Reviewer3 (AI review)

October 2, 2026

AI Review (reviewer3.com)

Comments (5)

1. Step 6 naming mismatch between text and Table 1 — Editorial

(6) alef-madda resolution, deleting a hamza written immediately before an alef, since Uthmani spells alef madda as hamza plus alef wherever it oc- curs and not only word-initially, together with ya- maqsura folding; (7) non-Arabic character removal

In Section 4.1, the sixth step of the normalization pipeline is described as 'alef-madda resolution... together with ya-maqsura folding'. However, in Table 1, Step 6 is labeled as 'Word-initial rules'. This creates an inconsistency in the naming of the pipeline steps.

2. Lossy majority-vote conflict resolution — Major

: exactly two keys occur with more than one Standard form across the corpus, and in both the majority is unambiguous, so the most frequent form is kept; no key ends in a tie, and no entry required manual curation.

For the two keys that occurred with multiple Standard forms, the authors resolved conflicts by keeping only the most frequent form. While this simplifies the mapping to a static 1-to-1 lookup, it introduces systematic errors when the minority form is the correct orthographic representation in context. The authors should identify these two keys, discuss the linguistic reasons for their multiple Standard forms, and evaluate the impact of discarding the minority forms on downstream NLP tasks or validation accuracy.

3. Unevaluated performance on non-identical verses — Major

The cross-form round trip is 90.9%, not 100%: for 567 verses the two forms still normalize to different strings.

The normalization pipeline fails to reduce 9.1% of Quranic verses to identical strings. While the authors acknowledge this limitation of the normalizer, they do not analyze how this failure affects the validator's performance on these 567 verses. If these verses are not represented in the evaluation suite, the reported 98.4% accuracy may not generalize to them. The authors should evaluate the validator's performance specifically on queries from these 567 non-identical verses to demonstrate whether the search layers can successfully overcome the normalization mismatch.

4. Unspecified excluded word form — Major

ing U+0670, so coverage is 2,290 of 2,291: one form, appearing in a single verse whose two read- ings admit no consistent span alignment, is delib- erately left out rather than guessed. Words without

The authors state that one Uthmani word form carrying U+0670 was deliberately excluded from the mapping because its two readings did not admit a consistent span alignment. However, the paper does not identify this specific word form or the verse in which it occurs. To ensure full reproducibility and completeness of the resource, the authors should explicitly name this excluded word form and explain the specific alignment conflict.

5. Unevaluated computational latency of fuzzy matching — Major

Layer 4 uses the standard li- brary’s difflib.SequenceMatcher rather than a faster third-party matcher deliberately: an op- tional dependency makes the validator’s output de- pend on which version of that library happens to be installed, which would defeat the determinism the rest of the design is built for. Candidates are

The fourth layer of the search pipeline uses Python's native difflib.SequenceMatcher to perform character-level fuzzy matching across all 6,236 verses when previous layers fail. Because SequenceMatcher has quadratic complexity and is implemented in pure Python, running it over thousands of verses without a pre-filtering step could introduce severe latency. The authors do not report any execution time or computational complexity analysis, which is critical for a validator intended for real-time deployment. The authors should provide latency benchmarks for queries that trigger the fuzzy matching layer.

Author response

Thank you for these focused comments. In v2: the Table 1 step-6 label will be corrected to match §4.1; the two majority-vote keys and the single excluded word form will be named, with their competing Standard forms, counts and verses; the validator will be evaluated specifically on queries drawn from the 567 non-converging verses; and latency benchmarks (median and tail) will be reported for queries that reach the fuzzy-matching layer, together with how often each layer is reached.

xPeerd (AI review)

October 2, 2026

AI Review (xPeerd — DAReview)

Summary

This manuscript presents two linked contributions: a corpus-aligned Uthmani-to-Standard Quranic word mapping and a deterministic recitation validator built on top of that mapping. The paper’s central technical motivation is clear and relevant: mainstream Arabic NLP tooling mishandles the Uthmani superscript alef U+0670, creating systematic byte-level mismatches between canonical Quranic text and Standard-Arabic-oriented NLP or ASR pipelines. The manuscript claims a 2,290-pair mapping built from alignment across all 6,236 Quranic verses and a seven-step normalization pipeline that yields identical normalized strings for 90.9% of verses across the two orthographic forms. On top of this, it introduces a four-layer, LLM-free validator and reports 98.4% accuracy on a released 124-case test suite, while openly releasing code and data. The work’s strengths are transparency, determinism, reproducibility for most reported numbers, careful failure analysis, and explicit acknowledgment of limits. Its main weaknesses are evaluation dependence on an internally authored test suite, lack of a held-out benchmark, limited real-world validation, incomplete normalization coverage beyond U+0670, and some novelty claims that are plausible but not independently substantiated within the manuscript. Overall, I judge the paper coherent and technically serious, with no strong paper-mill signals, but still needing stronger external validation and sharper framing of claims.

Potential Major Revisions

1. Strengthen the evidence for evaluation validity and generalization.

The manuscript openly states that “The evaluation suite is our own: we wrote the 124 cases, chose their grouping, and set the search thresholds, and there is no held-out split separating development from evaluation” and further concedes that it “measures coverage of anticipated failure modes rather than performance on an independent sample” . This is an important admission, but it also substantially weakens the force of the headline claim that the validator “scores 98.4% (122/124)” . I think the paper should revise its evidentiary framing more aggressively. At present, the accuracy figure is technically honest but rhetorically prominent. The manuscript should either downgrade the prominence of 98.4% in the abstract and introduction or supplement it with a more independent test design.

A major revision would be to add a genuine held-out evaluation partition, ideally one generated after freezing rules and thresholds, or to construct a benchmark from externally sourced recitation transcripts with verified verse labels. As written, the paper itself says the current suite was “built to cover the normalization edge cases we knew of” . That wording implies risk of overfitting to anticipated failure modes, even if no parameter tuning was done directly on the suite.

2. Reframe and stress-test the real-world deployment evidence.

The manuscript reports that “on 34 recitation transcripts drawn from a deployed Arabic ASR system the validator identifies the correct verse in every case” , but later clarifies that only transcripts whose normalized form is “an exact contiguous span of exactly one verse” were labeled and that “Labeling by exact containment selects transcripts the ASR happened to get right” . This is commendably transparent, yet it means the 34/34 result is much narrower than a casual reader may initially infer.

I think the manuscript should make this narrower interpretation unavoidable earlier in the paper. The current phrasing in the abstract risks over-reading because “identifies the correct verse in every case” sounds broader than the later qualification supports. A stronger revision would explicitly state in the abstract that these 34 cases are a selected, clean subset of real traffic, not a representative sample of noisy ASR recitation input. The manuscript itself gives the more honest summary: “real recitation transcripts, when they are transcribed correctly, are identified correctly, and that we do not yet have labelled ground truth for the ones that are not” . That sentence should influence the front-facing claims more strongly.

3. Address the unresolved mismatch between contribution framing and residual normalization gaps.

The introduction argues that the core underreported difficulty is concentrated in U+0670 and motivates a word-level resource rather than character-level heuristics . The mapping indeed appears useful, but the manuscript later reports that after normalization “5,669 verses (90.9%) reduce to identical strings,” leaving “567 verses” unresolved because of “other Uthmani-Standard divergences the character rules do not yet cover” . The limitations section further specifies residual classes such as doubled lam in the definite article, reduced-letter spellings, and medial hamza forms, adding that “Each is tractable and none is addressed here” .

I think this is a significant boundary on the practical completeness of the normalization pipeline, and it needs stronger integration into the contribution claims. At present, the paper’s first contribution can be read as if the orthographic interoperability gap is largely solved, whereas the results show a partial but important solution. I recommend reframing the contribution as a high-value first resource centered on U+0670 rather than as a more generally sufficient bridge across forms. That would align the claims more tightly with the reported 90.9% round-trip result.

4. Expand the analysis of failure modes in the search design and justify why they are acceptable.

The paper’s failure analysis is one of its strongest sections. It shows that in Case SS03, “a single substitution error turned a different verse into the exact match,” and in ME03 the system becomes “confidently and comprehensively wrong” after anchoring on the wrong opening verse . The manuscript concludes that “layer 1 has no notion of how surprising a match is” and that fixing this would require scoring candidates against alternatives rather than accepting the first exact hit .

This is an excellent diagnosis, but I think the paper should go one step further and either implement or at least simulate an alternative candidate-ranking strategy. Since the validator is intended to produce feedback on religious text, the consequences of confident misidentification deserve fuller treatment. The ethics section notes that declining an answer at high WER is preferable to authoritative-sounding incorrect correction , but the documented failures are not merely uncertainty failures; they are incorrect certainty failures created by the ranking logic. That distinction matters. A major revision should include a candidate-calibration mechanism, ambiguity flagging rule, or abstention policy for cases where exact-token matches are non-unique or semantically surprising.

5. Substantiate novelty claims with more careful wording.

The paper says, “To our knowledge, no published word-level resource maps between the two forms” and later calls the released mapping “the first published resource of its kind” . These claims may be correct, but within the manuscript they rely mainly on surrounding-resource discussion rather than a systematic review. Because novelty is central to the paper’s contribution, I think the wording should remain cautious and perhaps be narrowed to “no such published resource identified in the literature reviewed here.” This is especially important in a niche but potentially multilingual domain where relevant tools may exist outside the most visible CS.CL publication channels.

6. Improve reproducibility boundaries and auditability of non-reproducible claims.

The paper is notably strong in data/code release, stating that the repository reproduces the construction counts, coverage figures, ambiguity census, core tables, and the 98.4% total, while explicitly noting that the toolkit comparison in §1 and deployment measurement in §5.4 are not fully reproducible from the repository alone . This is good practice. However, because the abstract also says “every number in this paper except the deployment measurement... is reproduced by running them” , the unreproduced toolkit comparison should be foregrounded as a partial exception requiring additional external packages and exact versions. This is a smaller issue than the validation concerns above, but still important for methodological precision.

7. Anti-fraud / coherence assessment.

I do not see strong signs of fabrication or paper-mill behavior. The manuscript is unusually specific about mechanisms, thresholds, counts, and failure cases; it reports negative findings against its own stronger possible narrative; and it distinguishes reproducible from non-reproducible numbers carefully . Those are positive credibility indicators. Still, one cautionary feature is that several key claims depend on author-constructed evaluation assets and selected deployment subsets, which can unintentionally inflate apparent performance. I therefore do not issue a fraud warning, but I do advise readers to treat the current evaluation as promising engineering evidence rather than definitive external validation.

Potential Minor Revisions

1. Clarify wording that may overstate robustness.

In the abstract, “the validator identifies the correct verse in every case” for the 34 real ASR transcripts should be qualified immediately as applying only to labeled transcripts meeting a stringent exact-containment criterion, since the later section explains that these are selected clean cases rather than the full noisy distribution .

2. Improve consistency in how limitations are signposted.

The manuscript does a good job disclosing limitations in §8, but some of the most important limitations first appear only deep in the paper. For readability, the abstract or introduction should mention that the cross-form round trip is “90.9%, not 100%” and that the evaluation suite is internally authored with no held-out split .

3. Tighten a few potentially ambiguous claims.

The sentence “What Standard writes in its place is not fixed” is clear conceptually, but readers may benefit from a more compact formal description of the mapping variability classes already implied by the statistics in the footnote and later sections . Likewise, “fully deterministic, training-data-free, immediately deployable pipeline” is attractive wording, but “immediately deployable” may be too broad given acknowledged limitations in Tajweed detection and residual orthographic divergence .

4. Formatting and typographic issues visible in the uploaded file.

The PDF text extraction shows several broken line-wrap and encoding artifacts, such as “automatic-speech￾recognition,” “ref￾erence,” “valida￾tor,” and similar intraword breaks across many sections . These may be PDF extraction artifacts rather than source-manuscript errors, but if they reflect the submitted version, they should be corrected. The arXiv stamp also appears merged into running text near the introduction-to-related-work transition: “arXiv:2609.14967v1 [cs.CL] 14 Sep 2026a four-layer search...” . That suggests a layout interruption that should be checked in the source.

5. Table presentation could be more interpretable.

Table 3 is useful, but the sentence explaining why a better-performing threshold “is deliberately not adopted” is important enough to be summarized directly in the table caption or adjacent prose more explicitly . Table 4 is concise and helpful, but its juxtaposition of synthetic-suite accuracy and selected real-ASR cases may invite over-comparison unless a note clarifies the difference in labeling regime .

6. Statistical framing.

The manuscript uses percentages such as 98.4%, 90.9%, and 16.5% clearly, but adding raw counts consistently beside each percentage throughout the prose would improve precision. In several places this is already done well, e.g., “1,030 / 6,236 (16.5%)” . That practice should be made uniform.

7. AI content analysis.

Estimated AI-generated content: low.

I do not see obvious signs of post-2021 generative-AI drafting such as generic inflated claims, semantically hollow transitions, repetitive abstract phrasing, or fabricated citation patterns. On the contrary, the manuscript contains unusually concrete implementation details, self-critical failure analysis, and specific counts that are not typical of generic AI-authored text. Examples include “one form, appearing in a single verse whose two readings admit no consistent span alignment, is deliberately left out rather than guessed” and “The honest summary is that real recitation transcripts, when they are transcribed correctly, are identified correctly” . These are signs of disciplined human technical authorship.

Estimated percentage of AI-generated content: 5–15%.

Highlighted AI-detected sections: none obvious. If any assistance was used, it is not evident from stylistic incoherence. The epistemic impact of possible AI assistance appears negligible relative to the manuscript’s actual contribution, because the core value here lies in the released alignment resource, the deterministic pipeline, and the explicit evaluation logic rather than in polished expository prose.

Recommendations

1. Recast the headline empirical message more conservatively. Present the work as a reproducible resource paper with a promising validator prototype, not yet as a broadly validated performance result. This would better match the manuscript’s own evidence base.

2. Add an external or frozen held-out evaluation. Even a modest benchmark of independently labeled noisy recitation transcripts would materially strengthen the paper. The current suite is useful for regression testing but weaker as a publication-grade generalization claim.

3. Introduce a safer abstention or candidate-ranking mechanism. Since the failure mode can be “confidently and comprehensively wrong,” the design should prefer calibrated uncertainty over exact-match overcommitment in ambiguous cases.

4. Narrow contribution claims to what is fully demonstrated. The mapping centered on U+0670 is clearly valuable; the broader Uthmani-Standard normalization problem remains only partially solved, as the 90.9% round-trip figure already shows.

5. Make the deployment evidence easier to interpret responsibly. Move the selection-bias explanation closer to the first mention of the 34/34 result and clarify that the unlabeled degraded cases remain an open evaluation gap.

6. Preserve the current strong transparency practices. I think one of the manuscript’s best qualities is its refusal to hide limitations, especially around threshold selection, failure analysis, and reproducibility boundaries. Those features should remain central in any revision.

Overall, I regard this as a credible, coherent, and potentially useful contribution with solid engineering value, but one that still requires major revision before its evaluation claims can be considered fully persuasive.

Author response

I thank the reviewer for the thorough assessment and agree with the recommended reframing. v2 will present the work as a reproducible resource paper with a validator prototype. The abstract will state that the evaluation suite is author-constructed with no held-out split, that cross-form convergence is 90.9% rather than complete, and that the 34/34 deployment result covers a selected, clean subset. I will add a frozen held-out evaluation and introduce an abstention or candidate-ranking rule for non-unique exact matches, the mechanism behind both documented failures. Novelty wording will be narrowed to the literature reviewed, the toolkit comparison will be flagged as a reproducibility exception, and the layout artifacts will be fixed.

XSci (AI review)

October 2, 2026

Referee Report

Manuscript: A Corpus-Aligned Uthmani-to-Standard Quranic Word Mapping and a Deterministic Recitation Validator Field: Computational linguistics / Arabic natural language processing (language resources and evaluation) Recommendation: Major revision required

Summary and Recommendation

This manuscript addresses an orthographic interoperability problem that is real, under-documented, and of direct practical consequence. The Quran circulates in two byte-level distinct written forms, the Uthmani script of the printed mushaf and the Standard (Imla'i) Arabic that mainstream NLP tooling assumes, and the authors locate the difficulty principally in U+0670, the superscript alef, showing that no global character rule resolves it: Standard writes a full alef in 1,335 of their 2,290 cases and nothing at all in 362, so that deleting the character yields the correct Standard form for only 359 words and keeping it for none. In response the authors release a 2,290-pair word-level mapping, induced in four alignment passes over all 6,236 verses and covering 2,290 of the 2,291 attested U+0670-bearing word forms, together with a seven-step normalization pipeline and a deterministic, language-model-free recitation validator whose four search layers escalate from exact token matching through morphological expansion and relaxed coverage to character-level fuzzy similarity, with feedback graded by word error rate across five tiers. Reported results are 90.9 per cent cross-form convergence, 98.4 per cent (122/124) on a released evaluation suite, a census finding 16.5 per cent of verses ambiguous from a four-word opening, and correct identification on 34 of 34 selected deployment transcripts.

The engineering is sound and the manuscript is in several respects unusually self-critical: the authors decline a threshold setting that would have raised their headline number and explain why, they refuse to patch the search against the two cases that defeat it, they itemise the 567 residual divergent verses into named classes rather than gesturing at them, and they warn the reader unprompted that the deployment result measures the clean end of real input. These are good instincts and they recommend the work. The difficulty is that the same evidential discipline has not been applied to the paper's central claims. The mapping's correctness is established by a procedure that, for structural reasons, cannot produce a negative result; the validator's accuracy is measured on a suite that, by the authors' own sensitivity analysis, never exercises two of its four layers; the user-facing feedback component is never evaluated at all; and the architecture is never compared against the obvious alternative of indexing the Standard text the same corpus provider publishes. Major revision is recommended. None of what follows requires the authors to change what they built, and most of it can be addressed with the artifacts and data already in hand.

Major Concern Category 1: The Mapping and the Basis for Trusting It

1.1 The Standard side of the alignment may not be an independent observation

The paper's principal theoretical move, made in §2, is that the Uthmani-Standard correspondence is "descriptive and recoverable by alignment over parallel corpora rather than stipulated by guideline," which is what distinguishes the contribution from the prescriptive conventionalisation of CODA (Habash et al., 2012). That framing presupposes that the two sides of the parallel corpus are independent orthographic traditions that happen to be alignable. The Tanzil project's own published documentation appears to describe something different: a core Quran text extracted by comparison of authentic sources, followed by a rule-based verification stage that was able to produce Imlaei and Uthmani texts "all derived automatically from the core text." If that description holds, an alignment over the two would recover the generating program's conversion rules with high fidelity, which would in turn explain the remarkably clean statistics of §3.1, where exactly two keys conflict, neither ties, and no entry required manual curation. For a descriptive alignment over two independently maintained traditions, that consistency would be surprising; for one machine-generated correspondence, it is what one would expect. The authors should state the provenance of both texts explicitly, say whether they regard the two sides as independent, and validate a sample of pairs against a source outside the Tanzil family.

1.2 The cross-form figure is measured on the data the mapping was induced from

Section 4.1 reports that normalizing both forms of all 6,236 verses yields identical strings for 5,669 of them, and the abstract foregrounds the 90.9 per cent figure. The mapping, however, was built in §3.1 by aligning these same verses and retaining every U+0670-bearing Uthmani word as a key, and step 2 of the pipeline is a lookup into that table. For every corpus word carrying U+0670 whose two forms were alignable, the normalized Uthmani form is therefore defined to equal the normalized Standard form, and the measurement cannot fail on the phenomenon the paper is about. What the 567 residual verses actually reveal is the behaviour of the remaining character-level rules on divergences the mapping does not cover, which is a narrower and different thing. The number is not wrong and the authors do not overstate its completeness, but it carries no information about generalisation while the surrounding rhetoric — "we measure it rather than assert it" — invites the reader to treat it as if it did. Building the mapping on a held-out portion of the corpus and reporting convergence on the remainder would answer the question a reader actually has.

1.3 No one examined the pairs, and the artifact's verifiability rests on unstated preconditions

Section 3.1 presents the fully automatic construction as a virtue, noting that "no entry required manual curation." As a statement about the alignment's cleanliness this is fair; as a statement about validation it is the difficulty, because nobody read the 2,290 pairs. There is no sampled accuracy check by a reader competent in Quranic orthography, no inter-annotator agreement, and no adjudication protocol, and the same absence extends to the five feedback tiers, whose boundaries were set by the authors without reported review by anyone trained in recitation pedagogy. An automatic procedure is a reason to validate a sample, not to skip validation, and a stratified sample of two or three hundred pairs checked by two independent readers would yield a usable estimate at modest cost. Compounding this, the strong reproducibility claim of §6 depends on conditions the paper never records: the artifact is identified only by a repository URL with no tag, commit or archival deposit; no interpreter version is given for the main results, only for the toolkit comparison the repository cannot reproduce; and the upstream corpus is published in several Simple-family variants with download-time options and a revision history, none of which is specified. Archiving the inputs, pinning the environment, tagging a release, and adding a datasheet in the sense of Gebru et al. (2021) would make a strong claim a checkable one.

Major Concern Category 2: What the Evaluation Establishes

2.1 The suite exercises only the upper search layers, and no proportion carries an interval

The 98.4 per cent headline is broken down across twelve categories in Table 2, and the authors are forthright in §8 that the suite is their own and has no held-out split. A stronger consequence follows from their own Table 3 and goes unremarked. Suite accuracy is unchanged for every relaxed-coverage threshold from 0.40 to 0.80 and for every fuzzy threshold from 0.25 to 0.55, and §4.2 notes that disabling layer 4 entirely also yields 123 of 124. A parameter that can be varied by a factor of two without moving a single case is one that no case depends on, so the headline figure measures layers 1 and 2 on anticipated inputs, and nothing in the manuscript bears on the relaxed and fuzzy layers that exist precisely for degraded ASR output. Sweeping the generator's corruption rate upward until those layers fire, and reporting accuracy against injected word error rate alongside a per-layer firing distribution, would close this without new data. Relatedly, none of the paper's eight headline proportions carries an interval; 122 of 124 has a Wilson interval of roughly 94.5 to 99.6 per cent, and the paragraph in §4.2 reasoning about a one-case difference would be strengthened, not weakened, by noting that the difference is within noise.

2.2 No baseline or ablation isolates what the mapping contributes

Footnote 1 persuasively establishes that generic normalizers are poor word-level normalizers of U+0670. It does not establish that they are poor front ends to a four-layer fuzzy search, which is the claim the paper's architecture needs. Because both query and reference text pass through the same pipeline, a systematically wrong but consistent treatment of U+0670 may leave verse identification entirely unaffected: a global delete applied to both sides would still produce matching strings wherever U+0670 is the only divergence. The experiment that would settle this is a single ablation the released harness could run quickly — replace step 2 with a global delete, then with a global keep, then remove it so that step 4's contextual rule handles all occurrences, and report suite accuracy in each case. If accuracy collapses, the authors gain their strongest result; if it is unchanged, that is also a finding, and an important one, since it would show the mapping matters for cross-form identity and downstream interoperability but not for this retrieval task, allowing the claims to be scoped precisely. The same gap applies more broadly: no baseline of any kind is reported, not even a character n-gram or term-weighted index over the normalized corpus, so a reader cannot judge whether 98.4 per cent sits above the floor for a 6,236-document problem whose queries are drawn verbatim from the documents.

2.3 The feedback component is unevaluated and the deployment result is close to definitional

Section 4.3 occupies a single paragraph and nothing else in the manuscript concerns it, yet it describes the output users actually receive and is where the ethical argument of §7 is concentrated. That argument — that an incorrect authoritative-sounding correction is the serious failure mode and that the five tiers exist to prevent it — is a claim about calibration, and no evidence for that calibration is offered. Case ME03 is a counterexample in the authors' own data: a fifteen-word recitation reported as verse 112:4 with ten insertions should have exceeded the 0.60 abstention threshold, and the paper does not explain why it did not. Since the generator records the errors it injects, the precision and recall of the reported correction lists, and a confusion matrix of assigned against true tier, are directly computable. The deployment evaluation raises a parallel issue. Labelling transcripts by whether their normalized form is an exact contiguous span of exactly one verse is described as "independent of the validator's four-layer search," but such a transcript satisfies layer 1's AND-matching condition for the correct verse by construction, so 34 of 34 reports that the §5.2 collision did not occur in 34 clean draws. The informative data are present and discarded: of the 93 utterances overlapping a verse by 80 per cent or more, the 59 that are not exact are the degraded real cases for which no evidence exists, and they are few enough to label by hand.

Major Concern Category 3: Analysis, Positioning and Architecture

3.1 The normalization rules are stipulated, and the collisions they create are attributed to the text

Several steps in §4.1 encode substantive claims about Arabic orthography that are presented without derivation. Step 4 resolves unmapped U+0670 to alef "except after ya-maqsura, dhal, ha, or lam," a closed four-member exception list given with no counts, no citation, and no indication whether it exhausts the attested contexts or merely the cases encountered. Step 6 asserts that Uthmani spells alef madda as hamza plus alef "wherever it occurs and not only word-initially" and on that basis deletes the hamza unconditionally, again without a supporting count. The pipeline is also deliberately lossy — stripping all tashkeel, folding four alef variants and the connected hamza forms, folding ya-maqsura — and the aggregate effect is never quantified, although a count of word types and verses that become identical after normalization is readily computable and is the natural companion to the 90.9 per cent convergence figure. The omission has a visible consequence: §5.3 measures ambiguity over "identical normalized four-word openings" while §8 declares that ambiguity "an inherent property of the text, not a resolvable engineering gap." Recomputing the census over raw, fully diacritized, and normalized text would separate the inherent component from the normalizer's contribution, and could only help the paper. It is worth noting too that §5.4's input comes from a diacritizing ASR model whose diacritics step 3 discards, so the possibility that they carry usable disambiguating signal is never tested.

3.2 The ambiguity census measures a relation the documented failures do not instantiate

Section 5.3 is positioned as the paper's characterisation of the inherent limits of text-only validation and is reported as a headline statistic in Table 4. The relation it measures is identity of the first four normalized words between two verses. The failure mechanism documented in §5.2, however, is token containment rather than opening identity: case SS03 fails because the single query token wahid appears somewhere in 6:19, and ME03 fails because the four opening tokens are all present there. Verses 112:1 and 6:19 share no opening and would never fall into the same ambiguous group, so the census does not cover the only failure mode the paper has observed, while §5.3's statement that the ambiguity "does not affect the remaining 83.5 per cent of verses" invites a safety inference the data do not support. The quantity that would bound the observed mechanism is equally computable: for each verse, the number of other verses containing all of its opening tokens, or more generally the distribution over verses of superset candidates under layer 1's matching condition. Reporting the ambiguity curve as a function of opening length rather than at a fixed four words, which corresponds to no parameter the validator uses, would also give deployers something actionable.

3.3 Priority is asserted, prior art is omitted, and an obvious alternative goes unexamined

Variants of "to our knowledge, the first" appear four times, and §3.2 adds a substantive claim about the prior landscape, yet no search procedure, query set, or inclusion criterion is reported. The concern is sharpened by what Related Work omits: it covers corpora, dialectal conventionalisation, general surveys, diacritization restoration and acoustic recitation assessment, but passes over the literature on Quranic information retrieval entirely, which is where matching user text against verses across orthographic variants has been worked on for well over a decade. More consequentially, the paper never considers indexing the Standard (Imla'i) text that the same provider publishes for the same 6,236 verses and that the authors already hold as one side of their alignment. Under that alternative, ASR output in Standard form meets reference text in Standard form, verse identifiers are shared, and the U+0670 problem does not arise for retrieval at all. If there is a good reason it cannot serve — perhaps that feedback must be expressed in mushaf orthography, or that the morphological annotation is keyed to Uthmani tokenization — that reason is the architectural justification for the whole system and belongs in §1. Related constraints also deserve surfacing: the mapping is type-level, so majority-vote resolution cannot represent context-conditioned correspondence; the 58 one-to-many entries make it non-invertible and tokenization-dependent; and coverage against a non-Tanzil Uthmani encoding is untested.

Minor Issues

1. Footnote 1 reports that a global delete yields the correct Standard form for 359 words, while the same footnote states that Standard writes nothing where Uthmani has U+0670 in 362 words. Since deleting the character should succeed precisely where Standard writes nothing, absent other divergences, these counts appear to describe nearly the same set yet differ by three. A clause reconciling them would remove an apparent inconsistency in the paper's opening empirical claim.

2. Table 1's sixth row is labelled "Word-initial rules," whereas §4.1 describes step 6 as alef-madda resolution and argues explicitly that the rule applies "wherever it occurs and not only word-initially." As Table 1 is the paper's only worked example of the pipeline, the mismatch is likely to confuse readers tracing the correspondence between table and prose.

3. The coverage-weighted scoring formula is referenced but never stated. Section 4.2 says candidates are scored by a formula "that favors short verses where the query fills a large fraction of the verse text," and §5.2 attributes both failures to how candidates are ranked. The formula is therefore load-bearing for the failure analysis and should appear explicitly rather than only in the repository.

4. The multi-verse alignment mode is underspecified. Inputs over eight normalized words trigger a separate mode that "segments and scores each constituent verse independently," and case ME03 turns on how that mode selects its anchor and when it falls back. The segmentation procedure, fallback condition, and partial-alignment scoring deserve the level of detail given to the four layers.

5. No runtime figures are reported, although the paper emphasises deployability and layer 4 scores all 6,236 verses by sliding-window similarity using the standard library. Median and tail latency per query, together with how often each layer is reached, would substantiate "immediately deployable" and would also quantify the cost of the determinism argument made in §4.2.

6. The character range given for stripping Arabic Extended-A marks, U+08D3 to U+08FF, is broader than the combining marks it is described as covering, since the block also contains assigned non-mark characters. Stating whether the implementation filters by codepoint range or by Unicode combining category would clarify whether any letters are at risk.

7. Two small items of residue are described but not shown: the single U+0670 word form excluded for admitting no consistent span alignment, and the two keys resolved by majority vote. Naming them, with the competing Standard forms and their counts, would occupy a few lines and would let readers judge whether majority vote was appropriate or whether the variation was contextually conditioned.

8. The CC-BY 4.0 release of the mapping is not reconciled with the source corpus's terms of use, which permit verbatim copying and distribution but state that changing the text is not allowed and require the source to be clearly indicated with a link. Since CC-BY permits adaptation, a sentence addressing compatibility would protect downstream users. The annotation corpus's licence is likewise unnamed.

9. Transliteration is inconsistent and no romanisation scheme is stated, with forms such as yaqawmi, al-kafirin and Qul huwa llahu ahad mixing conventions. Adopting one documented scheme and giving the Arabic script alongside it for the examples in §3.1 and §5.2 would make the linguistic claims checkable by readers who cannot consult the repository.

10. The deployed system and its privacy policy are not identified, which makes the claims in §7 difficult to assess, and the retention, access and pseudonymisation arrangements for the 1,407 utterances are unstated. A brief characterisation of the ASR model would also help, since whether it is off-the-shelf, Quran-specific, or fine-tuned in-house materially affects what §5.4 transfers to.

Missing and Inadequately Integrated Literature

Cited but Inadequately Integrated

Tanzil Project (2024): The corpus underpins every count in §3 and §4.1 but is cited only as a project name, a year and a URL. The project documents multiple Simple-family variants, configurable download options, and a versioned release history, all of which determine the contents of the released mapping, and it describes a production process in which both the Imlaei and Uthmani texts were derived automatically from one core text. The latter bears directly on whether the alignment of §3 is descriptive in the sense §2 claims. The authors should cite the specific variant, option set and text version, and should address the provenance question explicitly.

Habash et al. (2012): CODA is the paper's main theoretical foil and the prescriptive-versus-descriptive contrast is well drawn, but the engagement stops short of the methodological lessons. CODA's development involved explicit guideline documentation and annotator agreement measurement precisely because orthographic normalization decisions are contestable, and that line of work also treats the word-boundary mismatches behind this paper's 58 one-to-many entries. The authors should address why an alignment-derived resource is exempt from the validation practices the work they position against found necessary.

Khan et al. (2021): Tarteel is described as addressing "the complementary half of the problem," which is accurate, but the complementarity is not exploited. Tarteel releases crowd-sourced recitation audio paired with its text, which makes it a natural source of independent, non-synthetic evaluation material for exactly the degraded-input regime §5.4 says the paper lacks, and the project has also had to confront the consent and governance questions §7 raises from first principles.

Dukes (2011): The Quranic Arabic Corpus supplies the lemma and root annotation on which layer 2 depends, yet no case in Table 2 is attributed to that layer and the paper never reports how often morphological expansion changes an outcome. The orthographic and tokenization interface between the two resources, which must be reconciled before any lookup can succeed, is likewise not described.

Habash (2010) and Darwish and Magdy (2014): Both are cited in a single sentence as documenting root-based morphology and diacritization as central challenges, which understates their relevance. Both treat orthographic normalization for Arabic information retrieval directly, including the alef and ya normalizations that steps 5 and 6 implement and the recall-versus-precision trade-off those operations entail. Engaging with that treatment would supply much of the justification the normalization rules currently lack.

Truly Missing Papers

Pasha et al. (2014): Not cited. MADAMIRA is the field's reference implementation for Arabic morphological analysis, disambiguation and orthographic normalization, and its normalization and tokenization schemes are the natural comparison point for the decisions made in steps 5 and 6. Including it would also make the toolkit survey in footnote 1 representative rather than partial. [Verified via WebSearch: LREC 2014, pages 1094–1101.]

Alqahtani and Atwell (2016): Not cited. This work reviews and classifies search techniques applied to the Quran, documents the deficiencies of keyword-based matching on this corpus, and proposes a framework with explicit matching and ranking modules. It is the closest published treatment of the retrieval problem the validator solves, and its absence is a principal reason the priority claim is currently unsupported. [Verified via WebSearch: NLDB 2016, LNCS 9612.]

Bashir et al. (2023): Not cited. A systematic review of Arabic NLP for Qur'anic research covering sixty-five works, which would provide the evidential basis that a repeated "first published" claim requires, and would also supply domain-appropriate conventions for how resources in this area are validated and reported. [Verified via WebSearch: Artificial Intelligence Review 56(7):6801–6854.]

Sharaf and Atwell (2012): Not cited. QurSim is a gold-standard corpus of expert-annotated related verse pairs over this exact text and is the established benchmark for verse-level matching. It bears both on the ambiguity census, offering an independent characterisation of which verses are genuinely confusable, and on the paper's need for evaluation material it did not author itself. [Verified via WebSearch: LREC 2012, pages 2295–2302.]

Alrabiah et al. (2013): Not cited. KSUCCA is a fifty-million-word corpus of Classical Arabic spanning the period of revelation and the following centuries, and is the most obvious external source against which a sample of the 2,290 mapped Standard forms could be independently checked — the validation the mapping currently lacks. [Verified via WebSearch: WACL-2, Lancaster, pages 5–8.]

González Martínez (2025): Not cited. This work analyses the orthographic layer of tajwid in Contemporary Quranic Orthography and presents a module that strips and reinserts it, addressing directly the question of which orthographic marks carry recoverable information and which can be removed without loss. It is the closest existing treatment of the information-destruction question raised in 3.1. [Verified via WebSearch: arXiv:2505.11379.]

Gebru et al. (2021): Not cited. "Datasheets for Datasets" sets out the documentation standard — motivation, composition, collection process, intended uses, limitations — that a released language resource is now generally expected to meet, and following it would surface several of the unstated preconditions identified in 1.3. [Verified via WebSearch: Communications of the ACM 64(12):86–92.]

References

Alqahtani, M., and E. Atwell. 2016. Arabic Quranic Search Tool Based on Ontology. In Natural Language Processing and Information Systems (NLDB 2016), Lecture Notes in Computer Science 9612:478-485. Springer.
Alrabiah, M., A. Al-Salman, and E. Atwell. 2013. The design and construction of the 50 million words KSUCCA King Saud University Corpus of Classical Arabic. In Proceedings of WACL-2: Second Workshop on Arabic Corpus Linguistics, pages 5-8, Lancaster, UK.
Bashir, M. H., A. M. Azmi, H. Nawaz, W. Zaghouani, M. Diab, A. Al-Fuqaha, and J. Qadir. 2023. Arabic natural language processing for Qur'anic research: a systematic review. Artificial Intelligence Review 56(7):6801-6854.
Gebru, T., J. Morgenstern, B. Vecchione, J. W. Vaughan, H. Wallach, H. Daume III, and K. Crawford. 2021. Datasheets for datasets. Communications of the ACM 64(12):86-92.
Gonzalez Martinez, A. 2025. A computational system to handle the orthographic layer of tajwid in contemporary Quranic Orthography. arXiv:2505.11379.
Pasha, A., M. Al-Badrashiny, M. Diab, A. El Kholy, R. Eskander, N. Habash, M. Pooleery, O. Rambow, and R. Roth. 2014. MADAMIRA: A Fast, Comprehensive Tool for Morphological Analysis and Disambiguation of Arabic. In Proceedings of LREC 2014, pages 1094-1101, Reykjavik, Iceland.
Sharaf, A.-B., and E. Atwell. 2012. QurSim: A corpus for evaluation of relatedness in short texts. In Proceedings of LREC 2012, pages 2295-2302, Istanbul, Turkey.

Conclusion and Path Forward

The problem this manuscript identifies is genuine, the artifact it releases is likely to be reused, and the authors' willingness to report inconvenient findings — the declined 0.65 threshold, the itemised residual verses, the acknowledged selection bias — suggests the revisions requested here are well within reach. The distance to publication is accounted for by a consistent pattern rather than by isolated errors: the claims that matter most are supported by procedures that cannot falsify them. The mapping's Standard side appears to derive from a text machine-generated alongside its Uthmani counterpart; the convergence figure is computed on the data the mapping was induced from; the suite accuracy, on the authors' own evidence, exercises only half the search architecture; the deployment labelling criterion is close to the matching condition it is meant to test independently; and the ambiguity census quantifies a relation that neither documented failure instantiates. Alongside these, the user-facing feedback component is unevaluated, no baseline or ablation isolates what the mapping contributes, and the architecture is never weighed against indexing the Standard text the same provider publishes.

A concrete programme would address most of this without new systems or new data. Building the mapping on a held-out portion and reporting convergence on the remainder would turn a coverage statistic into a generalisation estimate; validating a stratified sample of pairs against a source outside the Tanzil family, with two independent readers, would supply the human check the resource currently lacks. Replacing step 2 with a global delete and rerunning the suite would isolate the mapping's end-to-end contribution, and sweeping the generator's corruption rate upward would finally exercise layers 3 and 4; scoring the reported correction lists against the generator's injected errors would do the same for the feedback tiers, and hand-labelling the 59 imperfect high-overlap deployment transcripts would give the paper its first real measurement on degraded input. Recomputing the ambiguity census over raw, diacritized and normalized text, and adding a superset census under layer 1's matching condition, would both correct an overclaim and bound the known failure mode. Finally, a tagged and archived release with pinned inputs, a documented search behind the priority claim, and confidence intervals on the reported proportions would make the work citable with confidence. The research question is worth answering and the design judgement on display is sound; what the manuscript needs is an evaluation capable of telling the reader how well it has been answered.

Author response

I thank the reviewer for a careful and constructive report. I accept its central point: several key claims are currently supported by procedures that cannot falsify them, and the revision will address this directly.

Specifically, v2 will: (1) state the exact Tanzil variants, options and text version used, address explicitly whether the Uthmani and Standard texts are independent, and validate a stratified sample of mapping pairs against a source outside the Tanzil family; (2) induce the mapping on a held-out portion of the corpus and report cross-form convergence on the remainder; (3) add an ablation replacing step 2 with global delete, global keep, and no mapping, plus a simple term-weighted baseline; (4) sweep the injected corruption rate so that layers 3 and 4 are exercised, report per-layer firing counts, and add confidence intervals to all proportions; (5) score the feedback tiers against injected errors and explain the ME03 tier outcome; (6) hand-label the 59 imperfect high-overlap deployment transcripts; (7) recompute the ambiguity census over raw, diacritized and normalized text and add a superset census under layer 1's condition; and (8) discuss indexing the Standard text directly as an alternative architecture.

Related work will add Quranic IR and the cited resources (MADAMIRA, QurSim, Alqahtani & Atwell, Bashir et al., González Martínez, KSUCCA, Datasheets), and the release will be tagged and archived with pinned inputs. All minor issues, including Table 1, footnote 1 and licensing, will be corrected.