.48	5	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C3	`SINGLE_FEATURE_ABLATION_LOCAL_VECTORIZED` EXPLANATION_COMPUTE	2,393.15	2,600.995	600	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C3	`SINGLE_FEATURE_ABLATION_LOCAL_VECTORIZED` EXPLANATION_COMPUTE	2,400.9	2,521.345	450	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C3	`SINGLE_FEATURE_ABLATION_LOCAL_VECTORIZED` EXPLANATION_COMPUTE	2,379	2,532.37	570	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
GLOBAL	MODEL-C4	`PERMUTATION_IMPORTANCE_MACRO_F1` GLOBAL_EXPLANATION_TOTAL	1,448,625.8	1,455,475.61	3	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
GLOBAL	MODEL-C4	`PERMUTATION_IMPORTANCE_MACRO_F1` GLOBAL_EXPLANATION_TOTAL	1,444,864.4	1,462,304.6	3	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
GLOBAL	MODEL-C4	`PERMUTATION_IMPORTANCE_MACRO_F1` GLOBAL_EXPLANATION_TOTAL	1,468,170.9	1,475,774.37	3	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C4	`SINGLE_FEATURE_ABLATION_LOCAL_NAIVE_REFERENCE` EXPLANATION_COMPUTE	41,420.5	41,897.76	5	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C4	`SINGLE_FEATURE_ABLATION_LOCAL_NAIVE_REFERENCE` EXPLANATION_COMPUTE	41,724.4	42,124.82	5	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C4	`SINGLE_FEATURE_ABLATION_LOCAL_NAIVE_REFERENCE` EXPLANATION_COMPUTE	40,435.6	40,753.04	5	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C4	`SINGLE_FEATURE_ABLATION_LOCAL_VECTORIZED` EXPLANATION_COMPUTE	793.25	871.305	600	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C4	`SINGLE_FEATURE_ABLATION_LOCAL_VECTORIZED` EXPLANATION_COMPUTE	795.5	896.32	570	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`
LOCAL	MODEL-C4	`SINGLE_FEATURE_ABLATION_LOCAL_VECTORIZED` EXPLANATION_COMPUTE	779.45	854.315	600	HOST-STAGE12-001	`EXP-XAI-LATENCY-001`

Global dataset procedures and local per-sample methods are intentionally separate; this is not a universal method leaderboard. Top-k reduction and serialization are separate phases.

Claim evaluation

Claim	Status	Evidence
`C-XAI-COST-01`	SUPPORTED	At least one matched local-method overhead ratio differs by >=2x: True.
`C-XAI-COST-02`	UNRESOLVED	Additional-call criterion=True; model-level Spearman rho=1.000, but positive bootstrap CI is not established with N=4 models.
`C-XAI-COST-03`	SUPPORTED	Intrinsic extraction is below permutation p05 for both applicable models: True.
`C-XAI-COST-04`	UNRESOLVED	Stage-10/11 method coverage is incomplete for a fully matched dominance test; no universal composite score was constructed.

Primary artifact: results/xai/stage12_raw_timings.csv. Full report: docs/experiments/STAGE12_XAI_LATENCY.md.