Embedded deployment

Target: nRF52840, Arm Cortex-M4F, 64,000,000 Hz. Gate GATE-EMBED-EXPORT-001 freezes architecture and equivalence criteria; it did not export a model.

Execution boundary

Item	Status
Protocol	`FROZEN`
Model export	`HOST_EXECUTED`
Quantization	`NOT_EXECUTED`
Compiled Flash	`NOT_MEASURED`
SRAM	`NOT_MEASURED`
MCU latency	`NOT_MEASURED`
Energy	`NOT_MEASURED`

Candidate tiers

Model	Confirmed type	Initial path	XAI path	Tier	Reason
MODEL-C1	LogisticRegression	`MANUAL_STATIC_PARAMETERS`	LOCAL_LINEAR_CONTRIBUTION	`TIER_A_EXPORT_FIRST`	Transparent compact linear path and validated inexpensive local contribution; still requires FP32 equivalence.
MODEL-C2	RandomForestClassifier	`GENERATED_TREE_CODE`	STATIC_GLOBAL_IMPORTANCE	`TIER_B_EXPORT_IF_RESOURCES_ALLOW`	Tree comparator retained for scientific evidence; analytical structure is larger and compile size unknown.
MODEL-C3	SVC	`GENERATED_SVM_CODE`	HOST_ONLY_XAI	`TIER_C_RESEARCH_COMPARATOR`	Prediction export remains research comparator; current XAI evidence is weak and RBF/probability verification is difficult.
MODEL-C4	MLPClassifier	`DIRECT_C_CPP`	NO_ON_DEVICE_XAI	`TIER_A_EXPORT_FIRST`	Static dense FP32 path is plausible; on-device XAI deferred because only repeated ablation is validated.

Stage 14 host-compiled FP32 equivalence

Host-compiled FP32 equivalence proves numerical portability of the implementation. It does not prove nRF52840 performance or resource feasibility.

Experiment EXP-EMBED-FP32-EQUIV-001 completed with outcome FAILED. Both candidates retained every golden and boundary decision but failed frozen preprocessing criteria.

Model	Status	Preprocess max abs	Preprocess max rel	Score max abs	Probability max abs	Golden	Boundary
MODEL-C1	`FAIL`	7.995e-6	2.744e-3	8.315e-5	1.510e-6	100%	100%
MODEL-C4	`FAIL`	4.845e-4	1.392e-5	3.596e-7	3.596e-7	100%	100%

C1 local-XAI equivalence: FAIL. Quantization and MCU deployment remain NOT_EXECUTED.

Stage 15 physical MCU port

Experiment EXP-MCU-C1-FP32-PORT-001: BLOCKED_HARDWARE. No supported physical board/debug probe was detected, so no board target was guessed and no build, flash, physical correctness, linked ROM, or linked static-RAM claim was produced.

Review the hardware blocker, planned connection, and evidence artifacts.

Stage 14R C1 explicit-preprocessing repair

Outcome: FAILED. The original failure reproduced, and none of the five prospectively frozen all-FP32 candidates passed every unchanged preprocessing criterion. No candidate was selected.

Candidate	Max abs	Max relative	Failed rows	Score max abs	Probability max abs	Golden	Boundary	Mandatory
`C1-PREPROC-P0-BASELINE`	7.995e-6	2.744e-3	41	8.315e-5	1.510e-6	100%	100%	`FAIL`
`C1-PREPROC-P1-RECIPROCAL`	7.995e-6	2.744e-3	41	8.315e-5	2.017e-6	100%	100%	`FAIL`
`C1-PREPROC-P2-AFFINE`	7.995e-6	1.558e-2	60	8.315e-5	2.017e-6	100%	100%	`FAIL`
`C1-PREPROC-P3-SPLIT_MEAN`	7.995e-6	5.453e-4	29	8.315e-5	2.017e-6	100%	100%	`FAIL`
`C1-PREPROC-P4-COMPENSATED`	7.995e-6	5.452e-4	29	8.315e-5	2.017e-6	100%	100%	`FAIL`

Conclusion: STRICT_FP32_EXPLICIT_STANDARDIZATION_REPAIR_NOT_DEMONSTRATED. The next proposal is a separately frozen fused preprocessing/inference experiment—not a retroactive repair and not quantization.

C1 fused preprocessing architecture gate

Gate GATE-C1-FUSED-PREPROC-001: FROZEN. Experiment EXP-EMBED-C1-FUSED-EQUIV-001 subsequently executed and PASSED every frozen criterion on the host.

The architecture folds the frozen scaler into raw-domain linear weights and biases. It deliberately does not materialize standardized features, so it does not retroactively repair Stage 14 or Stage 14R.

Architecture	Scaler subtractions	Scaler divisions	Classifier multiplications	Transformed buffer	Evidence
`EXPLICIT_PIPELINE`	128	128	768	128 elements	`DERIVED_ANALYTICAL`
`C1-FUSED-F0`	0	0	768	0 elements	`DERIVED_ANALYTICAL`

Fused local XAI is not included. Inference must pass first; a baseline-preserving explanation representation requires a separate experiment.

Golden maximum score error: 4.853e-5. Claims C-EMBED-C1-FUSED-01 and C-EMBED-C1-FUSED-02 are supported; fused XAI remains NOT_EXECUTED.

Fused local XAI subsequently PASSED across 54 Stage-09 audit samples. Maximum vector L1 error was 5.205e-6; C-EMBED-C1-FUSED-XAI-01 is supported for the host representation.

Analytical storage — derived estimate

These values are not measured Flash or SRAM. They count raw scalar-equivalent parameters and omit compiled code, runtime metadata, stack, heap, buffers, and instrumentation.

Model	FP32 parameter bytes	Preprocessing bytes	Evidence type
MODEL-C1	3,096	1,024	`DERIVED_ANALYTICAL`
MODEL-C2	409,200	1,024	`DERIVED_ANALYTICAL`
MODEL-C3	69,872	1,024	`DERIVED_ANALYTICAL`
MODEL-C4	42,136	1,024	`DERIVED_ANALYTICAL`