Synthos 🧠⚡

A Self-Evolving Cognitive Operating System for AI-Augmented Scientific Research

Synthos reifies epistemological principles as executable cognitive atoms. Unlike conventional research agents that wrap Python libraries as tools, Synthos operates as a purely skill-driven architecture: the Agent itself is the runtime, with zero Python infrastructure code (core orchestration is 100% SKILL.md; helper automation scripts in scripts/ use Python/shell but are not part of the cognitive architecture).

📖 Full philosophy → docs/synthos-philosophy.md — Classical Chinese principles · 8-dimension cognitive framework · Three inviolable laws

Core Innovations

Innovation	Description
🧬 Epistemological Code	8 philosophical frameworks (First Principles, Falsificationism, Bayesian Reasoning, etc.) reified as executable SKILL.md atoms with formal I/O contracts
🏛️ Constitutional Hierarchy	CON ≫ MEM ≫ CMD ≫ SKL ≫ DEF — immutable principles enforced by a philosophical immune system
🔄 Self-Evolution Engine v2.39	11-step state machine with SEPL rollback, Git-as-Memory, structural probes, functional benchmarks, external absorption, drift detection, and active inference gates
🌱 Entelechy-Driven Absorption	Aristotelian self-realization as the absorption axiom — from "gap-filling" to "nutrition assessment"
📝 Zero-Python Architecture	All cognitive atoms + meta-components are pure SKILL.md markdown. No Python orchestration code

Architecture

┌─────────────────────────────────────────────────────┐
│              CONSTITUTION.md v5.0                    │
│        (CON ≫ MEM ≫ CMD ≫ SKL ≫ DEF)                │
├─────────────────────────────────────────────────────┤
│   ACQ → EXT → ASC → RIF → HYP → ARG/VER             │
│         ↕        ↕        ↕        ↕                 │
│   [Task Router] [Entelechy Gate] [Quality Gate]      │
├─────────────────────────────────────────────────────┤
│              Evolution Engine (11 steps)              │
│   PROBE → BENCHMARK → OPTIMIZE → EXTERNAL →          │
│   DIAGNOSE → IMPROVE → VERIFY → RECORD               │
└─────────────────────────────────────────────────────┘

7 Cognitive Atoms: Knowledge Acquisition → Knowledge Extraction → Association Discovery → Research Ideation → Hypothesis Generation → Argument Expression → Viewpoint Verification

3 Meta-Components: Task Router (shortest-path routing), Entelechy Gate (direction compatibility), Quality Gate (L0–L4 multi-layer verification)

3-Layer Skill Architecture

Synthos organizes 203 SKILL.md files into a strict layered hierarchy:

Layer	Skills	Purpose	Precedence
Core	88 skills	CON > MEM > CMD — immutable principles, core atoms	Highest
System	50 skills	SKL — system maintenance, evolution, quality gates	Medium
Optional	60 skills	DEF — auxiliary tools, creative, productivity	Lowest

6 Composite skills merge overlapping functionality into unified capabilities (github-workflow, creative-tools, mlops-toolchain, productivity-suite, social-monitor, paper-pipeline).

Paper Outputs

Synthos has produced 44 paper drafts across multiple biomedical domains and methodology:

Category	Papers	Size
🧠 Neurology	PD Torsion Review, BPPV Minimal Stimulus, PD Dysphagia Risk, HCS-3WT Breast Cancer	4 papers
👁️ Ophthalmology	Iris 3D Anatomical Segmentation, Iris YOLO, CutEye Model, Portable Eye Tracking	4 papers
🔄 VOR/Vestibular	VOR Sparse Modular NN, VOR Digital Twin	2 papers
📊 Methodology	PIMA CRISP-DM Helix	1 paper
📚 Literature Reviews	33 reviews across eye-tracking, VOR, BPPV, iris, eyeball, methods	33 reviews
Total	44 paper.tex files

Evolution History

Metric	Cycle 1	Cycle 53	Cycle 96
Structural Avg	0.861	1.00	0.92
Benchmark Score	0.66	1.00	0.82
Composite Score	0.86	0.98	0.95
External Absorptions	0	18	30+
SKILL.md Files	~50	~100	203

Comparison with Existing Systems

System	Epistemic Encoding	Constitutional	Self-Evolution	Zero-Python
GPT-Researcher	✗	✗	✗	✗
AI Scientist	✗	✗	△	✗
PaperQA2	✗	✗	✗	✗
Constitutional AI	✗	✓	✗	✗
DSPy	✗	✗	✓	✗
DeepResearchAgent	✗	✗	✓	✗
AI-Research-SKILLs	△	✗	✗	✓
Synthos (ours)	✓	✓	✓	✓

Quality Audits

Cycle	Paper	Score	Recommendation
97	synthos-paper.tex	62.5/100 (C)	MAJOR_REVISION — 7 orphan refs, 80% GitHub refs

See .evolution/reviews/ for detailed quality reports.

License

MIT License — see LICENSE.

Citation

If you use Synthos in your research, please cite:

@software{yang2026synthos,
  author = {Yang, Xiaokai},
  title = {Synthos: A Self-Evolving Cognitive Operating System for AI-Augmented Scientific Research},
  year = {2026},
  url = {https://github.com/yakeworld/Synthos}
}

Synthos v2.39.0 — 100 evolution cycles — Quality Score 0.90 (EXCELLENT) — Super Individual Engine ON