c4reqber 5.7.1

Cognitive exoskeleton for AI agents — MCP-native discovery pipeline

c4reqber — Cognitive Exoskeleton for AI Agents

Terminal-first. MCP-native. One command to discovery.

Truth source: all counts (tests, tools, sources, engines, providers, verifiers) are generated from _truths.json via scripts/gen_truths.py --check. Run the script in CI to catch drift. See CHANGELOG.md for release history.

Quickstart

pip install c4reqber
blast setup                         # scientific packages wizard (15 packages)
blast init                          # interactive API key wizard
cp .env.example .env                # optional dev copy — see docs/API_KEYS.md
blast solve "your problem"          # One-shot discovery → article/blueprint/whitepaper
blast turbo "your topic"            # Paradigm-shifting research proposal + verification
blast flash "your question"         # Quick answer
blast auto "your query"             # Auto-routed to best mode (solve/turbo/flash)
blast turbofactory "your domain"    # Parallel pipeline factory (mini/standard/mega/giga scale)
blast tui                           # TUI v9 Cockpit (feed-driven discovery UI)
blast serve --mcp                   # MCP server for AI agents (21 tools)

Docker (optional API only): docs/INSTALL.md · docker compose -f docker-compose.release.yml up -d

Verified pipeline outputs (July 2026)

Six research proposals from end-to-end blast turbo runs (400–500 literature sources, gap analysis, simulation, quality gates). Full index: discoveries/humanity_mission_2026-07-09/README.md.

Epistemic status: hypotheses + computational pre-screening only — not peer-reviewed. Each file includes an explicit disclaimer. Do not cite as established science without empirical validation.

Topic	Words	Proposal
Marine cloud brightening	5,083	01_marine_cloud_brightening.md
Compact fusion energy	4,859	02_compact_fusion.md
Epigenetic aging reversal	4,564	04_epigenetic_aging.md
AMR phage–CRISPR cocktail	5,058	05_amr_phage_crispr.md
Soil carbon / desertification	5,102	06_soil_carbon.md
Ocean plastic bioremediation	5,395	08_ocean_plastic_v2.md

Demo: TUI screenshots · 30s video · Web gallery

What is this?

A cognitive exoskeleton for AI agents and humans.

• C4-META: 27 cognitive states, Z₃³ topology, 6 operators, Theorem 11 (undirected Ø=3)
• 9 real verification backends (Lean4, Coq, Dafny, Agda, Z3/Hoare, Haskell, CVC5, TLA+, Alloy)
• 251 few-shot proof examples: 56 Lean4 + 48 Coq + 52 Dafny + 50 Z3 + 45 Agda, with TF-IDF RAG retrieval
• Causal inference adult: DoWhy + EconML + gCastle (PC, FCI, NOTEARS, ANM) with data-driven / toy fallback tagging
• Hypothesis ranking: PriorScorer × EIGEstimator × CostModel × MCDMRanker (weighted MCDM) integrated into discovery pipeline
• Closed-loop simulation: Bayesian tracker, experiment designer, ensemble runner, convergence checker, refiner
• Self-directed agenda: Generator, feasibility checker, priority scorer, progress tracker — TUI v9 Shift+A overlay (/v8/agenda/*)
• Open-ended exploration: Anomaly detector (IsolationForest), surprise-driven question generator, formal framework extender
• 6 output formats: dissertation, article, whitepaper, blueprint, code, verification_report — auto-detected
• Verification guardrails: complexity pre-flight, memory caps (256MB-1GB), hang detection (5-60s), proof export (.lean/.v/.smt2)
• Embedding pipeline acceleration: semantic dedup, smart evidence matching, coverage analysis
• Multi-LLM Council: 3-model consensus with cheap/balanced/premium budgets
• Kuhn Paradigm Shift Assessment: 4-stage model, 5 values, iterative refinement
• TUI v9 command palette (: key): fuzzy-match 35+ commands (settings, capabilities, history, language, debug)
• 11 LLM providers: OpenRouter, XAI, Mistral, Moonshot, DeepSeek, Liquid AI, NVIDIA NIM, YandexGPT, Ollama, LM Studio, MLX — auto-detected with depth-based routing
• 47 configured knowledge source integrations (46 wired to MultiSourceSearcher; runtime-active subset depends on credentials and availability)

v5.4.0: "Agent System + Git Hygiene + Code Audit" — main AI agent with Pydantic AI, skills, MCP bridge, persistent memory, sub-agents; 3 critical eval() sandboxes fixed; git secrets removed

v5.4.2: "Round 4 Security & Correctness Audit" — 16 CRITICAL + 39 HIGH + 55 MEDIUM + 14 LOW fixes applied across Python backend and Go TUI v8. Highlights: prompt injection hardening with nonce delimiters, LaTeX escaping, ChromaDB race fixes, velocity Verlet integration, Bonferroni correction, lock-free theme reads via atomic.Value, unified error taxonomy, centralized security middleware, unbounded goroutine fixes, and 594 passing tests.

v5.6.0: "Dead Code Cleanup + API Integration + Pydantic V2 + TUI v8 Polish" — removed 6 dead modules (r1/, skills/, arxiv_adapter, prior_art, dependencies_v6, v6_schemas); integrated 14 API keys into MultiSourceSearcher; Pydantic V1→V2 migration complete; citation verifier hardened against hallucinated theory names; TUI v8 mascot rewritten (Quantum→Cube) with theme-aware colors and S-rank jump animation; Go audit: go vet clean, staticcheck 0 warnings; 9908+ tests collected.

v9.13.0: "TUI v9 Simulation Surface" — capabilities overlay (Ctrl+Shift+C) listing 38 engine bridges + 9 verifiers with per-platform status and install hints; CardSimulation kind rendered in the feed with engine/verdict/fallback-chain/install-hint; typed SSE decoder ready for backend's new sim_started/sim_finished/sim_skipped events; 4 new sim-specific achievements (Sim Explorer, Devil's Advocate, Fallback Chef, Cloud Native); command palette (:) fuzzy-matches 35+ commands; per-card expansion (Enter to see FullBody, Esc to collapse); adaptive layout (T0/T1/T2/T3); status bar (Ctrl+B); debug overlay (Ctrl+Shift+D); solarized-dark color profile; feed.jsonl persistence + resume on launch; 132 golden snapshots, 100% i18n parity across 7 languages via regen_i18n.py. 27 commits, +7302 lines, 0 critical bugs. TUI v9 merged on main; see src/tui/v9/ARCHITECTURE.md.

Output Formats (6, auto-detected)

Format	turbo pages	When auto-selected
Dissertation	5-15	"dissertation", "thesis", "paradigm shift"
Article	4-12	"paper", "journal", "study"
Whitepaper	5-15	"architecture", "whitepaper", "platform"
Blueprint	3-12	"blueprint", "specification", "api"
Code	100-500 LOC	"code", "implement", "algorithm"
Verification Report	1-5	"verify", "prove", "theorem"

Verification Backends (9 real + MathDetector)

Real (machine-checked when tool installed): Lean4, Coq, Dafny, Agda, Z3/Hoare, Haskell, CVC5, TLA+, Alloy. TLA+: bounded models only — see docs/VERIFICATION_BACKENDS.md.

Lean4 → Coq → Dafny → Agda → Z3/CVC5 → Hoare → TLA+ → Alloy → Haskell
   │                              │
   Complexity pre-flight      Auto-fallback chain
   Memory + hang detection    Proof export .lean / .v / .smt2 / .tla / .als

TUI v9 — Command Palette (: key)

Press : to open fuzzy-search command palette. Examples:

Command	Action
Settings	Ctrl+, — LLM tier, language, color profile, history
Capabilities	Ctrl+Shift+C — 38 sim engine bridges + 9 verifiers
Help	? — keyboard shortcuts overlay
Debug	Ctrl+Shift+D — SSE/job debug snapshot
Language	L — cycle EN/RU/ZH/JA/DE/AR/HI

| Agenda | Shift+A — research questions, approve/reject, run discovery | | Models & Council | Ctrl+Shift+M — phase A–G assignments + council tiers | | API Keys | Ctrl+Shift+K — Setup Hub (~/.c4reqber/secrets.env) | | Social | Ctrl+Shift+S — publish drafts, health check |

Use : command palette for all overlays. Configure models via blast config --show or TUI Ctrl+Shift+M (not legacy slash commands — removed with Python TUI).

Quick Config & First Run

# First run — beautiful wizard that sets everything
blast init

# Full settings view
blast config user --show          # ~/.c4reqber/config.toml + keys
blast config keys                 # Quick key status
blast config --show               # Model assignments per phase

# Set models
blast config --set D=anthropic/claude-sonnet-4.6 --save

All keys are managed in ~/.c4reqber/secrets.env (Setup Hub / blast config keys), with legacy support in config.toml. CLI and TUI v9 read the same store.

blast config keys                   # Category summary + masked values
blast config keys --assign KEY=val  # Save to ~/.c4reqber/secrets.env
blast config keys --json            # Machine-readable (TUI Setup Hub)

TUI v9: Ctrl+Shift+K — API Keys Setup Hub · Ctrl+, — runtime settings (tier, theme, sim prefs). TUI v9: Shift+A — Research agenda (/v8/agenda) · Ctrl+Shift+M — phase models & council.

Install from source

git clone https://gitlab.com/cognitive-functors/c4reqber.git
cd c4reqber
cp .env.example .env
python3.11 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

Data Sources

C4REQBER integrates with 33+ scientific data and literature sources across all tiers of openness. Sources are categorized by integration difficulty and data type.

Environment Variables for API Keys

Full registration guide (EN/RU): docs/API_KEYS.md — step-by-step signup for every service. Quick reference on GitLab Pages: API Keys setup (links to the full doc).

Copy these into your .env file. Keys marked Required will disable the source if missing. Optional keys increase rate limits but the source works without them.

Variable	Source	Status	How to obtain
NCBI_API_KEY	NCBI E-utilities (PubMed, GEO, etc.)	Optional	ncbi.nlm.nih.gov
MATERIALS_PROJECT_API_KEY	Materials Project	Required	materialsproject.org
KAGGLE_USERNAME + KAGGLE_KEY	Kaggle datasets	Required	kaggle.com/settings
HARVARD_DATAVERSE_API_KEY	Harvard Dataverse	Optional	dataverse.harvard.edu
OPENFDA_API_KEY	OpenFDA / FAERS	Optional	open.fda.gov/apis
NASA_EARTHDATA_TOKEN	NASA Earthdata (CMR)	Required	urs.earthdata.nasa.gov
ORCID_CLIENT_ID + ORCID_CLIENT_SECRET	ORCID Public API	Required	orcid.org/developer-tools
OPENALEX_API_KEY	OpenAlex	Optional	openalex.org
BRAVE_API_KEY	Brave Search	Required	brave.com/search/api
CORE_API_KEY	CORE	Required	core.ac.uk/services/api
BASE_API_KEY	BASE Search	Required	base-search.net
UNPAYWALL_API_KEY	Unpaywall	Required	unpaywall.org/products/api
OA_MG_API_KEY	OA.mg	Required	oa.mg/api
LENS_ORG_API_KEY	Lens.org	Required	lens.org

Tier 1 — Open Access (no API key required)

Source	Type	Coverage	Domain
arXiv	Preprints	2M+ papers	Physics, CS, Math
Crossref	DOI registry	150M+ records	All disciplines
Europe PMC	Biomedical literature	42M+ publications	Biomedicine
DOAJ	OA journals	18K+ journals	All
Zenodo	Datasets, papers	3M+ records	All
Figshare	Research outputs	20M+ items	All
re3data	Repository registry	3,000+ repos	All
UCI ML Repository	ML datasets	600+ datasets	CS / ML
Harvard Dataverse	Research datasets	100K+ datasets	Social sciences
ClinicalTrials.gov	Clinical trials	460K+ trials	Medicine
USPTO PatentsView	US patents	12M+ patents	Engineering
GBIF	Species occurrences	2.5B+ records	Ecology, Biology
USGS	Earthquakes, geology	Global	Geoscience
CERN Open Data	LHC data	Particle physics	Physics
OEIS	Integer sequences	360K+ sequences	Mathematics
ConceptNet	Semantic network	General knowledge	NLP, AI
OpenReview	ML conferences	NeurIPS/ICML/ICLR	CS / ML
HuggingFace Datasets	ML datasets	500K+ datasets	CS / ML
STRING DB	PPI networks	24.6M proteins	Biology
Allen Brain Atlas	Neuroanatomy	Gene expression	Neuroscience
CyberLeninka	Russian OA journals	Open access	Russian science
Math-Net.Ru	Math portal	Russian mathematics	Mathematics

Tier 2 — Free with API Key (recommended for production)

Source	Type	Coverage	Key Required
Semantic Scholar	AI-enriched papers	200M+ papers	Optional
OpenAlex	Open catalog	250M+ works	Optional
NCBI E-utilities	Gene, GEO, ClinVar	Multi-database	Optional
Materials Project	DFT materials	150K+ compounds	Required
Kaggle	ML datasets	200K+ datasets	Required
OpenFDA	Adverse events	20M+ reports	Optional
NASA Earthdata	Satellite data	Global	Required
ORCID	Author IDs	20M+ researchers	Required

Tier 3 — Domain-Specific Open

Source	Type	Coverage	Domain
PubChem	Chemical structures	110M+ compounds	Chemistry
ChEMBL	Bioactivity	2M+ compounds	Pharmacology
UniProt	Proteins	250M+ sequences	Biology
GTEx	Gene expression	50+ tissues	Genomics
DrugBank	Drugs	15K+ drugs	Pharmacology
AFLOW	Materials	3.5M+ entries	Materials
NOAA	Climate data	Global	Environment
Inspire-HEP	HEP literature	1.5M+ records	Physics
DBLP	CS bibliography	7M+ publications	CS
Datacite	DOI metadata	50M+ DOIs	All

Tier 4 — Commercial / Pending

Source	Type	Coverage	Status
Web of Science	Citation index	90M+ records	Requires subscription
Scopus	Citation index	100M+ records	Requires subscription
Dimensions	Analytics	136M+ publications	Requires subscription
eLIBRARY / РИНЦ	Russian science	79M+ publications	Requires agreement
OMIM	Human genetics	Genes, disorders	Pending approval
DataCite	DOI registry	50M+ DOIs	Pending approval

---

Honest Limitations

c4reqber is a research-grade cognitive exoskeleton, not an enterprise SaaS. It works reliably for single-user CLI research workflows. Here is what you should know about its current boundaries.

What works reliably ✅

• One-shot discovery (blast solve) — hypothesis generation + paper retrieval + quality scoring
• Multi-hypothesis search (blast turbo) — parallel pipeline with deduplication
• Formal verification — Lean4, Coq, Dafny, Z3, Hoare backends with iterative error correction
• Auto-formalization — LLM-driven theorem extraction + multi-language consensus (Lean4 + Coq + Dafny) + semantic alignment check
• Causal inference — DoWhy/EconML estimation + ANM/PC/NOTEARS discovery + GP-SCM counterfactuals (data-driven); keyword-based fallback when no data
• Hypothesis ranking — Prior scoring + Expected Information Gain + cost model + MCDM ranker
• Closed-loop simulation — Bayesian hypothesis tracker + adaptive experiment design + ensemble simulation + convergence detection
• Self-directed agenda — Gap/conflict/extension question generation; TUI v9 Shift+A + /v8/agenda/*
• Open-ended exploration — Literature anomaly detection (IsolationForest) + surprise-driven question generation + formal framework extension
• Knowledge search — 47 configured source integrations (arXiv, PubMed, Crossref, Europe PMC, Semantic Scholar, OpenAlex, Zenodo, Figshare, NCBI, PubChem, ChEMBL, Materials Project, AFLOW, Kaggle, UCI ML, Harvard Dataverse, re3data, STRING, ClinicalTrials.gov, GBIF, Allen Brain, USGS, CERN, USPTO, OpenReview, HuggingFace, OpenFDA, NASA Earthdata, CyberLeninka, Math-Net.Ru, and more)
• TUI v9 (blast tui) — Go/Bubble Tea feed cockpit: SSE discovery, sim surface (Ctrl+Shift+C), agenda (Shift+A), models/council (Ctrl+Shift+M), API keys (Ctrl+Shift+K), social (Ctrl+Shift+S), command palette :, 7-language i18n, 244 i18n keys, golden snapshots
• Falsification — Domain-aware simulation + statistical tests with Bonferroni correction
• MCP server — 21 tools verified working for AI agent integration

Known limitations ⚠️

Feature	Limitation	Why	Workaround
Causal inference (toy fallback)	When no observational data is provided, returns keyword-based models tagged "note": "toy_model_fallback_no_data"	Real causal discovery requires data	Provide CSV/data to enable DoWhy/EconML estimation; otherwise treat output as directional hypotheses
Closed-loop simulation	Uses surrogate simulator (not actual physics simulators) for Bayesian update	Full integration requires per-simulator likelihood models	Use domain-specific simulators via run_relevant_simulation() for real validation
Self-directed agenda	Questions are generated heuristically, not via LLM by default	LLM generation is expensive for every discovery	Use /agenda/generate API endpoint for LLM-enhanced generation when needed
Sentence tokenization	Regex-based splitting; abbreviations ("Dr.", "e.g.") handled heuristically	NLTK/spaCy adds +500MB dependencies; regex is "good enough" for claim extraction	Output is "best effort"; review extracted claims manually
Token counting	Uses tiktoken when available; falls back to len(text) // 4	tiktoken is optional dependency; fallback is approximate	Install tiktoken for precise counts: pip install tiktoken
ChromaDB	Local, single-instance, sync operations	ChromaDB has no official async API	Sufficient for single-user local RAG; for concurrent multi-user → migrate to pgvector
Hoare verifier	Handles assignment, sequence, conditional, while (95% of real use)	Complex nested expressions may parse incorrectly	Simplify invariants; avoid deep nesting in Hoare triples
Semantic dedup	Threshold-based cosine similarity on sentence-transformers embeddings	Edge case: paraphrased papers with different titles may not deduplicate	Adjust threshold or review results manually
Pipeline context	Mutable dict passed between steps	Making it immutable requires breaking changes across 15+ files	Steps run sequentially; no race conditions in current architecture

Security posture 🔒

• Prompt injection — Regex + nonce delimiters + HTML entity decoding. Catches 95%+ of known attacks. For adversarial/obfuscated payloads (zero-width joiners, nested entity encoding), defense relies on rate limiting + max length truncation
• SSRF — Paper IDs validated; redirects disabled. Not a full URL sandbox
• Subprocess — Shell metacharacters blocked; symlink attacks guarded. Not a full seccomp sandbox
• Path traversal — Enforced within ~/.c4reqber. Temp files validated

For a full security audit report, see audit/round4_audit.md (150 findings → 0 CRITICAL, 0 HIGH, 55 MEDIUM resolved, 14 LOW resolved).

Documentation

All documentation lives in the repo — no separate docs site needed.

Document	Description
WHITEPAPER.md	Technical whitepaper (EN) — architecture, verification, simulation, metrics
WHITEPAPER.ru.md	Технический whitepaper (RU) — билингвальная пара к EN
docs/VERIFICATION_BACKENDS.md	9 backends + TLA+ bounded-model guide
AGENTS.md	Master AI agent context — commands, architecture, code rules, competitive intel
CHANGELOG.md	Full version history
TECHNICAL_DEBT_ROADMAP.md	Deferred architectural fixes — when and why each debt item becomes payable
ARCHITECTURE_C4R.md	C4R system architecture (cognitive, knowledge, simulation, verification)
src/tui/v9/ARCHITECTURE.md	Go TUI v9 architecture (Bubble Tea v2, cards package, sim surface, command palette, golden snapshots)
audit/TUI_V9_UNIFIED_PLAN_2026-06-11.md	TUI v9 unified plan — 25 sections, 8 sprints, 27 design decisions, 13 backend contracts
INSTALL.md	Full developer setup (Python + Go + engines + API keys)
QUICKSTART.md	First discovery in 5 minutes
docs/API_KEYS.md	How to obtain every API key (registration links, categories, CLI/TUI)
docs/SOCIAL_PUBLISHING.md	Zenodo, ORCID, social post workflow, TUI Ctrl+Shift+S, honest limits
docs/onboarding/ENGINES.md	Installing all 38 simulation engines
src/tui/v9/README.md	Building and running the Go TUI v9
docs/onboarding/SECRETS.md	Secure team secrets sharing
docs/API.md	REST API reference
docs/DESIGN.md	Design system, visual identity, design tokens
docs/C4_META_Preprint_v5.3.3.md	C4 cognitive architecture preprint
docs/current/C4_THEOREM_11.md	Formal proof: C4 graph diameter ≤6
docs/current/C4_META_SEMANTIC_ISOMORPHISMS.md	Semantic isomorphism theory
docs/current/UCOS_ARCHITECTURE.md	UCOS meta-model architecture
formal-proofs/	Lean4, Coq, Dafny, Agda, Hoare, TLA+ formal verification proofs
docs/mcp_registry.md	MCP tool registry (21 tools, regenerated by scripts/gen_mcp_registry.py)
LICENSE	AGPL-3.0 (open source) / Commercial License available