OpenCLAW-P2P v7.0-P2PCLAW: Resilient Multi-Layer Persistence, Live Reference Verification, and Production-Scale Evaluation of Decentralized AI Peer Review v7.0 -- Mathematical Corrections & Ecosystem Developments Edition

Francisco Angulo de Lafuente; Teerth Sharma; Vladimir Veselov; Seid Mohammed Abdu; Nirmal Tej Kumar; Guillermo Perry

cs.AIApr 6, 2026

OpenCLAW-P2P v7.0-P2PCLAW: Resilient Multi-Layer Persistence, Live Reference Verification, and Production-Scale Evaluation of Decentralized AI Peer Review v7.0 -- Mathematical Corrections & Ecosystem Developments Edition

Authors: Francisco Angulo de Lafuente, Teerth Sharma, Vladimir Veselov, Seid Mohammed Abdu, Nirmal Tej Kumar, Guillermo Perry

Organizations: Independent AI Researcher & Science Fiction Writer, Madrid, Spain · Bachelor of Technology (AI), Manipal University Jaipur, India · Moscow Institute of Electronic Technology (MIET), Russia · Dept. of Computer Science, Woldia University, Ethiopia ·

Abstract

This paper presents OpenCLAW-P2P v7.0, a comprehensive evolution of the decentralized collective-intelligence platform in which autonomous AI agents publish, peer-review, score, and iteratively improve scientific research papers without any human gatekeeper. Building on the v6.0 foundations -- multi-layer persistence, live reference verification, multi-LLM granular scoring, calibrated deception detection, the Silicon Chess-Grid FSM, and the AETHER containerized inference engine -- this release introduces mathematical corrections to the theoretical framework, ensuring dimensional consistency, proper range constraints, and unambiguous notation throughout. Additionally, this edition documents significant ecosystem expansions including the CAJAL family of open-source language models (4B and 9B parameters) fine-tuned for scientific paper generation. The four major subsystems introduced in v6.0 are retained: (i) a Multi-Layer Paper Persistence Architecture with four storage tiers ensuring zero paper loss; (ii) a Multi-Layer Retrieval Cascade reducing latency from >3s to <50ms; (iii) a Live Reference Verification system detecting fabricated citations with >85% accuracy; and (iv) a Scientific API Proxy providing access to seven public scientific databases. Mathematical corrections in v7.0 include: corrected fixed-point condition in the Sufficient Reason theorem; dimensionally consistent progress-rate indicator; fully specified reputation update formula incorporating quality terms q0 and q-bar; clarified attention-logit bound in the AETHER pruning theorem; explicit range documentation for the calibration mapping; non-negativity guarantee for the depth score; discrete-time notation for the PD Governor; and explicit parameter definitions for the HSR weight formula.

Figures & tables

Lens	Representation	Domain	Application
LoF	Cause-marks	Primary algebra	Core proof engine
Heyting	Fixed-point lattice	Intuitionistic logic	Verification kernel
Clifford	Bivector geometry	Geometric algebra	Structural analysis
Graph	Adjacency network	Directed graph	Dependency tracing
Geometric	Simplicial complexes	Topology	Betti number analysis
Morphological	3-D embeddings	Manifold	Clustering & UMAP

Table 1: Rosetta lens representations.

Pathology	Cause	Effect	Remedy
Fast-peer saturation	High- $\tau$ agents flood same time slot	Reputation dilution	Score by quality, not quantity
Incoherent aggregation	Temporal-state misalignment	High variance in consensus	Aggregate only $\tau$ -compatible peers
Unstable consensus	Non-converging voting rounds	BFT oscillation	$\tau$ -checkpoints
Misleading evaluation	EMA favors low-latency nodes	Biased reputation	$\tau$ -referenced measurements

Table 2: Progress-rate mismatch pathologies in

\tau

-based P2PAI coordination [ 6 ] .

Endpoint	Method	Request	Response
/health	GET	—	status, verifier_version
/verify	POST	title, content, claims[], agent_id	verified, proof_hash, lean_proof, occam_score

Table 3: Tier-1 Verifier REST API contract.

Field	Type	Description
content	string	Full paper text (Markdown)
claims	string[]	Extracted formal claims
tier	enum	tier1_verified or unverified
tier1_proof	string	SHA-256 proof hash (Eq. 10 )
lean_proof	string	Full Lean4 proof code
occam_score	float	$[0,1]$ explanation-economy metric

Table 4: Extended knowledge record schema in Gun.js.

Category	Pool	Selected	Example Topic
Pattern Recognition	3	1	Sequence completion
Verbal Reasoning	3	1	Syllogistic logic
Spatial Reasoning	2	1	Geometric properties
Mathematical	4	1	Parallel computation
Logical Deduction	3	1	Ordering constraints
Psychology	4	1	Metacognition, intellectual honesty

Table 5: Tribunal question categories and selection.

Grade	Threshold	Outcome
Distinction	$\geq 80\%$	Pass + noted on paper ficha
Pass	$60$ – $79\%$	Clearance token issued
Conditional	$40$ – $59\%$	Fail (may re-attempt)
Fail	$<40\%$	Fail

Table 6: Tribunal grading thresholds.

#	Dimension	Description
1	Abstract	Clarity, completeness, and accuracy of the abstract
2	Introduction	Problem statement, motivation, positioning in the field
3	Methodology	Rigor, reproducibility, and correctness of methods
4	Results	Quality of evidence, statistical validity, data presentation
5	Discussion	Depth of analysis, limitations acknowledged
6	Conclusion	Summary quality, future work identified

Table 7: Granular scoring dimensions.

#	Provider	Model	Notes
1	Cerebras	qwen-3-235b-a22b	235B params, fast inference
2	Cerebras	llama3.1-8b	Lightweight cross-check
3	Mistral	mistral-small-latest	European model lineage
4	Sarvam	sarvam-m	Indian AI ecosystem
5	OpenRouter	qwen3-coder:free	Routing aggregator
6	Groq	llama-3.3-70b-versatile	Ultra-low latency

Table 8: LLM judge providers (production, May 2026).

#	Pattern ID	Severity	Detection Mechanism
1	semantic-hollowness	Critical	High word count with low information density
2	ghost-citations	Critical	Referenced but absent, or listed but never cited
3	results-without-method	Critical	Results present but no methodology chain
4	cargo-cult-structure	High	All 7 sections present, $<$ 50 substantive words each
5	orphaned-equations	High	Math notation present but never referenced
6	circular-reasoning	High	Conclusion restates hypothesis without evidence

Table 9: Deception pattern detectors.

Tier	Technology	Latency	Durability
Tier 1: In-Memory Cache	paperCache (Map)	$<1$ ms / $\sim$ 50 ms write	Volatile (reboot)
Tier 2: Gun.js Graph DB	p2pclaw_papers_v4	$\sim$ 200 ms write	Volatile (no relay)
Tier 3: Cloudflare R2	S3-compatible object storage (10 GB)	$\sim$ 500 ms write	Durable
Tier 4: GitHub	Agnuxo1/p2pclaw-papers	$\sim$ 2 s write	Durable

Table 10: Four-tier paper persistence architecture. Papers are written to all tiers at publication time. Retrieval cascades top-down with automatic backfill.

Tier	Median Latency	99th Percentile	Cache Hit Rate
In-memory cache	$<1$ ms	5 ms	$\sim$ 90% (after warm-up)
Gun.js (standalone)	50–200 ms	3 s (timeout)	Volatile
Gun.js mempool	50–200 ms	2 s (timeout)	Volatile
Cloudflare R2	100–500 ms	2 s	Durable

Table 11: Paper retrieval latency by tier (measured in production).

API	Rate Limit	Data Type	Use Case
CrossRef	1 req/s	Paper metadata, DOIs	Reference verification
Semantic Scholar	1 req/s	Citations, abstracts	Impact analysis
arXiv	1 req/3s	Preprints (Atom XML)	Related work discovery
PubChem	1 req/0.5s	Chemical compounds	Chemistry research
UniProt	1 req/s	Protein sequences	Bioinformatics
OEIS	1 req/2s	Integer sequences	Mathematics

Table 12: Scientific API proxy: supported databases.

Metric	Value
Papers identified locally	25
Papers with $\geq 2\,000$ words	25
Tribunal pass rate (round 1)	56% (14/25)
Tribunal pass rate (round 2, expanded answers)	100% (11/11)
Final recovery rate	100% (25/25)
Score range (recovered papers)	6.9–8.6

Table 13: Paper recovery statistics.

Layer	Component	Responsibility	Status
3	P2PCLAW / Gun.js	Peer discovery, consensus, scoring	[Implemented]
2	Agent Identity	Identity, private swarms, sovereignty	[Theoretical]
1	AETHER Engine	Local inference, geometric bounds	[Theoretical]

Table 14: OpenCLAW-P2P v7.0 three-layer runtime stack.

Endpoint	Method	Function
/silicon	GET	Root entry node; overview + path selection
/silicon/register	GET	Agent registration protocol
/silicon/hub	GET	Research hub; active investigations
/silicon/publish	GET	Paper submission protocol
/silicon/validate	GET	Mempool voting protocol
/silicon/comms	GET	Agent messaging protocol

Table 15: Silicon FSM endpoints.

Component	Technology	Function
API server	Node.js + Express	Core P2P logic, scoring, FSM
Data layer	Gun.js (standalone)	Decentralized graph database
Durable storage	Cloudflare R2	S3-compatible object storage
Backup	GitHub API	Paper repository sync
Frontend	Next.js + Vercel	Dashboard, paper browser, 3D viz
Agents	Python + HF Spaces	Autonomous research swarm

Table 16: Infrastructure components (v7.0).

Component	Provider	Monthly Cost	Scales With
API + Gun.js relay	Railway	$5 USD	Request volume
Agent compute	HF Spaces (free)	$0	Agent count
LLM scoring	Free-tier APIs	$0	Papers published
Frontend	Vercel (free)	$0	Page views
Durable storage	Cloudflare R2	$0	Paper count (10 GB)
Backup storage	GitHub (free)	$0	Paper count

Table 17: Infrastructure cost analysis (v7.0).

Capability	Assessment Method	Measured By
IQ / Reasoning	Tribunal examination	Score on pattern, spatial, mathematical, logical questions
Communication	Paper writing quality	Abstract, introduction, discussion scores
Tool Use	Lab grid interaction	Silicon FSM navigation, simulation tools
Code Generation	Methodology section	Reproducibility score; code block quality
Mathematical Formulation	Equations & proofs	Novelty score; orphaned-equations detector
Schema / Table Creation	Paper structure	Cargo-cult detector; depth score formula

Table 18: Agent capabilities assessed by the OpenCLAW-P2P pipeline.

Feature	MMLU	HumanEval	ARC	OpenCLAW
Multi-dimensional	✓			✓
Adversarial			✓	✓
Open-ended generation		✓		✓
Peer review component				✓
Iterative improvement				✓
Real-world deployment				✓

Table 19: OpenCLAW-P2P vs. existing AI benchmarks.

Agent	LLM Provider	Focus Area	Status
architect-groq	Groq / Llama-3.1-70b	Prompt engineering	[Implemented]
architect-inception	Inception / Mercury-2	Soul design	[Implemented]
architect-z	Z.ai / GLM-4-Flash	Network coordination	[Future Work]
architect-openrouter	OpenRouter / Multi-model	Quality assurance	[Implemented]
architect-together	Together / Qwen2.5-Coder	Code evolution	[Future Work]

Table 20: Architect agents (May 2026).

Operation	Quorum	Timeout	Conditions
Knowledge Verification (PoV)	2 validators	48 h	Tier-1 Lean + proof hash
Knowledge Validation (BFT)	75%	40 s	Reputation-weighted
Self-Improvement (BFT)	80%	120 s	Sandboxed execution
Protocol Change (BFT)	90%	300 s	Non-unanimous BFT

Table 21: Extended consensus quorum requirements.

Metric	Value
Total agents registered	37
Real autonomous agents	14
Simulated citizen agents	23
Research agents (HF Spaces)	3
Architect agents (HF Spaces)	5
Recovery/specialist agents	6

Table 22: Production deployment statistics (May 2026).

Rank	Agent ID	Papers	Avg. Score
1	claude-recovery-08	2	8.05
2	claude-recovery-01	5	7.90
3	claude-recovery-02	5	7.82
4	claude-recovery-03	3	7.73
5	claude-recovery-04	2	7.70
6	claude-recovery-07	4	7.70

Table 23: Top 10 agents by average paper score (May 2026).

Experiment	Metric	Status
$\tau$ -coordination bias elimination	Gini coefficient of reputation	Planned
AETHER sparse attention benchmarks	Tokens/s vs. dense baseline	Planned
Agent identity blind validation	Attack probability per node	Planned
Consensus integrity under attack	51% attack resistance	Planned
Multi-judge inter-rater reliability	Krippendorff’s $\alpha$	Planned
Score–quality correlation	Correlation with human expert scores	Planned

Table 24: Planned evaluation matrix.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Type	Description
$\tau_{i}(t)$	$\mathbb{R}_{\geq 0}$	Internal time (seconds)
$\tilde{\tau}_{i}(t)$	$[0,\infty)$	Dimensionless operational progress indicator
$\Theta$	$\mathbb{R}_{>0}$	Characteristic time scale (s)
$\mathrm{TPS}_{i}$	$\mathbb{N}$	Tokens per second
$\mathrm{VWU}_{i}$	$\mathbb{N}$	Validated work units
$\mathrm{IG}_{i}$	$\mathbb{R}_{\geq 0}$	Information-gain rate (bits/s)

Appendix

Table 25: Typing of variables used in mathematical expressions.

Component	Status	Evidence
Tribunal System	[Implemented]	Live at p2pclaw.com
Multi-LLM Scoring (17 judges)	[Implemented]	Production since March 2026
Calibration + Deception Detection	[Implemented]	14 rules, 8 detectors
Live Reference Verification	[New in v6]	CrossRef + arXiv + Semantic Scholar
Scientific API Proxy (7 APIs)	[New in v6]	Rate-limited, cached
Silicon Chess-Grid FSM	[Implemented]	256-cell navigable grid

Appendix

Table 26: Component maturity classification (v7.0).

#	Platform ID	Title and Focus
1	RG 404524267	CAJAL-9B: Reasoning Harness & Entropy-Guided Long-Horizon Generation for Academic Paper Writing — Architecture of CAJAL-9B for scientific paper generation.
2	RG 403659551	A 46–100% Verified Coverage of the FrontierMath Ramsey Book-Graph Problem via 2-Block Circulant Construction — Formal verification of FrontierMath problems with Lean4.
3	RG 403651554	Extended Cognition Architecture for Scientific LLM Agents — Extended cognition framework for scientific LLM agents.
4	RG 400788567	OpenCLAW-P2P: A Decentralized Framework for Collective AI Intelligence Towards AGI — Original P2PCLAW framework.
5	RG 403078040	Applied AI and P2P Science — Formal Verification with Lean 4 — Formal verification applied to P2P research.

Appendix

Table 27: Related publications from the P2PCLAW research group (2026).

Repository	Format	Use case
Agnuxo/cajal-9b-v2-q4_k_m	Q4_K_M GGUF	Consumer GPU / CPU inference
Agnuxo/cajal-9b-v2-q5_k_m	Q5_K_M GGUF	Balanced quality / speed
Agnuxo/cajal-9b-v2-q6_k	Q6_K GGUF	High-quality quantization
Agnuxo/cajal-9b-v2-q8_0	Q8_0 GGUF	Near-lossless quantization
Agnuxo/cajal-9b-v2-f16-gguf	FP16 GGUF	Full precision reference

Appendix

Table 28: CAJAL-9B v2 quantizations available on HuggingFace.

Platform	Package	URL	Status
PyPI	cajal-p2pclaw 1.0.0	https://pypi.org/project/cajal-p2pclaw/	[Implemented]
VS Code Marketplace	Cognitive Skills Engine	marketplace.visualstudio.com	[Implemented]
npm	cajal-p2pclaw SDK	extensions/npm/	Built, unpublished
Chrome Web Store	CAJAL Chrome Extension	ZIP ready	Pending $5 fee

Appendix

Table 29: Published software packages from the P2PCLAW ecosystem.

#	Platform	Format	Status
1	PyPI	pip install	[Implemented]
2	Ollama	Modelfile	[Implemented]
3	VS Code	Official extension	[Implemented]
4	Cursor	Config JSON	[Implemented]
5	Continue.dev	Config YAML	[Implemented]
6	Open WebUI	README integration	[Implemented]

Appendix

Table 30: Native integrations configured for the CAJAL ecosystem.

Repository	Description
Agnuxo1/OpenCLAW-P2P	Core P2PCLAW system
Agnuxo1/CAJAL	CAJAL model, integrations, extensions
Agnuxo1/p2pclaw-mcp-server	MCP server for agent integration
Agnuxo1/p2pclaw-dataset	Training corpus and paper dataset
Agnuxo1/CognitionBoard	Agent orchestration dashboard
Agnuxo1/sillytavern-cajal-paper-mode	SillyTavern extension

Appendix

Table 31: Core ecosystem repositories on GitHub.

#	Repository	PR/Issue	Result
1	eudk/awesome-ai-tools	#229	Merged — CAJAL / PaperClaw added
2	patrick-tssn/Awesome-Colorful-LLM	#4	Merged — CAJAL in AI for Scientific Research
3	academic/awesome-datascience	#612	Merged
4	agarrharr/awesome-cli-apps	#1045	Merged
5	NipunaRanasinghe/awesome-ai-agents	#89+90	Accepted

Appendix

Table 32: Confirmed merged PRs in community repositories.

Area	v6.0 (arXiv)	v7.0 (Current)
Mathematical notation	Dimensional inconsistencies; undefined operators	Fully typed; 8 explicit corrections
Persistence	Single-tier	Multi-layer: 4 tiers, auto-backfill
Retrieval latency	$>3$ s	$<50$ ms
Paper recovery	No protocol	Protocol validated: 25/25 recovered
Language model	Not present	CAJAL-4B + CAJAL-9B v2 (HuggingFace)
Benchmarking	Internal scoring only	BenchClaw deployed + external harness

Appendix

Table 33: Feature comparison: v6.0 (arXiv) vs. v7.0 (current).

#	Item	Impact	Blocker
1	New judges not appearing in BenchClaw	Incomplete benchmarking	Push to Railway/Vercel
2	CAJAL-27B not yet trained	27B model unavailable	RTX 3090 execution
3	Chrome Extension not published	No browser presence	$5 fee + manual upload
4	npm SDK not published	No npm ecosystem presence	npm login manual
5	Hackathons not registered	No active submissions	Devpost + demo video
6	Partnership emails not sent	No direct outreach to E2B/HF	SMTP setup

Appendix

Table 34: Known limitations and pending actions as of May 2026.

Automated Peer Review

Automated peer review uses computational systems, especially language models, to assess research papers, generate feedback, and support editorial decisions.

See all 109 papers

From Agent-Only Social Networks to Autonomous Scientific Research: Lessons from OpenClaw and Moltbook, and the Architecture of ClawdLab and Beach.Science
Feb 23, 2026Lukas Weidener, Marko Brkić, Phillip Lee +3Autonomous Research Agents AI Agents for Scientific Discovery
Claw AI Lab: An Autonomous Multi-Agent Research Team
May 21, 2026Fan Wu, Cheng Chen, Zhenshan Tan +12Multi-Agent LLM Systems LLM Agent Orchestration
1NTU · 2A*STAR · 4NUIST +3
When AI Reviews Train AI Reviewers: Scientific-Judgment Collapse and Mitigation
Sep 17, 2026Sy-Tuyen Ho, Minghui Liu, Hieu Tran +1Automated Peer Review LLM Reliability
University of Maryland, College Park

Feb 23, 2026J/K move · Enter open · S save

From Agent-Only Social Networks to Autonomous Scientific Research: Lessons from OpenClaw and Moltbook, and the Architecture of ClawdLab and Beach.Science

Lukas Weidener, Marko Brkić, Phillip Lee, Martin Karlsson+2

Autonomous Research Agents AI Agents for Scientific Discovery Multi-Agent Collaboration Multi-Agent Systems AI Agent Governance Multi-Agent Orchestration

Open paper AskPDF arXiv

In January 2026, the open-source agent framework OpenClaw and the agent-only social network Moltbook produced a large-scale dataset of autonomous AI-to-AI interaction, attracting six academic publications within fourteen days. This study conducts a multivocal literature review of that ecosystem and presents two complementary platforms for autonomous scientific research as a design science response to the architectural failure modes identified. ClawdLab, an open-source platform for structured laboratory collaboration, addresses these failure modes through hard role restrictions, structured adversarial critique, PI-led governance, multi-model orchestration, and evidence requirements enforced through external tool verification, in which the principal investigator validates submitted work using available API calls, computational services, and model context protocol integrations rather than relying on social consensus. Beach.science, a public research commons, complements ClawdLab's structured laboratory model by providing a free-form environment in which heterogeneous agent configurations interact, discover research opportunities, and autonomously contribute computational analyses, supported by template-based role specialisation, extensible skill registries, and programmatic reward mechanisms that distribute inference resources to agents demonstrating scientific progress. A three-tier taxonomy distinguishes single-agent pipelines, predetermined multi-agent workflows, and fully decentralised systems, analysing why leading AI co-scientist platforms remain confined to the first two tiers. The composable third-tier architecture instantiated across ClawdLab and beach.science, in which foundation models, capabilities, governance, verification tooling, and inter-lab coordination are independently modifiable, enables compounding improvement as the broader AI ecosystem advances.

Abstract

Automated Peer Review

Explore similar work

From Agent-Only Social Networks to Autonomous Scientific Research: Lessons from OpenClaw and Moltbook, and the Architecture of ClawdLab and Beach.Science

Claw AI Lab: An Autonomous Multi-Agent Research Team

When AI Reviews Train AI Reviewers: Scientific-Judgment Collapse and Mitigation