Getting a QHPC paper into CUCAI and QMIND next year

QHPC CPU-Only Research Curriculum with Publication Pathways

Compiled 2026-08-04. Curriculum pairs each research topic with a concrete publication pathway. All literature verified (see verified-bibliography-quality-matrix.md). Updated 2026-08-08: all four topics now at phase1–7 pipeline depth with compiled PDF/DOCX manuscripts (phase7-final-manuscript.md/tcpuN-phase7-final-manuscript.md). Volume II expanded to full lesson plans (volume-ii-curriculum.md). Remaining work is measurement fills (T-CPU1/T-CPU2 [MEASURED] slots, T-CPU3 GPU baseline) and venue submission.

Curriculum Overview

Four CPU-only research topics (T-CPU1…T-CPU4) detailed in qphc-cpu-four-topic-track.md, plus Volume II Tiers 4–6 in volume-ii-curriculum.md. This document maps topics to venues and timelines.

Publication Pathways

T-CPU2 (Edge vs. Server CPU Energy Crossover) — Flagship

• Status (2026-08-08): manuscript complete + compiled; registered-report protocol pre-registered; measurement pending.
• Primary venue: IEEE/ACM conference on energy-efficient computing (e.g., IGSC, e-Energy) or Systems for ML workshop (MLSys/NeurIPS)
• Timeline: 6–9 months from benchmark harness to submission
• Positioning: first documented energy-per-token crossover curve "Pi vs. server CPU" for modern LLMs
• Adjacent contribution: reproducible 28-model Raspberry Pi 4 benchmark extension

T-CPU1 (CPU-Only Toolchain Optimization)

• Status (2026-08-08): registered report complete (tcpu1-phase7-*); measurement pending.
• Primary venue: workshop paper — Systems and Machine Learning (SysML/MLSys workshop track) or LLM efficiency workshops
• Timeline: 3–6 months; demo-ready for CUCAI/QMIND
• Positioning: quantifiable toolchain comparison on constrained hardware

T-CPU3 (Portable CPU-Only RAG)

• Status (2026-08-08): code + evaluation complete with real results (tcpu3-phase7-*); GPU-baseline comparison pending cluster access.
• Primary venue: short paper / demo at CIKM, SIGIR (resource track), or NLP-OSS workshop
• Timeline: 4–6 months
• Positioning: RAGdb-style stack portability + quality-vs-GPU study

T-CPU4 (Green-Aware Scheduling)

• Status (2026-08-08): simulator + full-factorial sweep + real results complete (tcpu4-phase7-*).
• Primary venue: IEEE/ACM CCGrid, JSSPP workshop (IPDPS), or Future Generation Computer Systems (journal)
• Timeline: 6–12 months; simulation-first then optional hardware validation
• Positioning: CPU-only cluster green scheduling with RL; extends GAS-MARL/DRL-MOSHRS line

Publication Readiness Gates (all topics)

1. Verified literature base (done — audit report + quality matrix)
2. Falsifiable question stated with threshold (done — per-topic)
3. Reproducible harness + measurements (in progress per topic)
4. Failure-mode fallback documented (RAPL → CPU-utilization; 28-model → stratified sample)
5. Citation audit before any submission (process established 2026-08-04)

Gate Status (2026-08-08)

[STATUS: 2026-08-08] Historical pipeline artifact for T-CPU2 (writing/review process, 2026-08-06/07). Retained for provenance. Content frozen as written; superseded by phase7-final-manuscript.md (compiled PDF/DOCX) and PROJECT_STATUS.md.

Peer Review Report

Manuscript Information

• Title: Energy-per-Token Crossover Between Raspberry Pi 4-Class Edge and Server-CPU LLM Inference
• Manuscript ID: T-CPU2 (registered-report submission)
• Review Date: 2026-08-06
• Review Round: Round 1 (external editorial review)

Reviewer Information

Reviewer Role *

Editor-in-Chief (EIC)

Reviewer Identity *

Senior green-computing editor (IGSC), 15 years editorial experience in energy-aware systems research; advocate of pre-registration as an integrity mechanism in computer systems research.

Review Focus *

(1) Journal fit of the registered-report format; (2) defensibility of the "first crossover curve" originality claim; (3) whether the contribution survives the study's own honesty constraints (Outcome B scoped-negative publishability).

Overall Assessment *

Recommendation *

• Accept —” Can be published directly, only minor formatting changes needed
• Minor Revision —” Minor revisions needed, no re-review after revision
• Major Revision —” Substantial revisions needed, re-review required after revision
• Reject —” Not suitable for publication in this journal

Confidence Score *

4 —” Mostly within my area of expertise, high confidence

Summary Assessment *

This registered-report manuscript specifies a pre-registered protocol for the first energy-per-token crossover comparison between Raspberry Pi 4-class edge hardware and a server CPU node across a model-size by quantization grid. The paper is structurally complete (8/8 sections), cites 22 verified references, and adheres to a disciplined reporting framework in which all empirical result slots remain marked [MEASURED] pending protocol execution. From an editorial perspective, the manuscript is unusually honest: the protected hedge, the scoped negative (Outcome B) framing, and the explicit G2 measurement-validity disclosure all indicate the authors understand that the contribution must survive even a null result. The most significant editorial concerns are (a) whether the "first crossover curve" claim can be sustained given the preprint-concentrated related-work cluster, and (b) whether the venue's readership will accept a registered-report paper whose Results section is a plan rather than data. These are addressable in revision without re-review. I recommend Minor Revision.

Strengths *

S1: Registered-report framing is an integrity differentiator

The manuscript marks every empirical slot [MEASURED] and the verdict slot [CROSSOVER VERDICT], pre-committing to analysis decisions before data collection (Methods, Section 3.7; Results, Section 4). This is a genuine strength: it converts the paper's main risk —” results-driven analysis —” into an auditable pre-registration. Evidence Anchor: text: §3.7 "The criterion is pre-registered" and §6.1 "[CROSSOVER VERDICT] filled after measurement"

S2: Honest scoped-negative pathway

The paper explicitly frames Outcome B (no crossover found) as publishable contribution (Section 4.2; Discussion 5.1), which is a strength for a green-computing venue that should welcome actionable negative results. Evidence Anchor: text: §4.2 "the figure still reports the margin map, which is the scoped-negative contribution defined in the blueprint's Outcome B"

S3: Bounded claims discipline

Section 5.6 explicitly limits all claims to the measured hardware class, and Section 7.4 disclaims carbon-footprint interpretation. This prevents the overgeneralization that plagues many edge-computing submissions. Evidence Anchor: text: §5.6 "All claims in this paper are bounded to the measured hardware class"

S4: Complete package for review

The submission includes abstracts (EN), AI disclosure, cover letter, and a final quality checklist —” unusual completeness for a workshop/conference submission. Evidence Anchor: absence: N/A —” package completeness verified across phase7-final-manuscript.md, phase7-cover-letter.md, phase7-final-quality-checklist.md

Weaknesses *

W1: Originality claim rests partly on undated preprints

Problem: The "first crossover curve" claim is defended against a related-work cluster ([5]-[12], [15]-[19]) that is preprint-concentrated and bibliographically incomplete (Phase 5a flag). Editorial acceptance of the novelty claim therefore depends on references the paper itself cannot fully verify. Evidence Anchor: absence: References § —” expected complete bibliographic entries for [5]-[12],[15]-[19]; checked reference list and Phase 5a compliance report Why it matters: If any preprint in the cluster has already measured edge-vs-server energy, the headline novelty claim collapses. The author-side bibliographic completion is not cosmetic. Suggestion: Complete the missing fields before submission; add one sentence in Section 2.5 acknowledging the G5 temporal constraint explicitly conditions the novelty claim on search-date cutoff. Severity: Major Confidence: 4 —” core expertise: editorial judgment on novelty claims

W2: Results section is a plan, not data —” venue risk

Problem: The Results section (Section 4, ~520 words against a 1,300 budget) is structurally complete but data-empty by design. Some IGSC reviewers may read this as an incomplete submission rather than a registered-report feature. Evidence Anchor: text: §4.1 "[MEASURED —” energy-per-token per device per model-size tier and quantization level" Why it matters: Venue fit is the EIC's primary responsibility; if the registered-report format is not clearly signaled, the paper may be desk-rejected on completeness grounds before reaching review. Suggestion: Add an explicit "Registered Report" designation to the title page and abstract header, and a one-paragraph editor's note explaining the format and the timeline for result completion. Severity: Minor Confidence: 4 —” core expertise: editorial fit

W3: Single-author, single-device-class scope limits readership draw

Problem: The paper measures one Pi SKU and one server node. For a conference readership, the sample breadth may reduce the paper's attractiveness as a headline contribution. Evidence Anchor: text: §5.5 "the model grid is stratified, not exhaustive, and the Pi leg is a single SKU" Why it matters: Venue-level significance is part of the editorial decision; the paper's own limitations honestly acknowledge this. Suggestion: Strengthen Section 6.3 (multi-SKU future work) with a concrete planned device list; optionally add a short "broader impact" paragraph positioning the protocol as reusable infrastructure. Severity: Minor Confidence: 3 —” adjacent field: editorial scope judgment

Coverage Receipt (conditional *)

Covers: both —” populated lists above; no receipt required.

Detailed Comments *

Title & Abstract

• Title (11 words) is accurate and encodes the core claim; front-loads "energy-per-token" and "crossover," which matches the keyword set.
• Abstract is structured (Background/Objective/Method/Results/Conclusion), 258 words, within the 250-word target; Results component honestly marked [MEASURED].
• Requirement met: 5/5 components complete.

Introduction

• Research background sufficient; the "two trends meet in an unmeasured place" framing is effective.
• Research question is verbatim, falsifiable, and repeated in abstract/objective —” good pre-registration hygiene.

Literature Review / Theoretical Framework

• Gap synthesis G1-G5 is clear and actionable; each gap maps to a section hand-off.
• Concern: the preprint cluster ([5]-[12]) is bibliographically incomplete —” see W1.

Methodology / Research Design

• Protocol completeness is high post-revision; pre-registered crossover criterion now names test, factor, sign rule, and multiple-comparison control.

Results / Findings

• As a registered report, Section 4 is a plan with marked slots —” legitimate but a venue-communication risk (see W2).

Discussion

• Conditional framing (5.1) is disciplined; mechanism hypothesis (5.2) is clearly labeled as "expected to confirm or refute."
• Limitations (5.5) and Scope (5.6) are honest and comprehensive.

Conclusion

• Answer to RQ is deferred to the verdict slot —” consistent with registered-report integrity.
• Future work (6.3) is concrete and valuable.

References

• IEEE numeric style; first-appearance order verified; 22/22 cited.
• Incomplete bibliographic fields in the preprint cluster remain the author-side blocker (W1).

Questions for Authors *

1. Can you confirm that completing the bibliographic fields for references [5]-[12] and [15]-[19] will not change the novelty claim in Section 2.5 (i.e., none of these preprints measured edge-vs-server energy)?
2. Do you intend to submit the completed results within a declared timeline after acceptance of the registered report, and what is that timeline?
3. Would you be willing to add an explicit "Registered Report" designation to the title page to prevent desk-rejection on completeness grounds?

Minor Issues

Language / Grammar

• None observed; prose is clean and TEEL-structured.

Citation Format

• Reference [4] notes an intentionally unedited identifier typo; the note is transparent but unusual for a submission —” consider a cleaner resolution before final submission.

Figures and Tables

• No figures yet (by design); Section 4 figure slots (heatmaps, crossover curve) are declared with content specifications —” acceptable for registered report.

Layout

• LaTeX source uses IEEEtran conference class; abstract compiles via pdflatex —” document this in submission instructions.

Dimension Scores *

[STATUS: 2026-08-08] Historical pipeline artifact for T-CPU2 (writing/review process, 2026-08-06/07). Retained for provenance. Content frozen as written; superseded by phase7-final-manuscript.md (compiled PDF/DOCX) and PROJECT_STATUS.md.

Peer Review Report

Manuscript Information

• Title: Energy-per-Token Crossover Between Raspberry Pi 4-Class Edge and Server-CPU LLM Inference
• Manuscript ID: T-CPU2 (registered-report submission)
• Review Date: 2026-08-06
• Review Round: Round 1 (external editorial review)

Reviewer Information

Reviewer Role *

Peer Reviewer 1 (Methodology)

Reviewer Identity *

Power-measurement methodology researcher; specializes in RAPL/SoC energy instrumentation validity, statistical treatment of noisy performance measurements, and pre-registration practice in systems experiments.

Review Focus *

(1) Statistical validity of the pre-registered crossover criterion; (2) measurement validity of the two-channel Pi energy approach and the G2 RAPL-on-ARM gap; (3) reproducibility of the protocol.

Overall Assessment *

Recommendation *

• Minor Revision —” Minor revisions needed, no re-review after revision
• Major Revision —” Substantial revisions needed, re-review required after revision
• Reject —” Not suitable for publication in this journal

Confidence Score *

5 —” Completely within my area of expertise, very confident

Summary Assessment *

This paper specifies a pre-registered protocol for comparing energy-per-token between a Raspberry Pi 4-class device and a server CPU node across a model-size by quantization grid, using RAPL on the server and a two-channel (SoC instrumentation + CPU-utilization estimation) approach on the Pi. From a methodology standpoint the protocol is substantially complete: the crossover criterion now names the statistical test, fixed factor, sign-consistency rule, and multiple-comparison control; the repetition budget, calibration subset, and sampling interval are quantified; and the estimation fallback is disclosed wherever used. The most serious methodological issue is the estimation fallback itself —” a CPU-utilization-based model calibrated on a 10% subset is a weak substitute for direct power measurement, and the paper's framing of G2 as an honest "gap" does not fully address the risk that crossover decisions near the boundary will be driven by estimation error. A second concern is the validity of the paired t-test design given that the two legs are different hardware platforms measured under different instrumentation —” the pairing assumption deserves explicit justification. These are addressable in revision.

Strengths *

S1: Complete pre-registered crossover criterion

The criterion names the primary test (paired t-test), the fallback (Welch's with per-cell substitution record), the fixed factor (2x standard error of the difference), the sign-consistency rule (≥80% of repetition blocks), and the multiple-comparison control (Benjamini-Hochberg at alpha=0.05 across the declared grid). Evidence Anchor: text: §3.7 "The primary test is a paired t-test on per-run energy-per-token at each cell"

S2: Explicit measurement-validity disclosure for the G2 gap

The paper discloses that RAPL is unavailable on ARM and presents the estimation fallback alongside the primary channel, labeled wherever used —” this is the honest handling of a known instrumentation gap. Evidence Anchor: text: §3.4 "On the Pi leg, RAPL is not available; this is the known G2 validity gap"

S3: Quantified repetition and calibration design

The 30-run repetition budget is justified against the sample-budget knee in the autotuning literature, and the calibration subset is sized (10%, minimum 8 cells) with stratification across tiers and quantization levels. Evidence Anchor: text: §3.5 "fixed per cell in advance at 30 measurement runs"

S4: Randomized measurement ordering decouples thermal drift

Measurement ordering randomized across cells; the robustness section (4.4) is pre-committed to reporting drift and warm-up effects rather than smoothing them. Evidence Anchor: text: §3.5 "Measurement ordering is randomized across cells to decouple thermal drift from model effects"

S5: Robustness section pre-committed to reporting flips

Section 4.4 explicitly reports cells where the crossover decision flips under the confidence-interval criterion, rather than hiding instability —” good registered-report hygiene. Evidence Anchor: text: §4.4 "Cells where the crossover decision flips under the confidence-interval criterion are reported, not smoothed"

Weaknesses *

W1: Estimation fallback validity near the crossover boundary

Problem: The CPU-utilization-based estimation model, calibrated on a 10% subset of grid cells, is used as the secondary channel. Near the crossover boundary —” where margins are smallest —” the decision will be most sensitive to estimation error. The paper does not specify an estimation-error tolerance or a decision-rule for when the fallback channel's margin exceeds its calibration uncertainty. Evidence Anchor: text: §3.4 "a CPU-utilization-based estimation model, calibrated on the Pi with an external power meter on a calibration subset of 10% of grid cells" Why it matters: A crossover declared on a boundary where the two channels disagree is a measurement artifact risk, not a finding. The registered-report criterion does not yet gate on channel agreement. Suggestion: Add a pre-registered channel-agreement rule: (a) report calibration error (e.g., MAE/RMSE of estimation vs. primary on the calibration subset); (b) declare a cell "fallback-limited" when the two channels disagree by more than the calibration error; (c) exclude fallback-limited cells from crossover adjudication or flag them explicitly. Severity: Major Confidence: 5 —” core expertise: energy measurement methodology

W2: Paired t-test pairing assumption unstated

Problem: The crossover criterion uses a paired t-test on per-run energy-per-token. The pairing structure across two different hardware platforms (Pi vs. server) is not defined: what pairs with what? If "paired" refers to matched runs under identical prompts, the pairing is legitimate but should be stated; if the two devices have different variance, the Welch fallback threshold needs a concrete trigger. Evidence Anchor: text: §3.7 "The primary test is a paired t-test on per-run energy-per-token at each cell" Why it matters: Paired designs assume correlated paired observations; without stating the pairing unit (e.g., run i on Pi vs. run i on server under the same prompt), a reviewer cannot verify the design is appropriate. Suggestion: State the pairing unit explicitly (e.g., "each pair is run i on the Pi leg and run i on the server leg for the same prompt, quantization, and model"); define the variance-difference trigger for the Welch fallback (e.g., ratio of sample variances > 2). Severity: Major Confidence: 5 —” core expertise: statistical design

W3: Accuracy spot-check sample not quantified

Problem: Section 3.6 mentions accuracy spot-checks "on a sample of cells" to detect quantization-level quality collapse, but the sample size and selection rule are not pre-registered. Evidence Anchor: text: §3.6 "Accuracy spot-checks are run on a sample of cells" Why it matters: The crossover is only meaningful at "quality-acceptable" cells; if the accuracy spot-check subset is chosen post hoc, the quality-acceptability gate is not pre-registered. Suggestion: Pre-register the spot-check subset (e.g., same 10% stratified calibration subset) and the pass threshold (e.g., no accuracy collapse on any quantized model vs. fp16 baseline within a stated tolerance). Severity: Minor Confidence: 4 —” core expertise: measurement methodology

W4: No pre-registered statement of environment equivalence

Problem: The paper fixes governor and thermal state "per leg" (Section 3.1) but does not pre-commit to what equivalence between legs means —” e.g., does the server leg run at default governor while the Pi is pinned? Different governor policies on each leg could confound the energy comparison. Evidence Anchor: text: §3.1 "Hardware state (clock frequency governor, thermal state, attached peripherals) is fixed and documented per leg" Why it matters: The paper's own rationale (governor policy materially affects ARM energy) applies symmetrically to the x86 leg. Suggestion: Pre-register the governor policy per leg explicitly (e.g., performance governor on both legs vs. ondemand; state which), and record actual sustained clocks in the artifact. Severity: Minor Confidence: 4 —” core expertise: measurement methodology

Coverage Receipt (conditional *)

Covers: both —” populated lists above; no receipt required.

Detailed Comments *

Title & Abstract

• Abstract method description is accurate: two-channel Pi measurement + RAPL server, 4 quantization levels, pre-registered criterion.

Introduction

• G2 gap motivation is sound; the methodological gap framing is one of the paper's better arguments.

Literature Review / Theoretical Framework

• (Primarily Reviewer 2's focus; noted the [20] sample-budget citation correctly motivates the repetition design.)

Methodology / Research Design

• Design appropriateness: pre-registered protocol is the correct design for this question.
• Sampling strategy: model grid stratification is defensible (three tiers, selection rule stated).
• Data collection: two-channel energy measurement is the correct response to RAPL-on-ARM absence, but see W1 for boundary-validity risk.
• Analysis methods: crossover criterion is operational after revision; pairing assumption needs explicit statement (W2).

Results / Findings

• As a registered report, no data yet; the pre-committed robustness plan (4.4) is methodologically sound.

Discussion

• The power-envelope asymmetry mechanism (5.2) is clearly labeled as a hypothesis "the measured traces are expected to confirm or refute" —” appropriate.

Conclusion

• Verdict slot deferred; consistent with pre-registration.

References

• (Primarily Reviewer 2's focus; [20] Tørring is the correct anchor for sample-budget sensitivity.)

Questions for Authors *

1. What is the pairing unit for the paired t-test in Section 3.7, and what triggers the switch to the Welch's test?
2. What tolerance does the calibration subset establish for the estimation fallback, and how will "fallback-limited" cells be treated in crossover adjudication?
3. What accuracy spot-check subset and pass threshold are pre-registered for the quality-acceptability gate?

Minor Issues

Language / Grammar

• None observed.

Citation Format

• None observed.

Figures and Tables

• Section 4 figure slots are well-specified (heatmaps, crossover curve with confidence intervals on boundary).

Layout

• None observed.

Dimension Scores *

[STATUS: 2026-08-08] Historical pipeline artifact for T-CPU2 (writing/review process, 2026-08-06/07). Retained for provenance. Content frozen as written; superseded by phase7-final-manuscript.md (compiled PDF/DOCX) and PROJECT_STATUS.md.

Peer Review Report

Manuscript Information

• Title: Energy-per-Token Crossover Between Raspberry Pi 4-Class Edge and Server-CPU LLM Inference
• Manuscript ID: T-CPU2 (registered-report submission)
• Review Date: 2026-08-06
• Review Round: Round 1 (external editorial review)

Reviewer Information

Reviewer Role *

Peer Reviewer 2 (Domain)

Reviewer Identity *

LLM inference systems researcher with edge-deployment expertise; published on CPU-only inference stacks and quantized model deployment on constrained hardware; familiar with Husom's 28-model Pi sweep and the kernel-generation cluster.

Review Focus *

(1) Literature coverage and correct positioning against the edge/CPU-only LLM inference energy corpus; (2) adequacy of the model grid stratification; (3) genuineness of the domain contribution ("first crossover curve" claim).

Overall Assessment *

Recommendation *

• Minor Revision —” Minor revisions needed, no re-review after revision
• Major Revision —” Substantial revisions needed, re-review required after revision
• Reject —” Not suitable for publication in this journal

Confidence Score *

4 —” Mostly within my area of expertise, high confidence

Summary Assessment *

This paper proposes the first energy-per-token crossover comparison between Raspberry Pi 4-class edge and server-CPU LLM inference, anchored to the reproducible Husom et al. 28-model Pi 4 sweep protocol and positioned against the kernel-generation and green-HPC-scheduling literatures. From a domain perspective, the literature positioning is strong: Husom 2025 and RAGdb are correctly identified as establishing edge feasibility without cross-device energy comparison, the kernel-generation cluster is explicitly scoped as GPU-side context (preventing a common overclaim), and gap G4 (CPU-only/edge fleets absent from green-HPC scheduling) is a genuine and valuable observation. The model grid (three tiers, four quantization levels, explicit selection rule) is defensible for the registered-report stage. The main domain concern is that the stratification —” two representative models per tier —” may be too coarse to support the "crossover curve" framing, which implies a continuous boundary; and the novelty claim is formally conditioned on an incomplete preprint cluster. Both are addressable in revision.

Strengths *

S1: Correct corpus positioning

The paper correctly distinguishes edge feasibility literature (Husom 2025 establishes Pi feasibility; RAGdb establishes CPU-only operation) from cross-device energy comparison, which it identifies as unmeasured —” the G1 claim is credible given the corpus reviewed. Evidence Anchor: text: §2.1 "Husom et al. swept 28 models on a Raspberry Pi 4, establishing both feasibility and the cost envelope of Pi-class inference"

S2: Kernel-generation cluster honestly scoped as GPU-side context

Section 2.2 explicitly states the cluster provides no energy measurements for CPU/edge targets and that the paper "does not rely on it for any energy claim" —” preventing the common failure of importing GPU-side efficiency claims into an edge energy argument. Evidence Anchor: text: §2.2 "It does not provide energy measurements for the CPU or edge targets under study, and we do not rely on it for any energy claim"

S3: G4 scheduling-population gap is a genuine domain contribution

The observation that CPU-only clusters and edge fleets are absent from the green-HPC scheduling population, and that no scheduler consumes an edge-vs-server crossover curve, is an original and actionable gap analysis. Evidence Anchor: text: §2.3 "CPU-only clusters and edge device fleets are not among the scheduled populations, and no scheduling study consumes an edge-versus-server energy crossover curve as input"

S4: G5 temporal constraint disclosed

The paper discloses that the kernel-generation cluster is preprint-concentrated and largely undated, constraining literature-review claims —” an honest limitation that pre-empts a common reviewer objection. Evidence Anchor: text: §2.5 "G5 is temporal and venue-based: the kernel-generation cluster is preprint-concentrated and largely undated"

S5: Reproducibility anchoring to Husom 2025

Reproducing the published 28-model Ollama sweep protocol on the Pi leg is a strong choice that makes the edge comparison auditable against the literature baseline. Evidence Anchor: text: §3.2 "reproducing the runtime used in the 28-model sweep [3] and thereby keeping the edge leg comparable to the published baseline"

Weaknesses *

W1: Grid coarseness vs. "crossover curve" framing

Problem: The model grid selects two representative models per tier (six models total) across four quantization levels. The paper's headline artifact is a "crossover curve," which implies a continuous boundary in the model-size x quantization plane; a six-model grid produces, at best, a coarse operating-point map, not a curve in the usual sense. Evidence Anchor: text: §3.3 "each selected model runs four quantization levels: q4_0, q5_0, q8_0, and fp16" Why it matters: If the field reads "crossover curve" as a densely-sampled boundary, the six-model design may be seen as overclaiming; if the authors intend a coarse map, the terminology should be adjusted. Suggestion: Either (a) rename the contribution to "operating-point map" or "crossover boundary estimate" in the title/abstract, or (b) justify the two-per-tier selection rule as sufficient to bracket the boundary (e.g., selection targeting the tier floor and the model most likely to cross) and state the boundary is interpolated between grid points. Severity: Major Confidence: 4 —” core expertise: edge LLM inference domain

W2: Novelty claim conditioned on incomplete bibliography

Problem: The "first crossover curve" claim is formally defensible only against the search-dated corpus; the preprint cluster [5]-[12], [15]-[19] is bibliographically incomplete (Phase 5a flag), so the domain reviewer cannot independently confirm that none of these preprints has measured edge-vs-server energy. Evidence Anchor: absence: References —” expected complete bibliographic entries for [5]-[12],[15]-[19]; checked reference list and Phase 5a compliance report Why it matters: Domain novelty judgments require verifiable references; incomplete entries force the reviewer to accept the novelty claim on faith. Suggestion: Complete the missing bibliographic fields before submission; add the search-date and database list to Section 2.5 so the novelty claim is explicitly time-stamped. Severity: Major Confidence: 4 —” core expertise: domain novelty assessment

W3: No llm.cpp cross-check beyond cell count

Problem: The llama.cpp cross-check (10% of cells) validates runtime robustness but the paper does not state what outcome would be treated as a runtime-dependent fragility (e.g., if Ollama and llama.cpp disagree on a crossover cell). Evidence Anchor: text: §3.2 "A llama.cpp cross-check is executed on 10% of grid cells" Why it matters: The cross-check's purpose is stated ("a single-runtime result would be a fragility rather than a finding"), but the decision rule for what constitutes a fragility is not pre-registered. Suggestion: Pre-register the agreement criterion (e.g., crossover cell must replicate under llama.cpp to be reported as robust; otherwise flagged as runtime-dependent in Section 4.4). Severity: Minor Confidence: 3 —” adjacent field: benchmarking practice

Coverage Receipt (conditional *)

Covers: both —” populated lists above; no receipt required.

Detailed Comments *

Title & Abstract

• Title accurately encodes the domain contribution; "crossover curve" framing flagged in W1.

Introduction

• Motivation is persuasive: feasibility (Husom) + CPU-only operation (RAGdb) + energy concern (Grange) all converge on the unmeasured comparison.

Literature Review / Theoretical Framework

• Coverage is strong and correctly scoped: edge feasibility, kernel generation (as GPU-side context), green scheduling (G4 gap), benchmarking methodology.
• G1-G5 synthesis is one of the paper's best features —” each gap maps to a section hand-off.
• W2: incomplete preprint bibliography constrains the novelty claim.

Methodology / Research Design

• Model grid: three tiers (small <4B, mid 4-8B, large >8B within 8 GB fp16) with explicit selection rule —” defensible; coarseness concern in W1.
• Runtime: Ollama primary + llama.cpp cross-check —” sound reproducibility design; agreement rule missing (W3).

Results / Findings

• (Primarily Reviewer 1's focus; noted Section 4 structure-complete with [MEASURED] slots.)

Discussion

• G4 implications (5.3) correctly positioned as "giving green schedulers a quantitative input" —” an inference, not a data claim.
• Kernel-gen future work (5.4) correctly scoped as CPU-port vector.

Conclusion

• Future work list (6.3) is concrete and domain-relevant.

References

• Husom 2025 [3] correctly cited as the feasibility anchor; RAGdb [1] correctly cited as CPU-only operation anchor.
• [20] Tørring is the right citation for sample-budget sensitivity; [19] SOL-ExecBench correctly frames per-device benchmarking.

Questions for Authors *

1. Given the six-model grid, do you intend "crossover curve" to mean an interpolated boundary estimate or a densely-sampled curve —” and does the title need adjustment to match?
2. Can you confirm that none of the preprint cluster references ([5]-[12], [15]-[19]) has measured edge-vs-server energy, once the bibliographic fields are completed?
3. What pre-registered agreement criterion will determine whether a crossover cell is "runtime-robust" under the llama.cpp cross-check?

Minor Issues

Language / Grammar

• None observed.

Citation Format

• Reference [4] (P12) carries a documented unedited identifier typo; the transparency is appreciated but the entry should be finalized before submission.

Figures and Tables

• Section 4 figure specifications (heatmaps, crossover curve with boundary confidence intervals) are clear; no figures expected in a registered report.

Layout

• None observed.

Dimension Scores *

[STATUS: 2026-08-08] Historical pipeline artifact for T-CPU2 (writing/review process, 2026-08-06/07). Retained for provenance. Content frozen as written; superseded by phase7-final-manuscript.md (compiled PDF/DOCX) and PROJECT_STATUS.md.

Peer Review Report

Manuscript Information

• Title: Energy-per-Token Crossover Between Raspberry Pi 4-Class Edge and Server-CPU LLM Inference
• Manuscript ID: T-CPU2 (registered-report submission)
• Review Date: 2026-08-06
• Review Round: Round 1 (external editorial review)

Reviewer Information

Reviewer Role *

Peer Reviewer 3 (Perspective)

Reviewer Identity *

Sustainable/green computing researcher specializing in HPC energy-aware scheduling; has published on scheduler energy-consumption models and workload-aware power management; evaluates papers for operational actionability of energy claims.

Review Focus *

(1) Actionability of the proposed crossover curve for green schedulers; (2) honesty of the single-SKU/single-node scope relative to the scheduling claims; (3) whether out-of-scope scheduler experiments are being demanded or correctly deferred.

Overall Assessment *

Recommendation *

• Minor Revision —” Minor revisions needed, no re-review after revision
• Major Revision —” Substantial revisions needed, re-review required after revision
• Reject —” Not suitable for publication in this journal

Confidence Score *

4 —” Mostly within my area of expertise, high confidence

Summary Assessment *

From the green-scheduling perspective, this paper's value is the proposal to produce a quantitative input (an edge-vs-server energy-per-token crossover curve) that schedulers currently lack —” G4 is a real gap and §5.3 correctly positions the future scheduling use. The scope honesty is commendable: the paper repeatedly confines its claims to a single SKU and single node, and it correctly does not promise a scheduler. My main concerns are that (a) the transfer from a per-token energy measurement to a scheduler-consumable decision rule requires stated operational assumptions (workload composition, idle draw, thermal drift, fleet heterogeneity) that the paper does not yet make explicit; (b) the single-SKU/single-node honesty, while good, may undersell or oversell depending on how the §5.3 claim is phrased. These are framing and assumption-explicitness issues, not design flaws.

Strengths *

S1: G4 gap is real and correctly framed

The observation that CPU-only clusters and edge fleets are absent from green-scheduler populations, and that no scheduler consumes an edge-vs-server crossover curve, is an original, field-relevant gap. Evidence Anchor: text: §2.3 "CPU-only clusters and edge device fleets are not among the scheduled populations, and no scheduling study consumes an edge-versus-server energy crossover curve as input"

S2: §5.3 correctly scopes the scheduling implication

The discussion explicitly frames the scheduling use as an inference ("offered as an inference, not a data claim"), giving schedulers a quantitative input rather than claiming a scheduler result. Evidence Anchor: text: §5.3 "giving green schedulers a quantitative input to such cost models"

S3: Single-SKU/single-node scope honestly declared

The paper repeatedly limits its reach (Pi 4-class SKU, single node, no device-fleet claims), pre-empting the classic overgeneralization objection. Evidence Anchor: text: §3.1 "single node, one SKU (Pi 4B 4 GB)"

S4: Energy model aligned with scheduler-consumable units

Energy-per-token, with the §3.6 breakdown into model/device/idle terms, maps to the cost-model inputs schedulers actually consume —” a good unit choice for the stated audience. Evidence Anchor: text: §3.6 "energy-per-token is decomposed into model-execution and baseline/idle contributions"

S5: Cross-check and sensitivity design supports operational confidence

The llama.cpp cross-check and the sample-budget sensitivity analysis (100 ms sampling, 30 runs) give schedulers a bounded error estimate rather than a point figure. Evidence Anchor: text: §4.4 "sensitivity of the reported energy-per-token to sampling interval and measurement-run count"

Weaknesses *

W1: Missing operational assumptions for scheduler transfer

Problem: The paper does not yet state the workload and operating assumptions under which a scheduler could consume the crossover curve: what fraction of tokens the batch is (batch-bound vs latency-bound), whether device idle draw is included in the crossover operating point, and what fleet heterogeneity is assumed (single-SKU only). Evidence Anchor: absence: §5.3 —” expected operational transfer conditions; checked §3.6 and §5.3 Why it matters: A scheduler integrating this curve needs to know the decision boundary assumes (e.g.) saturated-batch inference and a stable device temperature; otherwise the crossover operating point is misread as universal. Suggestion: Add to §5.3 the explicit operating conditions under which the crossover curve transfers to scheduling decisions, and state that fleet generalization (multi-SKU, multi-node) is out of scope. Severity: Major Confidence: 4 —” core expertise: green scheduling

W2: Scope-honesty phrasing may cut both ways

Problem: The paper correctly limits itself to a single SKU, but §5.3's scheduler implication implicitly invites fleet-level generalization. The reviewer needs the boundary between "measurement claim" (single SKU) and "implication" (fleet) stated in one sentence. Evidence Anchor: text: §5.3 "edging toward fleet-level device management" vs text: §3.1 "single node, one SKU" Why it matters: Without an explicit boundary sentence, the paper could be read as either underselling (referees discount the contribution) or overselling (readers generalize beyond the SKU). Suggestion: One explicit sentence in §5.3: the measured claim is single-SKU; the fleet implication is a hypothesis to be tested, not a measured result. Severity: Minor Confidence: 4 —” core expertise: scope-vs-implication framing

W3: No stated policy or operational recommendations

Problem: For a green-computing audience, the paper does not yet state what a scheduler should do with the curve (e.g., offload small-model tiers to edge when energy-per-token is lower, keep large tiers on server). This is an absence of a recommendations paragraph, not a design flaw. Evidence Anchor: absence: §5.3 —” expected at least one operational recommendation; checked §5.3 Why it matters: Green-computing venues value a crisp "so what" for practitioners; without it, the contribution reads as measurement-only. Suggestion: Add one conditional recommendation sentence, explicitly flagged as contingent on the measured crossover and on the operational assumptions in W1. Severity: Minor Confidence: 3 —” adjacent field: operational green computing

Coverage Receipt (conditional *)

Covers: both —” populated lists above; no receipt required.

Detailed Comments *

Title & Abstract

• Title is scope-honest (single device class vs server CPU).

Introduction

• Energy cost framing (Grange) + edge feasibility (Husom) + CPU-only operation (RAGdb) converges cleanly on the scheduling-adjacent motivation.

Literature Review / Theoretical Framework

• G4 (scheduling-population gap) is the strongest section; correctly distinguishes scheduler populations from the proposed crossover-curve input.

Methodology / Research Design

• Single-SKU, single-node design is appropriately scoped for a first measurement; the paper does not over-promise fleet results.
• Energy decomposition (§3.6) is scheduler-consumable.

Results / Findings

• (Reviewer 1's focus; noted Section 4 is structure-complete with [MEASURED] slots.)

Discussion

• §5.3 is the section to watch: W1 and W3 both land here. The "inference, not data claim" framing is good; it needs operational assumptions + one recommendation.

Conclusion

• Future work (6.3) correctly defers fleet and scheduler-integration studies.

References

• [13] Caux Datazero2 and [14] Koslovski correctly frame energy-aware HPC; [20]/[21] correctly frame measurement robustness.

Questions for Authors *

1. Under what workload and operating assumptions (batch saturation, idle draw inclusion, temperature stability) should a scheduler treat the crossover operating point as a decision boundary?
2. Can you add a one-sentence boundary in §5.3 distinguishing the measured single-SKU claim from the fleet-level implication?
3. What operational recommendation, if any, do you intend to derive from the curve (conditional on the measured result)?

Minor Issues

Language / Grammar

• None observed.

Citation Format

• None observed (per-device benchmarking framing in [19] is correctly cited).

Figures and Tables

• Crossover-curve figure spec (with boundary confidence intervals) is the right presentation for a scheduler-consumable input.

Layout

• None observed.

Dimension Scores *

[STATUS: 2026-08-08] Historical pipeline artifact for T-CPU2 (writing/review process, 2026-08-06/07). Retained for provenance. Content frozen as written; superseded by phase7-final-manuscript.md (compiled PDF/DOCX) and PROJECT_STATUS.md.

Peer Review Report

Manuscript Information

• Title: Energy-per-Token Crossover Between Raspberry Pi 4-Class Edge and Server-CPU LLM Inference
• Manuscript ID: T-CPU2 (registered-report submission)
• Review Date: 2026-08-06
• Review Round: Round 1 (external editorial review)

Reviewer Information

Reviewer Role *

Devil's Advocate (Adversarial)

Reviewer Identity *

Methodologically adversarial reviewer whose mandate is to stress-test the paper's central claims under the worst plausible reading: is the crossover finding real, or a grid/device/quantization artifact? Is the design capable of producing a negative result? Would the paper change any practitioner decision?

Review Focus *

(1) Novelty-vs-artifact: is "first crossover curve" a property of the hardware or of the grid/quantization choices? (2) Cherry-picking gate: is there a registered criterion that would force reporting an unfavorable result? (3) Actionability: does the paper change a "so what" decision? (4) Falsifiability: is §5.2's power-envelope reasoning circular?

Overall Assessment *

Recommendation *

• Minor Revision —” Minor revisions needed, no re-review after revision (tension noted below; the design is falsifiable, so minor is justified —” with the caveat that the novelty claim and the power-envelope framing must be tightened)

Confidence Score *

5 —” Fully within my area of expertise, high confidence

Summary Assessment *

The paper is unusually honest for the genre: it is a registered report with falsifiable questions, [MEASURED] slots, a conditional-discussion structure, and an explicit statement that the crossover is claimed as device-class-specific, not universal. The adversarial reading still surfaces four pressure points. First, the "crossover curve" could be an artifact of where the grid was placed and which quantization levels were chosen, not a property of the devices; the design must state how grid placement could manufacture a spurious crossover. Second, the cherry-picking gate is implicit rather than registered: nothing in §3 pre-commits the authors to report a crossover cell that contradicts the tier-trend hypothesis. Third, the "so what" is under-argued for a claims-heavy finding. Fourth, §5.2's power-envelope reasoning risks circularity (device class chosen because it is power-constrained, then energy advantage attributed to that constraint). None of these is fatal; all are fixable in revision.

Strengths *

S1: Genuinely falsifiable design

Four falsifiable questions with explicit [MEASURED] slots and conditional sections —” the single strongest safeguard against post-hoc rationalization. The paper would survive a null crossover. Evidence Anchor: text: §1.2 "F1—“F4 falsifiable questions with pre-specified decision rules"

S2: Device-class-scoped claim

The abstract and §5.1 explicitly limit the finding to "device-class-specific," pre-empting the universal-generalization reading. Evidence Anchor: text: §5.1 "crossover is claimed as device-class-specific, not universal"

S3: Surprising-cell rule exists

§4.5 pre-specifies a surprising-cell definition (largest absolute margin with direction contradicting tier-median sign), which is a genuine anti-cherry-picking mechanism. Evidence Anchor: text: §4.5 "surprising cell = largest absolute margin with direction contradicting the tier-median sign, per leg"

S4: Quantization axis honestly included

By testing q4_0/q5_0/q8_0/fp16 across all grid cells, the design acknowledges that the crossover may be quantization-dependent —” an honest treatment of the primary artifact risk. Evidence Anchor: text: §3.3 "each selected model runs four quantization levels: q4_0, q5_0, q8_0, and fp16"

S5: Negative-result capacity

The conditional structure (§5.3 "if the crossover is not confirmed, edge offload is not supported") makes the paper publishable on a null —” a rare and valuable property. Evidence Anchor: text: §5.3 "if the crossover is not confirmed —¦ offload is not supported"

Weaknesses *

W1: Grid-placement artifact risk on the headline claim

Problem: The crossover boundary's location depends on which models/quantization levels the grid contains. A six-model grid placed to bracket the boundary could manufacture a crossover that a denser grid would dissolve (or a sparser one would miss). The paper does not state what grid-configuration dependence would be reported as a limitation. Evidence Anchor: text: §3.3 "two representative models per tier" and §4.4 "sensitivity —¦ to sampling interval and measurement-run count" —” grid-configuration sensitivity is not in the §4.4 sensitivity list. Why it matters: The headline claim is "crossover curve"; the adversarial question is whether the curve is a property of the devices or of the chosen sample points. Suggestion: Add grid-configuration sensitivity to §4.4: report whether the crossover cells persist under (a) the alternative selection rule stated in §3.3, and (b) one additional model per tier if the alternative rule selects a different model. Severity: Major Confidence: 5 —” adversarial core

W2: Cherry-picking gate implicit, not registered

Problem: The surprising-cell rule (§4.5) exists, but nothing in §3 pre-commits the authors to report a crossover cell that contradicts the tier-trend hypothesis —” the failure mode of the paper would be suppressing the null and amplifying the crossover. Evidence Anchor: absence: §3 —” expected a pre-registered "any crossover cell that contradicts the tier-trend hypothesis must be reported with equal prominence"; §4.5 defines surprising cells but §3 does not mandate their reporting Why it matters: In a registered report, the gate must be in the protocol, not the results section. Suggestion: Move/add to §3: "Crossover cells whose direction contradicts the tier-trend hypothesis shall be reported with equal prominence to confirming cells, and the failure mode 'grid manufactured the crossover' shall be discussed if the alternative-selection sensitivity (W1) dissolves it." Severity: Major Confidence: 5 —” adversarial core

W3: "So what" under-argued for a claims-heavy finding

Problem: The paper demonstrates the gap and the method but does not crisply argue what a practitioner does differently if the crossover is confirmed vs refuted (beyond the §5.3 conditional). The energy-per-token difference must clear a decision-relevant threshold —” a 2% crossover is publishable but scheduler-irrelevant. Evidence Anchor: absence: §1.3/§5 —” expected a decision-relevant effect-size threshold; checked §1.3 and §5 Why it matters: The adversarial test of any empirical claim is whether it changes a decision at some effect size; the paper does not state the minimum crossover margin that would be operationally meaningful. Suggestion: Add to §1.3 or §5: the minimum crossover margin (e.g., ≥10% energy-per-token difference at the operating point) that would be claimed as operationally meaningful, and discuss whether the sampling error (§3.4) can resolve that margin. Severity: Minor Confidence: 4 —” adjacent: decision-oriented adversarial review

W4: §5.2 power-envelope circularity risk

Problem: The device class is selected because it is power-constrained; the paper then attributes the edge energy advantage (when confirmed) to the power envelope. That is partly circular: the constrained device wins the energy game partly because it was chosen for constraint. Evidence Anchor: text: §5.2 "the device class is selected for its power-constrained design; the energy advantage, when confirmed, is attributed to that constraint" Why it matters: If the edge advantage is partly an artifact of device selection, the "sustainability" reading of the finding is weaker than the paper's framing implies. Suggestion: Add one sentence distinguishing the descriptive claim (constrained device has lower energy-per-token at these operating points) from the causal claim (the constraint causes the advantage), and flag the causal reading as hypothesis. Severity: Minor Confidence: 4 —” adversarial reasoning

Coverage Receipt (conditional *)

Covers: both —” populated lists above; no receipt required.

Detailed Comments *

Title & Abstract

• "Crossover" is the right headline if the grid is dense enough to support it; see W1.

Introduction

• Motivation is sound but the "so what" threshold (W3) belongs in §1.3.

Literature Review / Theoretical Framework

• (Covered by Peer Reviewers 2 and 3; no independent finding.)

Methodology / Research Design

• Falsifiability is the design's strength (S1). The gap is the pre-registration of the cherry-picking gate (W2).

Results / Findings

• The surprising-cell rule (§4.5) is a genuine anti-cherry-picking mechanism; it belongs in the protocol (§3) to be a registered gate.

Discussion

• §5.2 circularity (W4) and the conditional "offload not supported" structure (S5) both live here; both are fixable with one sentence each.

Conclusion

• Future work (6.3) is concrete; add grid-density robustness there as a numbered item.

References

• (Covered by Peer Reviewer 2; no independent finding.)

Questions for Authors *

1. What grid-configuration dependence would you report as a limitation, and does §4.4's sensitivity list include grid placement?
2. Will the cherry-picking gate (crossover cells contradicting the tier trend reported with equal prominence) be registered in §3 before data collection?
3. What is the minimum energy-per-token crossover margin that you would claim as operationally meaningful, and can the sampling error resolve it?
4. How do you distinguish the descriptive claim (constrained device has lower energy-per-token) from the causal claim (the constraint causes it) in §5.2?

Minor Issues

Language / Grammar

• None observed.

Citation Format

• None observed.

Figures and Tables

• Crossover-curve figure must be accompanied by the grid-density caveat (W1) in the caption or §4.4.

Layout

• None observed.

Dimension Scores *

[STATUS: 2026-08-08] Historical pipeline artifact for T-CPU2 (writing/review process, 2026-08-06/07). Retained for provenance. Content frozen as written; superseded by phase7-final-manuscript.md (compiled PDF/DOCX) and PROJECT_STATUS.md.

Editorial Decision and Revision Roadmap

Manuscript Information

• Title: Energy-per-Token Crossover Between Raspberry Pi 4-Class Edge and Server-CPU LLM Inference
• Manuscript ID: T-CPU2 (registered-report submission)
• Decision Date: 2026-08-06
• Review Round: Round 1 (external editorial review —” 5 independent reviewers)

Reviewer Corpus

Corpus mean weighted score: 72.2 Corpus recommendation: Minor Revision (5/5 independent reviewers)

Editorial Decision

Decision *

Minor Revision —” the paper is conditionally accepted pending resolution of the consolidated revision items below. No re-review after revision; the associate editor will verify resolution against the roadmap.

Decision Rationale *

• All five reviewers independently returned Minor Revision with no Reject and no Major Revision —” an unusual consensus, driven by the registered-report format's built-in falsifiability and the paper's disciplined scoping of claims.
• The five independent weakness lists converge on three consolidated Major items (novelty-claim bibliographic conditioning; crossover-boundary validity under grid/fallback choices; protocol-level pre-registration gaps) and six Minor items (all framing/assumption-explicitness).
• No reviewer identified a design flaw requiring restart: the falsifiable-questions structure, [MEASURED] slots, conditional discussion, and single-SKU scope honesty were uniformly praised (8 of 10 consolidated strengths touch these).
• The Devil's Advocate explicitly notes the paper "would survive a null crossover" and "is publishable on a null" —” the strongest possible endorsement for the registered-report genre.

Editor's Note on Scoring *

The corpus mean (72.2) is depressed by the Evidence Sufficiency dimension (50/100 across all reviewers), which scores zero data by design in a registered report. The editor treats this as expected: for this submission type, evidence sufficiency is assessed on the planned measurement protocol's ability to deliver decisive evidence, which all reviewers rated positively (rigor 70-78).

Consolidated Revision Roadmap

MAJOR ITEMS (must address before resubmission)

M1. Novelty-claim bibliographic conditioning

Source: R-EIC W1; R2 W2 Problem: The "first crossover curve" claim is formally defensible only against a search-dated corpus; the preprint cluster [5]-[12],[15]-[19] is bibliographically incomplete, so the novelty claim cannot be independently verified. Action:

• Complete missing bibliographic fields for [5]-[12],[15]-[19] (venue, year, identifiers) before submission (Phase 5a Major-2 carry).
• Add the search date, database list (IEEE Xplore, ACM DL, Scopus, Google Scholar, WoS) and coverage window (2016-2026) to §2.5 so the novelty claim is explicitly time-stamped.
• Reconfirm, against the completed references, that no cluster paper measures edge-vs-server energy.

M2. Crossover-boundary validity (grid and estimation-fallback)

Source: R1 W1; R2 W1; R-DA W1 Problem: Three independent channels agree the crossover boundary's location depends on (a) grid placement/density and (b) the estimation-fallback policy when power data are incomplete near the boundary. Without stated controls, the headline "crossover curve" could be a grid/device/quantization artifact. Action:

• Add grid-configuration sensitivity to §4.4: report whether crossover cells persist under the alternative model-selection rule (§3.3) and, if the alternative rule selects a different model, one additional model per tier.
• Pre-register the estimation-fallback validity rule (R1 W1): (i) channel-agreement criterion for including fallback-estimated cells, (ii) calibration error bound, (iii) fallback-limited-cell exclusion rule near the boundary.
• Rename or reframe "crossover curve" to "crossover boundary estimate" or "operating-point map" in title/abstract, with interpolation between grid points stated explicitly (R2 W1).
• Add the grid-density caveat to the crossover-curve figure caption or §4.4 (R-DA minor issue).

M3. Protocol-level pre-registration gaps

Source: R1 W2; R-DA W2; R1 W3; R1 W4 Problem: Four protocol decisions exist only in the results/section structure, not in the pre-registration (§3): the paired t-test pairing unit + Welch trigger, the accuracy spot-check subset/threshold, per-leg governor equivalence, and the cherry-picking gate. Action:

• §3.7: state the pairing unit for the paired t-test (per-cell × per-leg) and the exact condition that triggers the Welch's fallback (R1 W2).
• §3.5: pre-register the accuracy spot-check subset (which cells) and pass threshold (R1 W3).
• §3.1: pre-commit per-leg governor equivalence (both legs use the default ondemand/schedutil governor, or state the difference) (R1 W4).
• §3: add the cherry-picking gate —” crossover cells whose direction contradicts the tier-trend hypothesis shall be reported with equal prominence; the "grid manufactured the crossover" failure mode shall be discussed if the M2 alternative-selection sensitivity dissolves it (R-DA W2).

MINOR ITEMS (address in same revision)

m1. Registered-report designation explicit

Source: R-EIC W2 Action: Add explicit "Registered Report" designation to the title page/author block so the format is not mistaken for a results paper (avoids desk-rejection misclassification).

m2. Operational assumptions for scheduler transfer

Source: R3 W1 Action: §5.3 —” state the workload/operating conditions under which the crossover curve transfers to scheduling decisions (batch saturation, idle-draw inclusion, temperature stability); fleet generalization stated as out of scope.

m3. Scope-boundary sentence

Source: R3 W2 Action: One sentence in §5.3: the measured claim is single-SKU; the fleet implication is a hypothesis to be tested, not a measured result.

m4. One operational recommendation

Source: R3 W3 Action: Add one conditional recommendation (contingent on measured crossover + m2 assumptions), e.g., offload small-model tiers to edge when energy-per-token is lower at the operating point.

m5. Decision-relevant effect-size threshold

Source: R-DA W3 Action: §1.3 or §5 —” state the minimum crossover margin claimed as operationally meaningful (e.g., ≥10% energy-per-token difference) and whether §3.4 sampling error can resolve it.

m6. §5.2 power-envelope circularity

Source: R-DA W4 Action: One sentence distinguishing the descriptive claim (constrained device has lower energy-per-token at these operating points) from the causal claim (constraint causes the advantage); flag causal reading as hypothesis.

m7. llama.cpp cross-check agreement criterion

Source: R2 W3 Action: Pre-register the agreement rule (crossover cell must replicate under llama.cpp to be reported as runtime-robust; otherwise flagged in §4.4).

Author-Side Carry (merged)

• Carry A (M1): Complete [5]-[12],[15]-[19] and [13]/[19] venue/identifier fields (Phase 5a Major-2; also fixes R2 citation-format note on [4] P12 typo at submission time).
• Carry B: Author block completion (affiliations, ORCID).
• Carry C: DOI/Zenodo persistent identifier slot (§7.3) activation at submission.
• Carry D: Fill [MEASURED] slots after execution (by design, post-review).

Reviewer Questions Consolidated for Authors

1. Under what workload and operating assumptions should a scheduler treat the crossover operating point as a decision boundary? (R3)
2. What is the minimum energy-per-token crossover margin that would be claimed as operationally meaningful, and can the sampling error resolve it? (R-DA)
3. What grid-configuration dependence would you report as a limitation? (R-DA)
4. Can you confirm, against the completed preprint cluster references, that none measures edge-vs-server energy? (R2)
5. What pre-registered agreement criterion will determine a "runtime-robust" crossover cell under the llama.cpp cross-check? (R2)

Roadmap Status

• Phase 1: 5 independent reviews (EIC + R1 Methodology + R2 Domain + R3 Perspective + R-DA)
• Phase 2: Synthesis —” decision (Minor Revision) + consolidated roadmap (3 Major, 7 Minor) + author-side carry
• Phase 3: Apply M1-M3 + m1-m7 to phase4-complete-draft.md (registered-report protocol edits) and regenerate Phase 7 package
• Phase 4: Re-verify against quality gates (0 em dashes, no AI terms, TEEL, CARS, IEEE numeric)
• Final: Author-side carry A-D at submission

[STATUS: 2026-08-08] Historical pipeline artifact for T-CPU2 (writing/review process, 2026-08-06/07). Retained for provenance. Content frozen as written; superseded by phase7-final-manuscript.md (compiled PDF/DOCX) and PROJECT_STATUS.md.

External Editorial Revision Log

Date: 2026-08-06 Round: External editorial review (5-reviewer corpus), Phase 2 synthesis decision: Minor Revision Source: external-review-editorial-decision.md (M1-M3, m1-m7) Target: phase4-complete-draft.md (registered-report protocol)

Applied Edits (14)

Metadata Updates

• Revision Round: 0 -> 1 (external editorial revision: M1-M3, m1-m7 applied)
• Total word count: 6,040 -> 6,905 (+15.1%); §3 +25.5%, §5 +31.2%, trim at Phase 5

Quality Gates Re-verified

• Em dashes in new text: 0 (all insertions use commas/colons/semicolons)
• No flagged AI terms introduced
• Protected hedge (#548) untouched; §1 prose unchanged
• All insertions TEEL-conformant, declarative, pre-registration phrasing

Not Applied (author-side carry, post-review)

• M1 bibliographic field completion: [5]-[12],[15]-[19], [13]/[19] venues (Carry A, Phase 5a Major-2) - cannot fabricate metadata

Carry A Resolution (2026-08-07, appended)

Carry A (preprint-cluster bibliographic completion) executed via websearch + Scholar Sidekick (Crossref/DataCite):

• Completed and verified against resolving records: refs [5]-[12], [15]-[19] (authors, years, arXiv IDs / DOIs). Applied to phase7-references.bib, phase7-final-manuscript.md, phase4-complete-draft.md, phase7-manuscript.tex (BibTeX keys unchanged; .bib enriched only, no new citations).
• µCUTLASS: resolved as separate paper (arXiv:2603.29010, Hari et al. 2026); added as metadata-only companion to [19] (micutlass key, not cited) rather than a new numbered reference, preserving IEEE first-appearance numbering.
• [13] identity correction (author decision delegated, integrity-first): the manuscript in-text claim is "negotiation-based power and IT co-optimization" but the entry was labeled "S. Caux et al." Verified records show the negotiation paper is Thi et al., "Negotiation game for joint IT and energy management in green datacenters," FGCS vol. 110, pp. 1116-1138, 2020, doi: 10.1016/j.future.2019.11.018 (Crossref-verified); Caux et al. Euro-Par 2018 (doi: 10.1007/978-3-319-96983-1_24) is RECO IT-scheduling, not negotiation. Ref [13] corrected to Thi et al. in all three manuscript files + .bib; Caux Euro-Par 2018 recorded in the bib note for provenance. All 22 numbered references now complete; zero remaining open items.
• §2.5 text updated in all three manuscript files: "bibliographic completion is a submission requirement" -> "cluster completed and verified on 7 Aug 2026"; G5 reworded from "temporal and venue-based, largely undated" to "venue-based, preprint-concentrated".
• [4] P12 identifier: left unedited; flag #4 in phase5a-citation-compliance-report.md resolved as NOP at Carry A (re-verification #558/#559 confirmed correct 2603.20224; prior typo claim withdrawn).
• Carry A closure —” audit-driven bibliographic corrections + [4] recast (2026-08-07): full-bibliography audit (auditBibliography, 23 claims) surfaced real errors beyond ref [13]; all corrected against resolving records and propagated to .bib, all three manuscripts, and tracking docs:
  • [1] RAGdb: title "Portable CPU-only retrieval-augmented generation database" was a paraphrase; verified record Khalid, A. B., "RAGdb: A Zero-Dependency, Embeddable Architecture for Multimodal Retrieval-Augmented Generation on the Edge," arXiv:2602.22217, 2025.
  • [2] Grange: claimed DOI 10.1016/j.future.2018.01.004 resolved to an unrelated HIB-tree paper; corrected to 10.1016/j.future.2018.03.049 (FGCS vol. 86, pp. 99-120, 2018).
  • [3] Husom: carried arXiv:2603.20224 (the P12 record); verified Husom record is arXiv:2504.03360, "Sustainable LLM Inference for Edge AI: Evaluating Quantized LLMs for Energy Efficiency, Output Accuracy, and Inference Latency" (Husom, Goknil, Astekin, Shar, Kåsen, Sen, Mithassel, Soylu).
  • [4] P12 recast: verified record at 2603.20224 is Wilhelm, P., Wittkopp, T., Kao, O., "Beyond Test-Time Compute Strategies: Advocating Energy-per-Token in LLM Inference" (2026), NOT "LLM-based CUDA kernel generation". Manuscript body previously cited [4] as the kernel-gen cluster anchor (misattribution: real ID, unsupported claims). Kernel-gen anchor moved [4] -> [5] (CuTeGen) in Sections 1, 2.2, 5.4 of all three manuscripts + outline + blueprint + corpus + phase1; [4] re-cited accurately for energy-per-token metric advocacy in Sections 1 and 3.6 (on-thesis for T-CPU2).
  • [14] Koslovski: corrected to "DAG-based workflows scheduling using Actor-Critic Deep Reinforcement Learning," FGCS 2024, doi: 10.1016/j.future.2023.09.018 (claimed 10.1016/j.future.2022.11.001 wrong).
  • [20] Tørring: corrected to "Analyzing Search Techniques for Autotuning Image-based GPU Kernels: The Impact of Sample Sizes," IPDPSW 2022, doi: 10.1109/ipdpsw55747.2022.00155 (Tørring, J. O., Elster, A. C.).
  • [21] Bolet: corrected to "An Exploration of Global Optimization Strategies for Autotuning OpenMP-based Codes," IPDPSW 2024, doi: 10.1109/ipdpsw63119.2024.00138.
  • [22] Alasandagutti: corrected DOI to 10.1109/hipc66333.2025.00013 (32nd IEEE HiPC 2025; claimed 10.1109/HiPC56045.2025.00012 wrong).
  • Refs [5]-[12] kernel-gen cluster (CuTeGen, KernelFoundry, QiMeng-Kernel, Kernel-Smith, R³, SwizzlePerf, CUDA-LLM, KForge) re-verified matched; no changes. Corpus-wide 0 retractions; [22] carries an erratum flag only. Phase 5a flag #4 now resolved as NOP-and-superseded; new flag #7 documents the [4] recast.
• Carry A propagation to planning/corpus docs (2026-08-07): stale "Caux (Datazero2)" / "Caux" labels synced to Thi et al. 2020 in phase1-literature-search-report.md (corpus entry #4 + matrix + Discussion list, with identity-correction provenance note), phase2-paper-outline.md (§2.3, §5, Evidence Map row), phase3-argument-blueprint.md (Discussion evidence), volume-ii-curriculum.md (Tier 6 core source). external-review-perspective-report.md:121 kept as-is (historical review snapshot, per phase6 precedent); grep confirms no stale labels remain in living docs.
• Post-closure re-verification (2026-08-07): remaining open identifiers resolved —” [7] QiMeng-Kernel (arXiv:2511.20100, DataCite confirms AAAI 2026 acceptance), [14] Koslovski (Crossref matched; bib enriched with full authors Koslovski/Pereira/Albuquerque + vol. 150, pp. 354-363), [15] GAS-MARL (Crossref matched), [16] Adaptive MARL (verified against official SCA/HPCAsia 2026 poster P-212, https://www.sc-asia.org/2026/data/poster/post212.pdf; URL added to bib), [17] DeepHCM (Crossref matched), [18] DRL-MOSHRS (Crossref matched), [19] SOL-ExecBench (DataCite matched). All 23 corpus entries verified post-closure; zero mismatches. citation-audit-report.md P12 rows annotated (Aug-4 report's "An Energy-Per-Token Evaluation of Reasoning SLMs" paraphrase corrected against abs-page title; verification of authors/date/ID stands).
• Final batch audit (2026-08-07, scholar-sidekick auditBibliography): full-bib audit returned 22/23 matched (high confidence), 0 mismatch, 0 ambiguous, 0 errored, 0 retracted. Adaptive MARL (index 16) not_found in Crossref/PubMed/OpenAlex, expected for a venue-only poster (no DOI/arXiv) —” verified at closure against official poster P-212 PDF; bib note carries the URL. No fabricated-title collisions corpus-wide.
• Carry B-D: author block, DOI/Zenodo slot activation, [MEASURED] filling after execution.

Next

• Carry A applied in place to all three manuscript files + .bib; no regeneration pending. All 22 numbered references complete and matching resolving records (audit-driven corrections at Carry A closure). Remaining author-side: Carry B-D (author block, DOI/Zenodo, [MEASURED] filling after execution).

Citation Audit Report

Scope: artifact-1.html … artifact-7.html in C:\Users\[Author]\Documents\August_2_2026_QHPC_Research_topics\ Method: Each citation verified independently (title + identifier + quoted content) against authoritative sources: arXiv abs pages, DOI resolution (Crossref/IEEE Xplore landing pages), publisher sites, official project/documentation sites. Nothing accepted on the basis of the artifact's own self-report. Verdicts: verified · likely-valid-paywalled · unverifiable · mismatched Audit date: 2026-08-04 Follow-up (2026-08-08): T-CPU1/T-CPU3/T-CPU4 manuscripts cited ONLY from the verified matrix (verified-bibliography-quality-matrix.md); per-manuscript citation audits in tcpuN-phase5a-citation-compliance-report.md (T-CPU4: 11 verified; T-CPU3: 4 verified + MiniLM prose-only; T-CPU1: 12 verified). Zero fabricated or retracted references in the new manuscripts.

Summary

¹ artifact-5's self-reported P-table claim of "zero mismatches, zero fabrications" is reconciled in the Reconciliation note below.

² Counting semantics: "17/17" = the 17 embedded Starter Pack references (external references — no local file), all verified. "14/14" = consensus.app referent-level validation — a verification count, not a raw link count (grep found 12 consensus.app link occurrences, artifact-4 only; artifact-5 has zero).

artifact-1.html — Citation-free

No citations, references, or source links beyond frame boilerplate. Nothing to verify.

artifact-2.html — Citation-free

No citations, references, or source links beyond frame boilerplate. Nothing to verify.

artifact-3.html — Citation-free

No citations, references, or source links beyond frame boilerplate. Nothing to verify.

artifact-4.html — References verified (1 mismatch corrected)

Exclusions (present in the artifact's "unresolved" self-report but NOT actual references of artifact-4):

• 10.1109/TDEL.2017.2729953
• 10.4991-…F (truncated/placeholder)
• 10.1023/A:1003381911194
• 10.5555/3291656.329197

Notes

• Kavianpour #11 was a genuine mismatch: the DOI resolved to a different paper (MacArthur et al.). Corrected in the artifact — citation rewritten to MacArthur et al. 2017, DOI unchanged. See Resolutions.
• HPL-MxP ICL SC'23 paper at https://www.netlib.org/utk/people/JackDongarra/PAPERS/146_2003_the-hpl-mxp-paper.pdf — large PDF (>5 MB), fetch fails on size not 404; paper itself is a real ICL technical report.
• The "OSU Micro-Benchmarks" line at artifact-4:656 is a task instruction, not a citation — non-issue.
• OpenMP 5.2 specification confirmed at openmp.org.

artifact-5.html — All 9 references verified

The artifact's own P-table (line 703) claims "zero mismatches, zero fabrications." This was independently checked per citation — all 9 pass.

Notes

• P12 ID: re-grep of artifact-5.html confirms the anchor cites the correct arXiv:2603.20224 (lines 351, 352, 715); zero occurrences of 2606. An earlier audit draft claimed a 2606 typo; that claim was itself erroneous and is superseded (see FLAG 2). All other fields (title, authors Wilhelm/Wittkopp/Kao, 4 Mar 2026, cs.CL, SLM + CoT/Majority Voting, MMLU) verified against the abs page.
• Embedded "Starter Pack" (17/17) verified: HPCG, HPX, Chapel, CUDA, SLURM, LibF, E3SM, SU2, GATK, Kurth (arXiv:1810.01993), energy research (arXiv:2408.01552), Frontier (OSTI), HIP, Lustre whitepaper, etc. (17/17 = the 17 embedded Starter Pack references — external AI Artifact referents, no local file; all verified. See footnote ².)

artifact-6.html — Citation-free

No citations, references, or source links beyond frame boilerplate. Nothing to verify.

artifact-7.html — Citation-free

No citations, references, or source links beyond frame boilerplate. Nothing to verify.

Related verification (outside the 7 artifacts)

• Standalone "Starter Pack" references (separate file): all verified — HPCG, HPX, Chapel, CUDA, SLURM, LibF, E3SM, SU2, GATK, Kurth (arXiv:1810.01993), energy (arXiv:2408.01552).
• consensus.app validation: 14/14 matches (referent-level verification count, not raw link count — 12 link occurrences in artifact-4, 0 in artifact-5; see footnote ²).

Actionable flags — resolved

1. artifact-4, ref #11 (Kavianpour) — mismatched. The DOI 10.1109/COMST.2017.2746083 resolved to MacArthur et al. 2017 (InfiniBand/Verbs/MPI tutorial), not the cited Kavianpour COMST article. RESOLVED: citation rewritten in artifact-4.html to MacArthur et al. 2017; DOI unchanged.
2. artifact-5, ref P12 — (superseded) earlier audit draft claimed a wrong arXiv ID (2606.20224). RESOLVED (NOP): re-grep confirms artifact-5 carries the correct 2603.20224 at all anchor sites (lines 351, 352, 715); zero 2606 occurrences; no in-file correction required.

Resolutions

FLAG 1 — artifact-4 ref #11 (RESOLVED)

• Decision 1A (direct edit): rewrote the citation in artifact-4.html to MacArthur, P., Russell, R., et al. (2017). "An Integrated Tutorial on InfiniBand, Verbs, and MPI." IEEE COMST 19(4):2894–2926.
• DOI link unchanged: https://doi.org/10.1109/COMST.2017.2746083 (verified).
• Applied at artifact-4.html lines 2488–2489; confirmed by re-read.

FLAG 2 — artifact-5 ref P12 (SUPERSEDED — erroneous flag, no action needed)

• Decision 2A (NOP): no in-file correction required; artifact-5.html unchanged.
• Verification: re-grep of artifact-5.html finds 2603.20224 at lines 351, 352, 715; zero occurrences of 2606.20224. The earlier FLAG 2 claim of a 2606 typo was a transcription error in the audit draft, not in the artifact.
• Paper metadata (Wilhelm, Wittkopp, Kao, arXiv:2603.20224, 4 Mar 2026, cs.CL) verified against the abs page. Annotation (Carry A closure, 2026-08-07): exact abs-page title is "Beyond Test-Time Compute Strategies: Advocating Energy-per-Token in LLM Inference"; the "An Energy-Per-Token Evaluation of Reasoning SLMs" variant used in this report was a paraphrase recorded in error.

Reconciliation note

Footnote ¹ (line 22) records the artifact-5 P-table self-claim: "zero mismatches, zero fabrications." This section reconciles that claim against the independent audit.

• Claim substantiated for the 9 verified references (P6–P14): every cited title, author set, and claim was independently confirmed against arXiv abs pages and official sources. No fabricated, predatory, or non-existent papers found in artifact-5.
• Caveat — identifier typo (FLAG 2, superseded): an earlier audit draft reported P12 carried the wrong arXiv ID (2606.20224). Re-verification (fresh grep) shows that report was erroneous: artifact-5 cites the correct 2603.20224. No identifier correction was needed; artifact-5.html untouched.
• Cross-artifact fix (FLAG 1): the corpus's single mismatched citation (artifact-4 ref #11, Kavianpour → MacArthur et al. 2017) was corrected in source at artifact-4.html lines 2488–2489. Outside artifact-5's scope but recorded here for unambiguous corpus status.
• Supporting checks: Starter Pack references 17/17 verified; consensus.app validation 14/14 matches (referent-level verification count, not raw link count — see footnote ²).

Bottom line: the artifact-5 self-claim of "zero mismatches, zero fabrications" holds under independent audit, qualified by one cross-artifact citation fix applied to artifact-4.html (FLAG 1). The earlier P12 identifier-typo flag (FLAG 2) was a draft error and is withdrawn.

Verified HPC Bibliography & Quality Matrix

Compiled 2026-08-04 from citation-audited sources across 7 research artifacts. Every entry below was verified against its resolving record (DOI/arXiv) via Scholar Sidekick. Fabricated/predatory entries removed during audit; corrections documented in citation-audit-report.md.

Updated 2026-08-08: now also the sole citation source for the T-CPU1/T-CPU3/T-CPU4 manuscripts (tcpuN-phase7-references.bib). Per-topic citation subsets: T-CPU4 → rows 4–9, 18–23; T-CPU3 → rows 1, 3, 4, 10; T-CPU1 → rows 1, 8–17, 20. MiniLM-L6-v2 is described in prose in T-CPU3 (no matrix entry — not added to avoid unverified attribution).

Quality Matrix Legend

• V = verified (identifier resolves, title matches)
• OA = open access status
• R = retraction/correction status (all clean unless noted)

Bibliography (Verified)

Audit Summary

• 7 artifacts audited (artifact-1 through artifact-7)
• 2 corrections applied:
  1. artifact-4 #11: Kavianpour et al. (2017) InfiniBand → MacArthur et al. (2017), DOI 10.1109/COMST.2017.2746083 (lines 2488–2489)
  2. artifact-5 P12: prior audit draft claimed an arXiv ID typo; re-verification showed the claim was erroneous and the ID (2603.20224) is correct. No correction applied; FLAG 2 withdrawn per audit reconciliation.
• Fabricated/predatory citations removed from research drafts during audit
• All remaining entries verified: identifier resolves, title matches resolved record
• Carry A enrichment (2026-08-07): preprint-cluster entries (rows 11-23) completed with authors, years, and resolving identifiers (arXiv/Crossref/DataCite); µCUTLASS recorded as companion to SOL-ExecBench. Ref [13] identity corrected: negotiation-based co-optimization claim maps to Thi et al. 2020 (FGCS, DOI 10.1016/j.future.2019.11.018); Caux et al. Euro-Par 2018 (DOI 10.1007/978-3-319-96983-1_24) retained for provenance only.
• Carry A closure (2026-08-07) audit-driven corrections: rows 1 (Husom ID/title/authors), 3 (RAGdb ID/title), 4 (Grange DOI), 6 (Koslovski title/DOI), 7 (Alasandagutti DOI), 8 (Tørring title/authors/DOI), 9 (Bolet title/DOI), 10 (P12 title/authors — verified record is the energy-per-token study; kernel-gen anchor moved to CuTeGen). All 22 numbered entries now complete and matching resolving records.
• Post-closure re-verification (2026-08-07): remaining open identifiers re-checked via Crossref/DataCite/arXiv abs — refs [7] QiMeng-Kernel (arXiv:2511.20100; AAAI 2026 confirmed on DataCite), [14] Koslovski (matched; bib authors expanded to Koslovski/Pereira/Albuquerque, vol. 150, pp. 354-363 added), [15] GAS-MARL (matched), [17] DeepHCM (matched), [18] DRL-MOSHRS (matched), [19] SOL-ExecBench (matched). Ref [16] Adaptive MARL verified against official SCA/HPCAsia 2026 poster P-212 (https://www.sc-asia.org/2026/data/poster/post212.pdf). Zero mismatches across all 23 entries post-closure.
• Final batch audit (2026-08-07, scholar-sidekick auditBibliography): 23 claims (title + identifier) submitted against full bib. Result: 22 matched (high confidence), 0 mismatch, 0 ambiguous, 0 errored, 0 retracted/corrected/concern. Index 16 (Adaptive MARL) not_found in Crossref/PubMed/OpenAlex — expected: venue-only poster with no DOI/arXiv; verified at closure via official SCA/HPCAsia 2026 poster P-212 PDF. No fabricated-title collisions detected.

Quality Matrix

Verified HPC Education Reference Corpus (17 Sources)

Delivered 2026-08-04 for QHPC club curriculum. All 17 sources citation-verified (identifiers resolve, titles match). No retractions/corrections flagged. Updated 2026-08-08: corpus now grounds all four research topics plus Volume II full lesson plans (Tiers 4–6 in volume-ii-curriculum.md). Source 2 (RAGdb) used by T-CPU3; sources 13–17 used by T-CPU4 and Tier 6; sources 1, 17 used by Tier 4; sources 3–12 used by Tier 5. MiniLM-L6-v2 (T-CPU3 embedder) deliberately omitted — no verified matrix entry exists; described in prose instead.

Corpus

A. LLM Inference on Edge/CPU Hardware

1. Husom, E. J., Goknil, A., Astekin, M., Shar, L. K., Kåsen, A., Sen, S., Mithassel, A. K., Soylu, A. (2025). "Sustainable LLM Inference for Edge AI: Evaluating Quantized LLMs for Energy Efficiency, Output Accuracy, and Inference Latency." arXiv:2504.03360. — Foundational benchmark; baseline for T-CPU2 reproduction.
2. RAGdb (Khalid, A. B., 2025). "RAGdb: A Zero-Dependency, Embeddable Architecture for Multimodal Retrieval-Augmented Generation on the Edge." arXiv:2602.22217. — Reference for CPU-only RAG stacks (T-CPU3).

B. CPU-Only Kernel Generation & Autotuning

3. Wilhelm, P., Wittkopp, T., Kao, O. (2026). "Beyond Test-Time Compute Strategies: Advocating Energy-per-Token in LLM Inference." arXiv:2603.20224. — Energy-per-token metric advocacy for LLM inference; kernel-gen anchor is CuTeGen (below).
4. CuTeGen. "LLM kernel generation via CuTe abstraction layer with staged optimization." — Abstraction-layer generation strategy; kernel-gen anchor.
5. KernelFoundry. "Hardware-aware evolutionary kernel optimization using MAP-Elites and meta-prompt co-evolution." — Evolutionary + LLM hybrid.
6. QiMeng-Kernel MTMC. "Hierarchical decoupling of optimization strategy from implementation for kernel generation." — Strategy/implementation separation.
7. Kernel-Smith. "Evolution-oriented post-training for stable iterative kernel optimization." — Post-training stability.
8. Record-Remix-Replay (R³). "Hierarchical GPU kernel optimization spanning source implementation, compiler passes, and launch config." — Cross-layer optimization.
9. SwizzlePerf. "Hardware-aware LLM optimization of GPU kernel spatial locality via workgroup remapping." — Locality optimization.
10. SOL-ExecBench and µCUTLASS. "Benchmarking GPU kernels against hardware Speed-of-Light with domain-specific language guidance." — Benchmark methodology.
11. CUDA-LLM FSR. "Feature Search and Reinforcement framework for LLM-generated CUDA kernel optimization." — Search+RL framework.
12. KForge. "Cross-platform LLM kernel generation via dual-agent refinement and profiling interpretation." — Cross-platform generation.

C. Scheduling, Green HPC & Autotuning Studies

13. GAS-MARL. "Green-aware HPC scheduling via multi-action deep reinforcement learning with proactive delay decisions." — Green scheduling RL.
14. Adaptive MARL for Real-Time GPU Job Scheduling. "Two-agent architecture for heterogeneous GPU clusters." — Heterogeneous scheduling.
15. DeepHCM. "Deep RL for job scheduling on load-aware heterogeneous clusters with state aggregation." — State aggregation RL.
16. DRL-MOSHRS. "Multi-objective scheduling for heterogeneous HPC systems via D3QN with priority-based elastic orchestration." — Multi-objective D3QN.
17. Tørring, J. O., Elster, A. C. (IPDPSW 2022). "Analyzing Search Techniques for Autotuning Image-based GPU Kernels: The Impact of Sample Sizes." doi: 10.1109/ipdpsw55747.2022.00155. — Autotuning methodology (sample-budget dependence).

Teaching Use Map