Reviewer3 (AI review)

October 2, 2026

AI Review (reviewer3.com)

Comments (6)

1. Incorrect section reference for deployment trade-off — Editorial

ity. The trade-off, discussed in §7.4, is that the web tier and the voice agent can now fail independently: the website can be fully up while no conversation is possible.

Section 3.2 states that the trade-off where the web tier and voice agent can fail independently is discussed in §7.4. However, Section 7.4 ('Reliability Evidence') discusses the automated test suite and CI workflow. The independent failure trade-off is actually discussed in the unnumbered 'Limitations' section under 'Single-host availability'.

2. Incorrect section reference for retrieval layer — Editorial

The 'Ethical Considerations' section references §7.4 for the domains explicitly covered by the retrieval layer. However, Section 7.4 ('Reliability Evidence') is about the automated test suite. The retrieval layer and its coverage are described in Section 3.3 ('Knowledge Retrieval: Six MCP Servers') and evaluated in Section 7.3 ('Retrieval Grounding').

3. Incomplete latency measurements — Minor

mean latency 519.7ms. Estimated end-to-end la- tency from speech end to first audio response is 0.9–1.7s under typical conditions – competitive with commercial assistants despite running the ASR/TTS stages on operator-controlled rather than hyperscale infrastructure.

The authors claim that the system achieves an estimated end-to-end voice latency of 0.9–1.7s. However, they only report empirical measurements for ASR mean latency (235.5ms) and LLM first-token mean latency (519.7ms). They do not provide any empirical measurements for TTS generation latency (the time required for Fasih-TTS-V1 to synthesize and return the first audio chunk) or network transmission overhead. Because TTS synthesis is a major contributor to voice agent latency, omitting this measurement makes the end-to-end latency claim speculative. The authors should measure and report the latency of the TTS component and the overall network round-trip time to substantiate the end-to-end latency figures.

4. No empirical validation of observability stack — Major

ing. We address this with a heartbeat that travels outward from the agent on a fixed interval to a judge that sits outside both halves of the system; staleness alone reveals an outage, with no crash- detection logic required. A public health endpoint

The authors propose a three-layer observability design to detect when the GPU-bound agent host goes silent. However, the paper does not provide any empirical data or analysis demonstrating the effectiveness of this design in production, such as the number of actual or simulated outages detected, false positive/negative rates, or the latency of detection. Without empirical validation, it is unclear whether the proposed heartbeat and health endpoint design successfully resolves the targeted failure mode in practice. The authors should include evaluation data from simulated failure injections or production logs to demonstrate the reliability of the observability stack.

5. Vague operationalization of retrieval success — Minor

On 30 held-out Islamic questions, retrieval suc- ceeds for 100% of queries, and the retrieval- augmented condition adds only ∼ 50ms mean la- tency over answering from parametric memory alone (786ms vs. 736ms) – confirming that ground- ing, which is the paper’s central integrity claim, is not purchased at a meaningful latency cost.

The authors assert that retrieval succeeds for 100% of queries on 30 held-out Islamic questions. However, the paper does not define what constitutes "retrieval success" (e.g., whether it means successful tool execution, retrieval of relevant context, or accurate grounding of the final response). Furthermore, there is no description of the 30 held-out questions, how they were curated, or which of the six MCP servers they targeted. This lack of detail makes it impossible to evaluate the rigor or validity of the 100% success rate. The authors should clearly define the criteria for retrieval success and provide details on the evaluation dataset and the specific tools triggered.

6. Missing evaluation of Muslim-6B-PRO — Major

Muslim-6B-PRO’s model card does not yet report quantified benchmark results; we describe its architecture and training honestly rather than assert an accuracy figure it does not have.

The authors claim to release a family of fine-tuned Arabic Islamic model artifacts, specifically highlighting Muslim-6B-PRO as an efficient tool-routing LLM. However, the paper contains no evaluation, benchmark results, or performance metrics for this model. Furthermore, the authors note that the production pipeline currently uses a larger general-purpose model instead of Muslim-6B-PRO, and that its model card does not yet report quantified benchmark results. This leaves the utility and performance of Muslim-6B-PRO completely unvalidated. The authors should provide quantitative evaluation results (such as tool-calling accuracy or routing success rates) for Muslim-6B-PRO to support its inclusion as a functional artifact.

Author response

Thank you. v2 will correct both section cross-references, complete Section 7.2 with full latency measurements including dispersion and sample sizes, define retrieval success precisely, report operational evidence for the observability stack (observed uptime, outage counts, detection times), and add a quantitative evaluation of Muslim-6B-PRO's tool selection.

xPeerd (AI review)

October 2, 2026

AI Review (xPeerd — DAReview)

Summary

This manuscript presents MUSLIM as a deployed Arabic voice AI system for grounded Islamic knowledge, with emphasis on production engineering rather than only model construction. Its main strengths are the clear system-level motivation, a concrete architecture with explicit retrieval grounding, candid reporting of operational constraints, and deployment-aware evaluation. The manuscript is especially strong where it explains why “Islamic knowledge is retrieved, never generated from parametric memory” and how this design choice is implemented through a six-server MCP retrieval layer and a production observability stack . It also usefully reports an operational evaluation mistake involving the validator, which increases credibility by showing self-correction rather than selective reporting .

Its main weaknesses concern scope and evidentiary sufficiency. Several headline claims are only partially validated, especially around user-facing quality, faithfulness outside covered corpora, and the practical contribution of the released model family. The paper explicitly acknowledges “No formal user study” and that “Questions outside the indexed Tafsir/Hadith/metadata corpora fall back to the LLM’s parametric knowledge and may be inaccurate,” which materially limits the strength of claims about end-user reliability and groundedness in realistic settings . Overall, this is a promising and unusually deployment-conscious systems paper, but it would benefit from stronger validation, tighter claim calibration, and clearer separation between demonstrated results and forward-looking assertions.

Potential Major Revisions

1. Clarify the boundary between demonstrated grounding and residual ungrounded generation.

The manuscript’s central integrity claim is strong, but the evidence currently supports it only within a bounded subset of queries. In §3.3, the author states that “Islamic knowledge is retrieved, never generated from parametric memory” . However, the Limitations section later states that “Questions outside the indexed Tafsir/Hadith/metadata corpora fall back to the LLM’s parametric knowledge and may be inaccurate” . These two statements are not fully reconciled. As written, the first reads as a system-wide rule, while the second admits a meaningful exception for out-of-coverage questions. This is not necessarily a flaw in the system, but it is a coherence issue in the paper’s framing. The paper should explicitly distinguish between in-coverage grounded answering and out-of-coverage fallback behavior, and quantify how often such fallback occurs in real use if logs or offline test prompts are available. Without that clarification, the manuscript risks overstating the degree of grounding actually guaranteed.

2. Strengthen evaluation of answer quality and faithfulness beyond retrieval success.

The evaluation in §7.3 reports that “On 30 held-out Islamic questions, retrieval succeeds for 100% of queries” and that retrieval adds only about 50 ms latency . This is encouraging but insufficient as evidence for user-facing correctness. Retrieval success is not the same as answer faithfulness, answer completeness, or correct citation use in the final response. The manuscript itself acknowledges that “a formal faithfulness evaluation (e.g. RAGAS; Es et al., 2023) against expert-annotated ground truth is future work” . For a domain where the paper argues that misattribution is harmful, this gap is substantial. A major revision should add either expert adjudication of answer correctness on a representative benchmark or a structured faithfulness analysis comparing retrieved evidence, generated response, and final citations. At minimum, the paper should avoid using retrieval success as a proxy for epistemic reliability.

3. Provide a more rigorous and self-contained evaluation of the released model family.

Section 4 makes visible claims about released artifacts, but the evidentiary basis is uneven. For Muslim-6B-PRO, the manuscript says it is “Built for tool-call routing (31 tools) and persona discipline” using 2,731 curated examples , yet the Limitations section admits that “Muslim-6B-PRO’s model card does not yet report quantified benchmark results” . That means one of the headline contributions is presently under-evaluated within the paper. Similarly, the TTS model’s rank is based on a “community-voted, ELO-rated leaderboard” and a dated snapshot rather than a stable standardized benchmark alone . The CER benchmark helps, but the paper would be stronger if it separated community ranking evidence from controlled evaluation evidence more carefully. As a major revision, the manuscript should add quantitative evaluation for tool-routing accuracy, tool-selection precision/recall, failure rates, and error typology for Muslim-6B-PRO, rather than presenting the release itself as sufficient evidence of contribution.

4. Address limited empirical support for deployment claims involving real users.

The abstract frames the platform as “serving grounded, sourced Islamic knowledge to real users” and emphasizes production readiness . Yet the paper later concedes “No formal user study. All results are technical (accuracy, latency, retrieval success); no controlled study of learning outcomes or perceived response quality has been conducted” . For a systems paper this may be acceptable, but some claims presently read closer to product validation than infrastructure validation. If the contribution is engineering rather than human-subject effectiveness, the paper should narrow its claims accordingly. If the author wishes to preserve stronger deployment claims, then a more direct user-centered evaluation is needed: even simple statistics such as session counts, refusal rates, fallback rates, median conversation length, or anonymized satisfaction proxies would materially strengthen the empirical basis.

5. Expand reproducibility details for the live production configuration.

The architecture description is generally clear and commendably concrete. For example, the paper details the use of LiveKit, a stateless voice agent, OpenAI-compatible LLM serving, and six MCP servers . However, several practically important details remain too high-level for full reproducibility. The manuscript reports “small startup patches” for tool-schema and tool-call-token quirks in the backend without specifying their nature or failure mode . Likewise, latency figures are reported, but the hardware, network conditions, concurrency level, and measurement protocol are not sufficiently specified in the retrieved text. Since the paper’s novelty is tied to deployment realism, reproducibility should include exact serving conditions, host specifications, request concurrency, and whether latency numbers reflect cold starts, warm starts, or median live traffic conditions.

6. Tighten novelty claims relative to prior work.

The paper states, “To our knowledge, no prior published system combines real-time Arabic voice interaction, source-attributed multi-corpus retrieval, deterministic recitation validation, and a production account/metering/observability layer in one deployed platform” . This may be true, but it is a compound novelty claim assembled from multiple components, some of which are individually known. Such claims are difficult to verify and can appear overstated unless the manuscript clearly distinguishes architectural integration novelty from algorithmic novelty. The strongest novelty appears to be deployment integration in a high-stakes Arabic religious domain, not necessarily new modeling. I would recommend reframing the contribution more explicitly as systems integration, deployment methodology, and risk-aware domain adaptation, unless a stronger comparative literature review can substantiate broader originality.

7. Expand anti-fraud and coherence assessment: low paper-mill risk, but some self-referential dependencies need scrutiny.

I do not see obvious paper-mill traits such as meaningless jargon, fabricated pseudo-technical sections, or internally incoherent claims. On the contrary, the paper includes specific architectural choices, candid operational failures, and concrete limitations, all of which support authenticity . However, several key validations depend on self-authored or project-linked artifacts: a companion validator paper, Hugging Face model cards, and a live leaderboard snapshot . This does not imply fraud, but it does increase dependence on author-controlled evidence. The revision should therefore strengthen third-party or independently reproducible evaluation and make clearer which results are externally benchmarked versus internally reported.

Potential Minor Revisions

1. Title and terminology consistency.

The manuscript title begins with “Muslim” while the abstract consistently uses “MUSLIM” as the system name. The casing should be standardized across the title, abstract, headings, figures, and model names unless the distinction is intentional and explained .

2. Clarify potentially overloaded or ambiguous product language.

In the abstract, the phrase “a free per-account turn allowance, capacity-aware refusal, and email verification deferred to the point it actually matters” is vivid but somewhat informal for an academic manuscript . Similar phrasing appears in the introduction: “prior work did not need to survive real traffic” . These formulations are readable, but a more neutral register would fit better in archival writing.

3. Resolve formatting and extraction artifacts.

There are visible hyphenation and line-break artifacts throughout the text, such as “self￾hosted,” “Con￾text,” “de￾ployed,” and “la￾tency” . If these are from PDF extraction only, they should still be checked in the source file because similar artifacts sometimes persist into camera-ready versions.

4. Improve citation typography and arXiv line intrusion.

In the Related Work section, the line around “Radford arXiv:2609.31511v1 [cs.CL] 25 Sep 2026et al., 2023” appears malformed, suggesting a typesetting or bibliography insertion problem . This should be corrected because it interrupts readability and weakens presentation quality.

5. Define evaluation protocol more explicitly.

The manuscript states “Measured directly on real data” for latency and gives mean values . It should specify sample size, distributional spread, hardware, network conditions, and whether “real data” means production traffic, held-out scripted queries, or offline replay. This is partly a major issue, but it also belongs as a local reporting improvement.

6. Figure and table integration.

Figure 1 and Tables 1–2 appear central to the paper’s contributions, but the prose around them could more directly interpret them rather than merely reference them . For example, Table 2 lists licensing differences, yet the paper does not discuss how the “non-commercial” status of Fasih-TTS-V1 affects deployment, reproducibility, or reuse.

7. Ethical considerations need one additional clarification.

The ethics section states that “The system captures voice conversations for service improvement” and that capture requires both an operator-level switch and user choice . This is a thoughtful safeguard, but the paper should clarify whether any human annotation, retention duration, or religious-content sensitivity review process exists for captured conversations. Given the domain, this would strengthen the ethics discussion.

8. Minor rhetorical overstatement.

The sentence “an accuracy number that moves with which packages are installed is not a property of the system being measured” is insightful and memorable , but several similarly emphatic sentences throughout the manuscript border on editorial style rather than neutral exposition. A slightly more restrained tone would improve fit for peer-reviewed publication.

9. Typographic and grammatical issues noted from retrieved text.

10. AI content analysis for a post-2021 manuscript.

Estimated percentage of AI-generated content: low to moderate probability, approximately 10–25%.

Rationale: The manuscript does not exhibit the strongest common markers of generic AI-generated academic prose. It contains domain-specific engineering details, candid admission of an evaluation error, concrete deployment trade-offs, and non-formulaic limitations, all of which support human authorship . However, some passages are rhetorically polished in a way consistent with assisted drafting, especially compact, contrastive sentences in the introduction and abstract such as “For a domain where a misattributed narration is a real harm, these are not minor gaps” and “This paper reports on MUSLIM as a deployed system, not a prototype” .

Highlighted possibly AI-assisted sections:

• Abstract: highly compressed contribution framing with polished parallel structure
• Introduction: persuasive contrastive prose and contribution enumeration
• Some parts of Limitations and Evaluation: unusually aphoristic sentences, though still content-rich

Assessed epistemic impact: limited if assistance was only editorial. The paper’s main evidentiary weaknesses arise from validation gaps, not from probable drafting assistance. If AI tools were used for wording, the larger concern would be ensuring that every technical claim, benchmark number, and novelty statement is directly supported by verifiable evidence.

Recommendations

1. Reframe the core contribution more precisely as a deployed, retrieval-grounded Arabic Islamic voice system with bounded coverage, rather than implying universal retrieval grounding across all Islamic knowledge queries.

2. Add a faithfulness evaluation. Since the manuscript’s central claim is about safe, sourced answering, the most important missing evidence is not more latency reporting but a rigorous assessment of whether final answers remain faithful to retrieved evidence.

3. Add explicit benchmark evaluation for Muslim-6B-PRO. If this model is a headline contribution, its routing quality should be measured directly.

4. Report deployment metrics that reflect actual operation: number of sessions, refusal categories, saturation frequency, fallback frequency, and perhaps anonymized query-type breakdowns. This would substantiate the “deployed system” contribution.

5. Make reproducibility stronger by documenting hardware, software versions, configuration details, concurrency assumptions, and the exact measurement setup for latency and retrieval results.

6. Narrow or qualify novelty claims unless they are backed by a more systematic comparison table against closely related systems.

7. Preserve the manuscript’s strongest quality: its candor. The discussion of the earlier 99.2% figure and dependency-related evaluation artifact is one of the most credible parts of the paper and should remain central rather than peripheral .

8. Add one concise paragraph explicitly explaining the safety policy for out-of-coverage questions, since this is where the paper’s strongest practical and ethical risk lies .

Overall, I find the manuscript coherent, technically informed, and unlikely to be fraudulent or paper-mill generated. My main reservation is not integrity of intent, but evidentiary mismatch between ambitious reliability claims and the currently narrow validation regime.

Author response

I thank the reviewer and agree that the grounding claim needs precise scoping. v2 will distinguish in-coverage grounded answering from out-of-coverage fallback, quantify how often fallback occurs, and state the safety policy for those questions. It will add a faithfulness evaluation, a benchmark for Muslim-6B-PRO, and deployment metrics (sessions, refusal categories, saturation and fallback frequency). Reproducibility details (hardware, versions, concurrency, measurement protocol) will be documented. Novelty will be reframed as systems integration in a high-stakes Arabic domain, and naming consistency and formatting will be fixed.

XSci (AI review)

October 2, 2026

Referee Report

Manuscript: "Muslim: A Deployed Arabic Voice AI Platform for Grounded Islamic Knowledge" Field: Computational linguistics / natural language processing — applied Arabic speech-and-language systems Recommendation: Major revision required

Summary and Recommendation

This manuscript describes MUSLIM, a deployed Arabic voice assistant built on the principle that Islamic knowledge should reach users only through retrieval from source corpora rather than through a language model's parametric memory. A WebRTC-mediated pipeline combining Silero VAD, an NVIDIA NeMo Arabic FastConformer recogniser, an OpenAI-compatible language model, and self-hosted speech synthesis dispatches tool calls across six Model Context Protocol servers, three of them operator-controlled: a deterministic recitation validator, a tafsir and audio server covering eight classical works, and a hadith server holding more than 50,000 narrations across 17 collections. The author also presents a released family of fine-tuned Arabic artifacts, an account and metering layer providing a thirty-turn rolling allowance with deferred email verification and a fail-open capacity wall, and a three-layer observability design targeting the failure mode of a silent GPU host behind a healthy web tier. Reported results include 98.4% recitation-validation accuracy on 124 cases, ASR mean latency of 235.5 ms, language-model first-token latency of 519.7 ms, and 100% retrieval success on 30 held-out questions.

The work has real merits. Several design commitments reflect genuine domain understanding rather than generic engineering, most visibly the decision never to synthesise Qur'anic recitation and instead to stream authentic human recordings. The account in Section 7.1 of how an earlier 99.2% figure proved to be an artifact of a silently absent dependency, and of the decision to remove the dependency class rather than patch the instance, is a model of honest measurement reporting. The obstacles to publication are nonetheless evidentiary and scholarly rather than architectural. A paper identifying itself as a deployment report presents no deployment data; its one new empirical result measures something other than what the paper claims; a load-bearing citation appears misattributed and the premise it supports is contradicted by existing work; and the knowledge base on which the integrity claim rests is counted but not documented. Major revision is recommended.

Major Concern Category 1: Evidence, Measurement, and the Deployment Claim

1.1 A deployment paper that reports no deployment data

The manuscript stakes its identity on production experience: the title says "deployed," the abstract speaks of real users, and the third contribution is justified because prior work did not need to survive real traffic. That framing creates an obligation the paper does not meet. No account counts, session volumes, refusal-reason breakdowns, verification rates, or availability figures appear anywhere, which is hard to reconcile with Section 6, where analytics are described as instrumented for exactly these events. Several claims depend on the withheld data: the conversion argument for deferred verification is a behavioural hypothesis testable against the author's own funnel, and the thirty-turn ceiling is attributed to GPU capacity without any capacity measurement from which it could be derived. Reporting the aggregate funnel over a stated window, however modest the numbers, would turn Sections 5 and 6 from design descriptions into findings.

1.2 The new evaluation measures tool invocation rather than grounding

The paper's central claim is that Islamic content is retrieved and sourced rather than generated; the evidence is Section 7.3, where retrieval succeeds for 100% of 30 held-out questions. This is a tool-invocation metric: it indicates that a call returned something, not that the passage was correct, that the spoken answer was faithful to it, or that the citation identified the right source. A system can retrieve successfully on every query and still misattribute a narration, the harm the introduction names. The scale is also uninformative, since the exact binomial interval for 30 of 30 extends to roughly 88%, and the items are undescribed as to provenance and coverage. Because Section 7.1 defers the 98.4% figure to a companion paper, this probe constitutes nearly all of the manuscript's own empirical content.

1.3 Latency characterised as measured where the body describes an estimate

The abstract promises "real, measured latency and accuracy figures," naming an end-to-end latency of 0.9–1.7 s, and contribution 5 repeats the characterisation. The body states that this figure is estimated under typical conditions, with only the 235.5 ms ASR mean and the 519.7 ms first-token mean measured directly. The estimate is also unreconstructible from what is reported: those components sum to roughly 755 ms, while trailing silence, generation beyond the first token, synthesis, and transport remain unaccounted, and no TTS latency appears anywhere despite a self-hosted TTS model being a headline artifact. The claim of being competitive with commercial assistants is likewise comparative without a comparison. Given Section 7.1's own argument that a number which moves with its environment is not a property of the system, aligning the abstract with the body is both easy and consistent.

1.4 Reliability evidenced by a count of tests rather than their results

Section 7.4 offers more than 1,000 tests — 889 Python, 119 TypeScript — as reliability evidence, but reports no pass rate, no coverage, no distribution across the three named subsystems, and no indication of whether integration tests exercise the voice pipeline end to end. Enforcement is a pre-push git hook, which runs in the developer's own environment and is bypassable, while the continuous-integration workflow that would supply independent verification is disclosed as inactive. The candour is welcome but does not substitute for evidence, and Section 7.1 supplies the counterexample: an entire search layer was silently skipped and the suite did not catch it. Stronger evidence is within reach, since the Section 6 instrumentation already produces observed uptime, outage counts, and tallies of swallowed agent-side errors.

Major Concern Category 2: Scholarly Positioning and the Content of the Grounding Claim

2.1 A misattributed citation supporting a contradicted premise

Section 2 states that prior computational work on Quranic recitation "has focused primarily on acoustic phoneme assessment via forced alignment or end-to-end classification (Radford et al., 2023)," and that the alternatives require corpora "that do not yet exist at production quality." The reference list identifies Radford et al. (2023) as the Whisper paper, a general multilingual recogniser that is neither a recitation study nor an instance of either named approach. The premise is also questionable: QDAT is a public benchmark of roughly 1,500 recitations from 150 readers, Al Harere and Al Jallad (2023) report mid-nineties accuracies on it, and Abdelfattah, Khalil and Abbas (2025) describe a pipeline yielding 848 hours and some 286,000 annotated utterances. None of this invalidates the design choice, which has excellent independent justification in a latency-bounded pipeline with no GPU headroom; that is the argument the paper should make.

2.2 A novelty claim that surveys fields the manuscript does not engage

Section 2 closes by asserting that no prior system combines real-time Arabic voice interaction, source-attributed multi-corpus retrieval, deterministic recitation validation, and a production metering and observability layer. Such claims are acceptable when the survey behind them is visible, and here it is not: one paragraph cites four works, none of them domain prior art, while the nearest alternatives appear without citation and three commercial assistants are characterised in a single uncited sentence. Directly relevant work is absent, including the Qur'an QA shared tasks at OSACT 2022 and ArabicNLP 2023 with their AyaTEC and QRCD foundations, and the Fanar platform, which pairs Arabic speech with an attribution service for verifying fact-based generated content. A structured discussion of these strands would let the author state precisely which combination remains unoccupied.

2.3 Grounding asserted as architecture rather than established as a guarantee

The organising principle, in the introduction and again in Section 3.3, is that no Islamic content is generated without retrieval from a verified source. A retrieval-first architecture constrains where content originates, not what the model says. Between retrieval and the user sits a model that summarises, paraphrases, selects among passages and composes connective prose, and the manuscript describes no citation-span verification, entailment check, or low-overlap refusal that would catch content introduced at that stage; misattribution survives perfect retrieval. Two design features also strain the absolute phrasing: web search appears in Table 1 as one of the six retrieval servers, with no account of how open-web material is distinguished to the user from classical corpora, and Limitations concedes that out-of-corpus questions fall back to parametric memory. The principle should be scoped precisely, and whatever output-level verification exists described.

Major Concern Category 3: The Knowledge Base, the Released Artifacts, and Operational Claims

3.1 Retrieval corpora that are counted but not documented

Section 3.3 is the paper's account of its knowledge base and runs, on the substantive point, to three sentences. Eight classical tafsir works and more than 50,000 narrations across 17 collections are counted; none is named. The composition is not neutral, since an index weighted toward one group of exegetes presents a materially different interpretive profile from another, and the identity of the hadith collections determines what the system can say and on whose authority. The digitisation matters too, as classical texts circulate in editions with variant readings and uneven freely available transcriptions, and no verification against a reference edition is reported. Licensing is specified carefully for the released models but not for the redistributed corpora or the streamed audio, and the selection procedure is undescribed although it forms part of the method here.

3.2 Authenticity grading and jurisprudential stance are absent

Within the hadith sciences a narration's grade is not appended metadata but the determination of whether it may be relied upon at all, and a corpus of this size necessarily spans the full range. Grading appears nowhere in the manuscript: not in the description of the index, not in retrieval filtering or ranking, and not in what the system speaks aloud. The voice-first interface sharpens this, since a listener cannot see a grade badge and receives only what the model chose to say. A parallel issue arises on the exegetical side, where returning one work's commentary without indicating it is one among several presents a single position as the position. Relatedly, Table 1's optional fiqh server is undisclosed as to school, although the Sunni schools differ on many practical questions and silently adopting one makes a choice on the user's behalf.

3.3 No domain expert evaluates any output

Every result in Section 7 is a systems metric: whether a recited verse matched the right verse, whether a retrieval call returned something, how fast each stage ran. None addresses whether an answer a user heard was correct, appropriately hedged, correctly attributed, or acceptable to a qualified reader. Limitations acknowledges the absence of a formal user study, which understates the position, since what is missing is any expert review at all. This matters because the dominant failure mode is not an isolated wrong fact but a subtly wrong framing: a ruling without its conditions, a narration applied outside its usual context, a minority position presented without that qualification. Such failures are invisible to retrieval-success metrics and visible to a trained reader in seconds. A few hundred queries rated by two or more qualified reviewers would supply what the domain requires.

3.4 The released model family rests on one evaluated artifact, whose evidence is over-read

Contribution 2 presents a released family, but the family thins under examination. The two additional ASR checkpoints receive one sentence between them, with no data, hyperparameters, error rates, or licences. Muslim-6B-PRO is specified in configuration detail and evaluated not at all — no tool-selection accuracy across its 31 tools, no comparison against its Qwen3 base — and Table 2 discloses that it does not serve live traffic. What remains is Fasih-TTS-V1, and both of its figures need qualification. An ELO rank from 245 battles across 17 systems yields intervals wide enough that adjacent ranks are often indistinguishable. More consequentially, a 1.3% character error rate below a 1.8% human-recording baseline, on a metric computed by transcribing synthesised audio, reflects that canonical, disfluency-free synthesis is easier for a recogniser than natural speech rather than superhuman intelligibility. No listening test and no Qur'anic-register evaluation accompany either.

3.5 Operational claims without a threat model, named components, or ethical specifics

The abstract describes the metering layer as producing an abuse-resistant product, yet Section 5 states no threat model and reports no adversarial testing. Account multiplication is the clearest gap: verification is deferred and an unverified account receives a full thirty-turn window, but no account-creation control is described; the concurrency semantics of the signed allowance are unclear; and the capacity check, soundly motivated as fail-open, therefore yields precisely when admission control matters. Separately, the production language model, the serving backend to which two startup patches are applied, and the GPU hardware are all unnamed although every quantitative claim depends on them, and no code, harness, or evaluation set is released. The ethical discussion likewise asserts consent signed into the access token, no session replay, and internal-identifier analytics without retention period, access controls, deletion path, or ethics review, and never reconciles Table 2's "production-derived examples" with the consent model.

Minor Issues

1. Section 7.2, titled "Latency," appears with no body text; the heading for Section 7.3 follows immediately, and the latency figures then appear beneath it, with the retrieval results in the paragraph after. The published version should restore the latency results to Section 7.2.

2. Two cross-references point to the wrong section. Section 3.2 states that the independent-failure trade-off is discussed in Section 7.4, which covers test-suite evidence; it is in fact treated under Limitations. Ethical Considerations likewise cites Section 7.4 for the retrieval layer's covered domains, where Section 3.3 appears intended.

3. Every latency figure is a bare mean. The 235.5 ms and 519.7 ms values, and the 786 ms against 736 ms comparison, appear without the number of observations, the collection window, or any dispersion statistic. For an interactive voice product the tail is the user-perceptible quantity, so median and p95 would be considerably more informative.

4. The term "verified source" carries the manuscript's integrity claim but is never defined. The reader is not told what makes a corpus verified, which verification was performed, by whom, or against what reference. A stated criterion would make the central claim checkable and would clarify how the web-search server relates to it.

5. The reciters, readings, recordings, and rights status of the streamed recitation audio are unidentified. Section 3.3 explains persuasively why authentic human recitation is used rather than synthesis, but the audio is redistributed to users, and both whose recitation is served and under what terms deserve a sentence.

6. Qira'at support is listed in Table 1 as a capability of the remote metadata server, while the recitation validator presumably normalises against a single reading. How the validator behaves when a user recites in a different canonical reading is a correctness question arising from the manuscript's own feature list.

7. The health endpoint is offered as a reusable design but cannot be implemented from the paper. The heartbeat interval, the staleness threshold at which the external judge declares an outage, and the status codes encoding the three-way verdict are all omitted, although Section 6 is framed as transferable guidance.

8. Deliberately swallowed agent-side errors are mentioned but never quantified. Which classes are suppressed to keep a conversation alive, how often, and with what effect on the resulting answer would be among the most useful operational facts the paper could report, and the instrumentation already exists.

9. The token-minting route's signing scheme, token lifetime, and replay protection are unstated. The claim that the allowance is signed into the token as attributes the client cannot forge is the foundation of the enforcement design, and unforgeability depends on details the manuscript does not supply.

10. The 2,731 training examples behind Muslim-6B-PRO are uncharacterised. The split between synthetic and production-derived material, the distribution across the 31 target tools, the curation procedure, and whether a held-out portion exists are all unreported; per-tool coverage is the first quantity a reader will want.

Missing and Inadequately Integrated Literature

Cited but Inadequately Integrated

Radford et al. (2023) is cited in a role it cannot fill, as the representative of prior work on Quranic recitation assessment. Its genuine relevance here is as a baseline against which the chosen FastConformer could be compared on Arabic and Qur'anic audio, which would support the model-selection argument in Section 3.1, where the claim that the training distribution includes religious Arabic is unsourced and no Arabic word error rate is reported.

Es et al. (2023) appears only in Limitations as deferred work, although its faithfulness measure addresses exactly the gap between the retrieval-success metric and the attribution claim; applied to even a modest expert-annotated set it would let the paper report something about output faithfulness, using tooling the author has already identified. Lewis et al. (2020) is invoked for the general case for retrieval without engaging its finding that retrieval reduces rather than eliminates unsupported generation, and in this domain the residual rate is the quantity of interest.

Casanova et al. (2024) is cited as the base for Fasih-TTS-V1 without engaging the relationship between a massively multilingual zero-shot system and a 2.4-hour single-speaker fine-tune; reporting the base model's Arabic behaviour would isolate what the fine-tuning and the custom front-end contributed. Navid AI (2026) and SILMA AI (2026) supply both headline synthesis figures but are cited only by link, and since the interpretation of each depends on the protocol, both instruments deserve description. Elnawasany (2026) carries the strongest accuracy figure and the text edition underlying the 2,290-pair mapping, and should be summarised sufficiently that the quoted result is interpretable without retrieving the preprint.

Truly Missing Papers

Al Harere and Al Jallad (2023) and Abdelfattah, Khalil and Abbas (2025) are both uncited and both bear on the premise of Section 2. The former evaluates MFCC and LSTM models on the public QDAT corpus of roughly 1,500 recitations from 150 readers; the latter presents a largely automated pipeline producing 848 hours of recitation audio and some 286,000 annotated utterances with a reannotated Tajweed benchmark. Together they establish that labelled recitation-error resources exist, and they are the natural points of contrast for the text-level validator.

Malhas and Elsayed (2020) and Malhas, Mansour and Elsayed (2023) supply the field's standard evaluation apparatus for Qur'anic question answering: AyaTEC, a reusable verse-based collection of 207 questions and 1,762 answers built with the participation of tafsir specialists, and the Qur'an QA 2023 shared task over a 1,266-passage collection. Either would replace the undescribed 30-question probe with a community benchmark. Fanar Team (2025) describes an Arabic-centric multimodal platform including Arabic speech and an attribution service for verifying fact-based generated content, and is essential to the novelty claim. Kulkarni et al. (2023) provides ClArTTS, a 12-hour single-speaker Classical Arabic corpus with diacritised transcriptions and mean-opinion-score baselines, both a larger training resource than the 2.4 hours used here and the reference point for the subjective evaluation the manuscript omits. Bashir et al. (2023) releases the ABCD dataset of roughly 29,000 hadith with 18,000 narrators and their biographical information, a natural starting point for adding grade and narrator metadata to the indexed corpus. All citations in this section were verified via WebSearch.

Abdelfattah, M., M. Khalil, and H. Abbas. 2025. Automatic pronunciation error detection and correction of the Holy Quran's learners using deep learning. arXiv:2509.00094.
Al Harere, A., and K. Al Jallad. 2023. Mispronunciation detection of basic Quranic recitation rules using deep learning. arXiv:2305.06429.
Bashir, M. H., et al. 2023. Hadiths classification using a novel author-based hadith classification dataset (ABCD). Big Data and Cognitive Computing 7(3):141.
Fanar Team. 2025. Fanar: An Arabic-centric multimodal generative AI platform. arXiv:2501.13944.
Kulkarni, A., A. Kulkarni, S. A. M. Shatnawi, and H. Aldarmaki. 2023. ClArTTS: An open-source Classical Arabic text-to-speech corpus. In Interspeech 2023, pages 5511-5515.
Malhas, R., and T. Elsayed. 2020. AyaTEC: Building a reusable verse-based test collection for Arabic question answering on the Holy Qur'an. ACM TALLIP 19(6):1-21.
Malhas, R., W. Mansour, and T. Elsayed. 2023. Qur'an QA 2023 shared task. In Proceedings of ArabicNLP 2023, pages 690-701.

Conclusion and Path Forward

The manuscript documents a system whose architectural commitments are sound and whose author writes about engineering reality with a candour that is genuinely valuable. Making retrieval the only path to Islamic content is the right decision for this domain; concentrating access control at the single token-minting route and routing liveness outward to a judge beyond both halves of a split topology reflect operational experience rather than textbook defaults; and the account of how an absent optional dependency inflated an earlier accuracy figure is the kind of self-correction the field benefits from reading.

What stands between the manuscript and publication is the distance between its claims and its evidence. The deployment framing needs deployment numbers, and the instrumentation to produce them is already running. The grounding claim needs an evaluation of output faithfulness and attribution correctness at meaningful scale, ideally with domain-expert review, rather than a thirty-item tool-invocation check. The positioning needs a related-work section engaging Qur'anic question answering, recitation assessment, Arabic speech resources, and deployed Arabic platforms, with the misattributed citation replaced by the literature it currently stands in for. The knowledge base needs to be named, sourced, licensed, and honest about authenticity grading. And the abstract needs to describe as estimated what the body describes as estimated. None of this requires new architecture; most requires reporting what the author already knows or can measure within weeks. The system appears to merit that effort, and a revision along these lines could make a solid contribution.

Author response

I thank the reviewer for a detailed and fair report and accept that the main gap is between the paper's claims and its evidence. v2 will: report aggregate deployment data (accounts, sessions, refusal reasons, verification and availability) over a stated window; replace the 30-question retrieval check with a faithfulness and attribution evaluation, with domain-expert review, using community benchmarks such as AyaTEC/QRCD; name every tafsir work and hadith collection with sources, editions and licences, and address hadith grading and the fiqh server's school; describe end-to-end latency as estimated in the abstract and report median, p95 and TTS latency; evaluate Muslim-6B-PRO's tool routing; state a threat model for the metering layer; and name the production LLM, serving backend and hardware.

The misattributed citation in §2 will be replaced with the recitation-assessment literature it should point to (Al Harere & Al Jallad; Abdelfattah et al.), and related work will add Qur'an QA, ClArTTS, Fanar and ABCD. Section 7.2 and the cross-references will be fixed.