The paper proposes Parallel Science, an open infrastructure intended to host and curate a dedicated AI-generated scientific literature that runs in parallel to (but interoperates with) the human literature. The concrete implementation described centers on Parallel ArXiv (papers.parallelscience.org), which ingests AI-generated papers from GitHub Pages, assigns stable PX identifiers (PX:YYMM.NNNNN), versions papers using content hashing, and classifies them into an arXiv-compatible taxonomy via a two-step LLM-based classifier (Secs. 3–5). The manuscript also outlines a production stack of supervised and autonomous AI “scientist” fleets (Denario/CosmoEvolve) with containerized deployments across multiple providers plus monitoring/cost accounting (“Mission Control”), and introduces Parallel Open Review (reviews.parallelscience.org) as an AI review layer envisioned to connect to replication engines and drive a production–evaluation–selection loop (Secs. 4, 7–8). Overall, the systems/institutional framing is timely and the Parallel ArXiv mechanics are described with useful implementation detail, but the review/replication loop, governance/safety posture, security/threat model, and quantitative evaluation of system behavior are currently under-specified relative to the central claims, making it hard to assess robustness, quality control effectiveness, and real-world scalability.

• ✔ Large language models and multi-agent systems are now capable of generating scientific hypotheses, writing and executing code, analyzing data, and producing coherent research papers, as demonstrated in recent work on fully automated scientific discovery, autonomous chemical research, and benchmarking language agents on machine learning experimentation.
• _Reference(s):_ Lu et al., 2024, Boiko et al., 2023, Huang et al., 2024
• _Justification:_ Verification failed with gpt-5: Error code: 400 - {'error': {'message': 'Your input exceeds the context window of this model. Please adjust your input and try again.', 'type': 'invalid_request_error', 'param': 'input', 'code': 'context_length_exceeded'}}

• ⚠ The research backend CMBAgent implements a two-phase planning-and-control strategy in which a planner agent designs an execution plan and a controller agent delegates subtasks to specialized agents, and is built on prior work on multi-agent systems for cosmological parameter analysis and open-source planning and control systems with language agents for autonomous scientific discovery, orchestrated using AG2 for multi-agent conversation patterns and LangGraph for structured generation tasks.
• _Reference(s):_ Laverick et al., 2024, Xu et al., 2025, Wu et al., 2023
• _Justification:_ Xu et al., 2025 describes cmbagent as a two-phase Planning & Control system: a planner designs a multi-step plan and, in the Control phase, a controller delegates each sub-task to specialized agents such as the researcher and engineer (Xu et al., 2025). It is powered by AG2/autogen (Xu et al., 2025; Laverick et al., 2024), and builds on earlier MAS work for cosmological parameter analysis (Laverick et al., 2024). However, while LangGraph is used in the denario system to turn results into manuscripts (a separate subsystem where cmbagent is the backend), the papers do not show cmbagent itself being orchestrated with LangGraph (Xu et al., 2025). Thus the claim about LangGraph within cmbagent is not directly supported, making the statement only partially supported.

• ✔ The Denario research pipeline used in this infrastructure is based on the Denario project framework for deep knowledge AI agents for scientific discovery, originally described by Villaescusa-Navarro et al. (2025), which introduced a dual-backend architecture that the present work extends with iterative refinement across multiple research cycles and fleet-scale containerized deployment.
• _Reference(s):_ Villaescusa-Navarro et al., 2025
• _Justification:_ Verification failed with gpt-5: Error code: 400 - {'error': {'message': 'Failed to download file. File urls cannot be larger than 32MB.', 'type': 'invalid_request_error', 'param': 'url', 'code': 'file_above_max_size'}}

• ✖ The CosmoEvolve Virtual Lab system, which operates through the same CMBAgent stack and OpenClaw gateway described here, builds on prior work on multi-agent systems for cosmological parameter analysis to perform autonomous cosmological data analysis and will be described in detail in a forthcoming publication referenced as a multi-agent system for cosmological parameter analysis (2024).
• _Reference(s):_ Laverick et al., 2024
• _Justification:_ Laverick et al., 2024 describes the cmbagent/CMBAgent multi-agent system and its architecture but does not mention any 'CosmoEvolve Virtual Lab' or an 'OpenClaw gateway.' It also states the current system requires human feedback at every step rather than being autonomous. No forthcoming publication about a CosmoEvolve system is referenced. Thus, the statement is not supported by the paper.

This section audits symbolic/analytic mathematical consistency (algebra, derivations, dimensional/unit checks, definition consistency).

Maths relevance: light

The paper is primarily a systems/infrastructure description with very limited mathematics: identifier-format notation, hashing/versioning rules, occasional complexity notation ($O(1)$), and brief summaries of external example papers that mention standard statistical/ML quantities (PCA variance explained, $R^2$, negative log-likelihood minimization). There are no displayed equations or derivations to audit for algebraic correctness.

### Checked items

• ✔ PX identifier format and example consistency (Sec. 3.2, p.3 (parsed pages 3/12))

• Claim: Identifiers have the format PX:YYMM.NNNNN, with an example PX:2604.00001 described as the first paper indexed in April 2026.
• Checks: notation consistency, constraint/sanity check
• Verdict: PASS; confidence: high; impact: minor
• Assumptions/inputs: YYMM corresponds to year (two digits) and month (two digits)., NNNNN is a zero-padded sequence number within that month.
• Notes: The example PX:2604.00001 matches April 2026 under the stated YYMM convention, and the prefix PX is used consistently as the Parallel ArXiv distinguisher.

• ✔ Append-only registry + sequence table logic (Sec. 3.2, p.3 (parsed pages 3/12))

• Claim: An append-only registry maps repository names to identifiers, and an ID sequence table provides $O(1)$ allocation of the next available number within each month.
• Checks: logic consistency, asymptotic-claim plausibility (non-numeric)
• Verdict: PASS; confidence: medium; impact: minor
• Assumptions/inputs: A per-(year,month) counter/sequence is maintained in the database., Allocation increments the stored counter and returns the previous/new value deterministically.
• Notes: No internal contradictions: an explicit per-month sequence record can yield constant-time lookup/update. Formal proof is not provided, but the claim is consistent with the described data structure.

• ✔ Content-hash versioning rule consistency (Sec. 3.2, p.3 (parsed pages 3/12))

• Claim: A SHA-256 hash of title, author, abstract, and categories is used to detect content changes; hash changes trigger a new version and prevent spurious version bumps when rebuilding pages without content changes.
• Checks: definition consistency, logic/sanity check
• Verdict: PASS; confidence: high; impact: minor
• Assumptions/inputs: Hash input fields are exactly those listed (title, author, abstract, categories)., Rebuilds that do not alter these fields should keep the hash stable.
• Notes: Given the stated hashed fields, rebuild-only changes (e.g., HTML regeneration) would not affect the hash, matching the stated goal. The rule is self-consistent as described.

• ⚠ Turbulent-fluid example: grid-size notation and method chain (Sec. 6.1, p.8 (parsed pages 8/12))

• Claim: Example paper uses density/velocity fields on a “1283” periodic grid over 10 time slices; computes spatial derivatives spectrally and temporal derivatives via finite differences; builds a 66-term library; applies cross-validated LASSO with OLS refinement; reports $R^2$ and stability under integration.
• Checks: notation clarity, workflow logical consistency
• Verdict: UNCERTAIN; confidence: medium; impact: minor
• Assumptions/inputs: “1283” is intended to mean a 3D grid with 128 points per dimension ($128^3$)., The pipeline description is a narrative summary rather than a derivation.
• Notes: The chain of methods is narratively coherent, but the symbol-like token “1283” is ambiguous ($128^3$ vs 1,283). This ambiguity prevents a clean symbolic interpretation, though it does not create a direct contradiction elsewhere in this paper.

• ⚠ VCSF anomaly score definition (min over parameter space) (Sec. 6.2, p.8 (parsed pages 8/12))

• Claim: Anomaly scores are computed as the minimum negative log-likelihood over the parameter space, found via gradient-based optimization, after PCA and whitening against nuisance parameters and modeling with a conditional normalizing flow.
• Checks: definition consistency, missing-symbol audit
• Verdict: UNCERTAIN; confidence: low; impact: minor
• Assumptions/inputs: A conditional likelihood $p(x|\theta)$ (or similar) is defined by the flow, and the anomaly score is $s(x)= \min_\theta [-\log p(x|\theta)]$., The “parameter space” refers to the nuisance/conditioning variables referenced earlier in the paragraph.
• Notes: The minimization-based score is plausible and internally consistent at the narrative level, but the paper does not introduce symbols/domains (what is optimized over, constraints, differentiability assumptions) needed to verify formal correctness of the objective and its relationship to “invariance to known physical parameter variations.”

### Limitations

• The manuscript contains no displayed equations, numbered formulas, or step-by-step derivations; therefore an algebraic/derivational consistency audit is largely inapplicable.
• Sections 6.1–6.3 summarize other papers’ methodologies and metrics without providing the underlying definitions/equations in this PDF, so only surface-level notation/logic checks are possible.
• No unit/dimensional analysis is feasible because the paper does not define physical quantities symbolically (it discusses infrastructure rather than presenting physical models).

This section audits numerical/empirical consistency: reported metrics, experimental design, baseline comparisons, statistical evidence, leakage risks, and reproducibility.

Of 20 candidate numerical/consistency checks, 15 passed and 5 were uncertain; no failures were detected. Verified items include exact parsing/format checks (identifier example), exact arithmetic/unit conversions (seconds$\leftrightarrow$minutes; polling period), exact integer recomputations ($128^3$ grid and totals), simple inequality/bounds checks ($R^2$ range, percentage and metric bounds, $\sigma$ threshold comparison, frequency ordering), and a combinatorial parts-to-total index coverage check for $N=12$ in Table 1. Uncertain items require additional document text/table extraction or external quantities (e.g., DB size vs paper count; document-wide repeated-constant scans; cross-table identifier matching).

### Checked items

• ✔ C1 (Page 3, Section 3.2 (Identifier scheme and versioning))

• Claim: Identifier format PX:YYMM.NNNNN; example: "PX:2604.00001 for the first paper indexed in April 2026".
• Checks: format_consistency (example parsing vs claimed meaning)
• Verdict: PASS
• Notes: Parsed YYMM and sequence compared to claimed meaning.

• ✔ C2 (Page 3, Section 3.3 (Category classification))

• Claim: Classification selects one of "21 top-level archives" and assigns "up to three" secondary categories for cross-listing.
• Checks: integer_reasonableness / bounds check from stated limits
• Verdict: PASS
• Notes: Constraint constants checked; schema-level validation only.

• ⚠ C3 (Page 4, Section 3.4 (Database and storage))

• Claim: “The database is compact, at approximately 2 KB per paper.”
• Checks: cheap_recomputation (order-of-magnitude check given paper counts if available)
• Verdict: UNCERTAIN
• Notes: Need number_of_papers and/or total DB size to recompute per-paper storage.

• ✔ C4 (Page 6, Table 1 caption + rows (Supervised fleet configuration))

• Claim: Table 1 describes supervised denario-$i$ fleet with $N = 12$ and lists scientist indices $1,4,5$; $2$; $3$; $6$; $7$–$12$.
• Checks: parts_to_total (index coverage equals $N$)
• Verdict: PASS
• Notes: Checked that listed indices/range cover exactly $1..N$ with no omissions/extras.

• ✔ C5 (Page 6, Table 1 (Scientist 3 row))

• Claim: Scientist 3 has "CPU 32", "RAM 64 GB", "GPU RTX PRO 6000", "Timeout 1800 s".
• Checks: unit_consistency (time conversion) + sanity ratio
• Verdict: PASS
• Notes: Converted seconds to minutes and compared.

• ✔ C6 (Page 6, Table 1 (most rows))

• Claim: Several rows list "Timeout 300 s".
• Checks: unit_consistency (time conversion) + repeated constant check
• Verdict: PASS
• Notes: Converted seconds to minutes and compared.

• ✔ C7 (Page 6, Section 4.2 (Fleet configuration and isolation))

• Claim: Workstation specs include "two NVIDIA RTX PRO 6000 Blackwell GPUs (96 GB VRAM each)" and system RAM "512 GB".
• Checks: parts_to_total (GPU aggregate VRAM)
• Verdict: PASS
• Notes: Computed total VRAM and checked optional inequality RAM $>$ total VRAM.

• ✔ C8 (Page 8, Section 6.1 (Governing equation discovery))

• Claim: Grid is "$128^3$" across "10 time slices".
• Checks: cheap_recomputation (power and multiplication)
• Verdict: PASS
• Notes: Exact integer recomputation of $128^3$ and multiplication by time slices.

• ⚠ C9 (Page 8, Section 6.1 (Governing equation discovery))

• Claim: “constructed a library of 66 candidate terms”.
• Checks: repeated_constant_cross_check (within-document references)
• Verdict: UNCERTAIN
• Notes: Within-document cross-occurrence check requires full document text; not available in PAYLOAD.

• ✔ C10 (Page 8, Section 6.1 (Governing equation discovery))

• Claim: $R^2$ is reported as a range: "$R2 = 0.57$–$0.71$".
• Checks: range_validity (bounded metric)
• Verdict: PASS
• Notes: Checked metric bounds and ordering.

• ✔ C11 (Page 8, Section 6.2 (Simulation mismatch detection))

• Claim: "top 3 components capturing $97.35\%$ of variance".
• Checks: percentage_bounds
• Verdict: PASS
• Notes: Checked percentage bounds.

• ✔ C12 (Page 8, Section 6.2 (Simulation mismatch detection))

• Claim: Partial AUC reported as "$0.1488$".
• Checks: metric_bounds
• Verdict: PASS
• Notes: Checked metric bounds $[0,1]$.

• ⚠ C13 (Page 8, Section 6.2 (Simulation mismatch detection) vs Page 6 Table 1)

• Claim: Section 6.2 says the paper was produced by "denario-3 (Claude Sonnet 4.6 with GPU access)"; Table 1 lists Scientist 3 with GPU "RTX PRO 6000" and model "Claude Sonnet 4.6".
• Checks: cross_section_entity_consistency
• Verdict: UNCERTAIN
• Notes: Cross-section entity consistency requires extracted Table 1 text mapping scientist$\to$(model,gpu); not available beyond isolated fields.

• ✔ C14 (Page 8, Section 6.3 (Multi-frequency analysis of tSZ maps))

• Claim: Blind source detection yielded "200 candidates above $5\sigma$"; Bullet Cluster recovered at "$49\sigma$".
• Checks: unit-consistent comparison (sigma thresholds)
• Verdict: PASS
• Notes: Checked sigma inequality and that count is a non-negative integer.

• ✔ C15 (Page 8, Section 6.3 (Multi-frequency analysis of tSZ maps))

• Claim: Spectral diagnostics across "$90$, $150$, and $220$ GHz" bands.
• Checks: count_and_order check
• Verdict: PASS
• Notes: Checked count=$3$ and strict increasing order.

• ✔ C16 (Page 8, Section 6.3 (Multi-frequency analysis of tSZ maps))

• Claim: “This 18-page paper with 16 figures …”
• Checks: integer_relation (figures $\leq$ pages)
• Verdict: PASS
• Notes: Sanity inequality figures $\leq$ pages.

• ✔ C17 (Page 8, Section 7 (Monitoring and Cost Transparency))

• Claim: Dashboard polls the supervised fleet "every 10 seconds".
• Checks: unit_consistency (frequency conversion)
• Verdict: PASS
• Notes: Converted polling period to polls/min and compared.

• ✔ C18 (Page 9, Table 3 (Per-paper costs))

• Claim: Table 3 lists three papers with total costs \$0.61, \$2.40, and \$4.10.
• Checks: cheap_recomputation (sum, mean, min/max)
• Verdict: PASS
• Notes: Computed sum/mean/min/max; diff_* refers to sum vs $7.11$. Mean checked within $0.01$ as instructed.

• ⚠ C19 (Page 9, Table 3 (Per-paper costs) vs Page 8, Sections 6.1–6.3)

• Claim: The three example papers in Sections 6.1–6.3 correspond to PX:2604.00016, PX:2604.00009, PX:2604.00015, which appear in Table 3.
• Checks: cross_section_identifier_match
• Verdict: UNCERTAIN
• Notes: Requires Table 3 paper IDs to compare sets; PAYLOAD does not include Table 3 IDs (only costs).

• ⚠ C20 (Page 7, Table 2 + Page 6, Section 4.2)

• Claim: Table 2 lists 3 connected systems; text states supervised denario-$i$ fleet runs $N = 12$ scientists.
• Checks: cross_table_consistency (counts stated in different locations)
• Verdict: UNCERTAIN
• Notes: Document-wide duplicate-number consistency requires full document text; not available in PAYLOAD.

### Limitations

• Only parsed text provided from the PDF was used; no additional PDF structure (e.g., exact table cell boundaries) was available beyond the text transcript.
• No checks rely on external URLs, repositories, datasets, runtime logs, or executing the described systems; such claims are listed as unverified.
• No figure value extraction was performed; checks avoid reading plot pixels or inferring quantitative values from images.
• Some candidates (e.g., repeated-constant scans) are only actionable if you run text-search over the full PDF text; the provided transcript may omit formatting nuances.
• Some checks were uncertain because required quantities or structured table content were not available in the provided payload (e.g., DB size vs number of papers; Table 3 paper IDs for cross-matching; Table 1 scientist-to-(model,GPU) mapping for cross-section validation; full-document text for duplicate-number consistency).

## Paper Ratings

| Dimension | Score |

|-----------|:-----:|

| Overall | 6/10 ██████░░░░ |

| Soundness | 6/10 ██████░░░░ |

| Novelty | 8/10 ████████░░ |

| Significance | 6/10 ██████░░░░ |

| Clarity | 7/10 ███████░░░ |

| Evidence Quality | 4/10 ████░░░░░░ |

Justification: The paper presents a timely and original institutional/system design for a parallel AI-generated literature with concrete implementation details for Parallel ArXiv, identifier/versioning, and fleet operations. Mathematical and numerical audits found no major inconsistencies, but statement verification flagged at least one unsupported claim (CosmoEvolve reference), and several core components of the claimed production–evaluation–selection loop (AI review, replication engine, governance/security) are under-specified. The evidence base is limited to case studies and cost snapshots without systematic operational metrics, classification accuracy, review calibration, or reliability data, which constrains confidence in robustness and scalability. Strong conceptual novelty and clear exposition are offset by gaps in quantitative evaluation, threat modeling, and governance, leading to an overall solid but not yet top-tier assessment.