Local-first research evidence map

Edge Intelligence Reproducibility Map

Which reported gains in edge intelligence can be independently reconstructed with available code, data, hardware, and result provenance?

PASS

0 errors and 0 review warnings. The source set is content-addressed below.

sha256:f81c893af76059a40fd6b942fbe289c87f5a39b2df7099343d94594334bbbfe6

5Papers mapped

10Claims anchored

63.7Mean coverage / 100

12Replication tasks

6Current blockers

Replication candidate triage

A rule-based queue combines evidence gaps, required resources, task dependencies, and remaining effort. No hidden composite score.

Rank	Candidate	Decision	Coverage	Effort	Unresolved resources	Next action
1	ChannelCraft: A Procedural Benchmark for Robust Channel Estimation `channelcraft-2023`	complete	100.0	0h	—	Archive the reproduced outputs and record claim-level outcomes.
2	ThermalTiny: Calibrated TinyML under Dynamic Thermal Budgets `thermaltiny-2025`	evidence first	68.2	20h	thermal-board	Resolve Custom thermal-control evaluation board (partial).
3	EdgeFormer: Event-Driven Transformers for On-Device RF Classification `edgeformer-2025`	evidence first	75.8	28h	rf-spectra-v2	Resolve RF Spectra v2 dataset (partial).
4	VoltGuard: Uncertainty-Aware Fault Detection in Converter Telemetry `voltguard-2024`	evidence first	36.4	2h	converter-telemetry, voltguard-model	Resolve Converter telemetry traces (unknown).
5	SparseBeam: Low-Overhead Beam Selection with Structured Priors `sparsebeam-2024`	evidence first	37.9	24h	mmwave-traces, sparsebeam-code	Resolve Urban mmWave channel traces (unavailable).

What can actually be rebuilt?

A weighted summary of documented reconstructability. The score is a navigation aid, never a verdict on research quality.

Assessment coverage

63.7 / 100 mean

5 of 5 papers have explicit cards.

Scores measure documented reconstructability. They do not judge correctness, importance, novelty, statistical validity, or research integrity.

Recurring evidence gaps

data

environment

compute

code

Evidence matrix

Every cell comes from an evidence locator, not a model-generated guess.

Paper	Year	Score	Method specificity	Data availability	Code availability	Environment capture	Metric definition	Baseline traceability	Compute disclosure	Result traceability
ChannelCraft: A Procedural Benchmark for Robust Channel Estimation `channelcraft-2023`	2023	100.0	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
EdgeFormer: Event-Driven Transformers for On-Device RF Classification `edgeformer-2025`	2025	75.8	Yes	Partial	Yes	Yes	Yes	Partial	Partial	Partial
SparseBeam: Low-Overhead Beam Selection with Structured Priors `sparsebeam-2024`	2024	37.9	Partial	No	Partial	?	Yes	Partial	?	Partial
ThermalTiny: Calibrated TinyML under Dynamic Thermal Budgets `thermaltiny-2025`	2025	68.2	Yes	Partial	Partial	Partial	Yes	Yes	Partial	Partial
VoltGuard: Uncertainty-Aware Fault Detection in Converter Telemetry `voltguard-2024`	2024	36.4	Partial	?	No	?	Yes	Yes	No	Partial

YES = sufficiently documentedPARTIAL = usable with assumptions NO = explicitly unavailable? = not yet established

Claim anchors

Statements are linked to page, figure, table, appendix, or repository evidence supplied by the reviewer.

empirical

The method claims lower pilot overhead at the stated recall target.

SparseBeam: Low-Overhead Beam Selection with Structured Priors

Abstract and §5.4 (synthetic)

uncertain

empirical

Structured priors reportedly improve top-3 beam recall in dense layouts.

SparseBeam: Low-Overhead Beam Selection with Structured Priors

Results §5, Table 1 (synthetic)

reported

empirical

Estimator rankings change across procedural channel families.

ChannelCraft: A Procedural Benchmark for Robust Channel Estimation

Table 3 and artifact results.csv (synthetic)

corroborated

empirical

Ten released seeds reproduce the reported aggregate within tolerance.

ChannelCraft: A Procedural Benchmark for Robust Channel Estimation

Appendix C and result bundle (synthetic)

corroborated

empirical

Event gating reportedly lowers median inference energy on the target board.

EdgeFormer: Event-Driven Transformers for On-Device RF Classification

Results §4.3, Figure 5 (synthetic)

uncertain

empirical

EdgeFormer reports higher macro F1 than the compact convolutional baseline.

EdgeFormer: Event-Driven Transformers for On-Device RF Classification

Results §4.2, Table 2 (synthetic)

reported

empirical

Temperature-aware recalibration reportedly reduces expected calibration error.

ThermalTiny: Calibrated TinyML under Dynamic Thermal Budgets

Figure 6 and Appendix B (synthetic)

reported

limitation

The calibration benefit degrades outside the board's characterized range.

ThermalTiny: Calibrated TinyML under Dynamic Thermal Budgets

Limitations §7 (synthetic)

corroborated

empirical

VoltGuard reports higher in-distribution AUROC than a deterministic network.

VoltGuard: Uncertainty-Aware Fault Detection in Converter Telemetry

Results §4.1, Table 2 (synthetic)

reported

empirical

Selective prediction reportedly preserves precision under load shift.

VoltGuard: Uncertainty-Aware Fault Detection in Converter Telemetry

Results §4.4, Figure 7 (synthetic)

uncertain

Replication execution plan

Dependency waves expose what can run in parallel and which missing artifact blocks later work.

Wave	Task	State	Priority	Estimate	Dependencies
1	Freeze paper and supplement versions `freeze-sources`	done	critical	1.5h	—
2	Resolve VoltGuard telemetry availability `request-volt-data`	blocked	critical	2h	freeze-sources
2	Verify code and data reuse conditions `verify-licenses`	done	critical	2h	freeze-sources
3	Resolve RF Spectra raw-capture access `acquire-rf-data`	blocked	critical	3h	verify-licenses
3	Rebuild ChannelCraft environment `rebuild-channel-env`	done	high	4h	verify-licenses
3	Reconstruct ThermalTiny evaluation board `rebuild-thermal-board`	ready	high	12h	verify-licenses
3	Design an open trace substitute for SparseBeam `replace-mmwave-traces`	ready	high	10h	verify-licenses
4	Rerun ChannelCraft released seeds `rerun-channel`	done	high	6h	rebuild-channel-env
4	Train EdgeFormer across five seeds `train-edgeformer`	blocked	high	18h	acquire-rf-data
4	Reconstruct missing SparseBeam training path `implement-sparsebeam`	ready	medium	14h	replace-mmwave-traces
4	Run the thermal calibration sweep `run-thermal-sweep`	ready	medium	8h	rebuild-thermal-board
5	Profile EdgeFormer board energy `profile-edge-energy`	blocked	high	7h	train-edgeformer

Audit boundary

Machine checks validate structure and references. They cannot establish truth, fairness, statistical validity, or author intent.

Structural result

PASS

55 artifacts · 45 graph nodes · 66 graph edges · 56 sealed files.

Actionable backlog

22 partial, missing, or unknown evidence items remain.

Open the JSON source to see the exact evidence, next action, and human-entered decision behind every cell.