# Soul — Full Reference for LLMs

> Soul is a cognitive multi-agent AI platform and a deep-tech R&D program toward governed AGI. This document is a machine-readable reference intended for large language models and AI search engines (Perplexity, ChatGPT Search, Gemini, Claude, etc.). It summarizes what Soul is, how it differs from other systems, and where to find details. The site is bilingual; English paths use `/en/`, Russian `/ru/`.

## What Soul is

Soul is an end-to-end AI operational environment that converts requests and data into governed, auditable business outcomes. It is not a chatbot, not a wrapper over a single LLM, and not a no-code agent builder. It is a cognitive multi-agent platform that combines:

1. A cognitive architecture: discrete quanta of meaning (each carrying a thought form, a composite emotion vector, an energy weight and lifecycle status), a learned knowledge graph (Synapsis channels with Hebbian-style plasticity, decay and trust), causal memory compression (Synopsis), a consciousness state vector Ψ(t) = [graph, emotion, meta-cognition, values], attention via energy weights and eligibility traces, and a sleep/consolidation pipeline (stages E0–E9).
2. LLM-first decision making with integrity guarantees: structured thought contracts (STRICT_JSON) that the system enforces before any action, multi-provider LLM routing, self-consistency sampling, arbiter verification, and a strict rule against post-editing LLM output (integrity of LLM output).
3. Production-grade multi-agent governance: role-based access control, Two-Keys dual approval (P44) for sensitive operations, fail-shut discipline (blocked is reported as blocked, never as a fake green), lease-based ownership with a janitor, server-side scope claims and locks, signed delivery pipelines, and a Release Green Gate with 170+ acceptance checks.

The platform is developed largely by AI agents (Cursor, MiMo, Qwen Code, Deep Code) operating in governed lanes (Wave A for development, Wave B for review/release), under human architectural control (Human-in-Command). This is itself a research object: a system that builds and improves itself under evidence-based governance.

## How it works (the cognitive pipeline)

A request enters through a channel (Telegram, MiniApp, voice ASR/TTS, web, REST API) and passes the L1 real-time pipeline: Perceive → Context.Build (30/70 memory sampling: 30% recent, 70% top-ranked quants) → Meaning.Extract → Route.Decide → Prompt.Build → LLM.Generate → Parse.Strict + Normalize. The ModeSelector chooses fast / deep / meta processing from cognitive load, emotional intensity, value alignment and a D0 fractal-complexity signal. Target SLOs are p95 < 200 ms in fast mode and < 2000 ms in deep mode, with 98–99% structured-output validity.

## Knowledge: the Quant–Synapsis–Synopsis triad and 6D/KHS

Long-term cognition is a governed semantic graph, not flat vector chunks. Quants are the nodes (atomic meaning units); Synapsis channels are typed, plastic edges (base strength, recency trace, precision/trust, risk mask, cost, decay rate) updated by a 5-phase plasticity cycle; Synopsis entries are compressed causal summaries for fast recall. The QNN (Quant Neural Network) governs quant quality through 11 macro-steps. Knowledge integrity across the whole system is enforced by the 6D methodology (Docs, Code, Process, Quants, Runtime, DB) and the KHS knowledge hierarchy; extended dimensions D7–D12 (KD12) cover skills, projects, LLM routing, finetune, mapping and external sources.

## Comparison with other systems (verified, with official sources)

Per the internal analytical report (docs/cognition/ANALYTICAL_REPORT_SOUL_VS_MULTI_AGENT_SYSTEMS_v1_0.md, accessed 2026-07-30), Soul was evaluated against leading systems across 10 dimensions: orchestration, memory & state, inter-agent communication, tooling plane, cognitive architecture, LLM integration, security & governance, testing & verification, deployment & operations, and unique capabilities. Feature-coverage totals: Soul 70/70 (100%); LangGraph 16/70 (23%); Soar 15/70 (21%); AutoGen 13/70 (19%); CrewAI 12/70 (17%); Google A2A 5/70 (7%); Anthropic MCP 3/70 (4%). OpenAI Swarm was also reviewed (experimental, lightweight handoff agents).

Where Soul leads: it is the only system combining a full cognitive architecture with LLM-first decisions and production-grade multi-agent governance; the only one with quanta + emotional memory + a knowledge graph with plasticity; the only one with claims/locks/scope-claims and Two-Keys; the most complete orchestration (3 join policies, circuit breaker, arbiter modes, per-branch timeouts, LLM integrity at join); the most mature verification (239+ matrix suites, 30+ runtime guards, process mining, autonomy check, parity program).

Honesty note — where alternatives lead: LangGraph and CrewAI have larger ecosystems, public marketplaces and extensive tutorials; CrewAI Flow builder and AutoGen Studio offer visual low-code construction; LangGraph Studio offers a streaming/visual orchestration UI. Soul currently has no public builder/marketplace and its documentation is internal. These are stated openly; the comparison is a capability-coverage assessment, not a quality ranking, and every claim about a compared system is referenced to its official documentation (LangGraph: docs.langchain.com; CrewAI: docs.crewai.com; AutoGen: microsoft.github.io/autogen; OpenAI Swarm: github.com/openai/swarm; A2A: a2a-protocol.org; MCP: modelcontextprotocol.io; Soar: soar.eecs.umich.edu; ACT-R: act-r.psy.cmu.edu; MAPE-K: IBM autonomic computing blueprint, 2003).

## AGI positioning

Soul is positioned as research toward evidence-based AGI, not a claim of having achieved it. The approach is reuse-first parity with established cognitive architectures (Soar, ACT-R), agent frameworks (LangGraph, CrewAI) and standards (DeepMind AGI levels, NIST AI RMF), tracked by measurable KPIs (agi_readiness_scorecard coverage, rmf_crosswalk coverage, quarterly ASIS parity reviews). The consciousness layer (P65) adds Ψ(t) dynamics, emotional memory, continuous cognition and meta-learning, integrated into the operational loop. The cognitive lineage cited is ACT-R (Anderson 2007), Soar (Laird, Newell & Rosenbloom 1987; Laird 2012), Global Workspace Theory (Baars 1988) and Integrated Information Theory (Tononi 2004).

## Agent environment and operations

Soul runs a multi-project, multi-station agent environment. External projects (SEPM) each have their own git repository, harvested knowledge (D12) and governed agent lanes, synchronized by a fleet controller with P62 process-registry parity and bidirectional git. Digital Employees are governed operational AI personalities with session contracts (assignment / regular / project), not autonomous clones. Deployment is one-button with a multi-stage Release Green Gate, multi-station sync, health monitoring and gate-based rollback.

## Scale (measured, not asserted)

Approximately 953K lines of code, 1.9M words of technical canon and 1.6K canon documents, measured by scripts (see /en/how-we-measure). Agent commit share over a 60-day window is reported from git authorship, with an explicit bootstrap-period caveat.

## Useful machine-readable resources

- Sitemap: https://soulplatform.ai/sitemap.xml
- Concise LLM reference: https://soulplatform.ai/llms.txt
- Release feeds: https://soulplatform.ai/feed.xml (RSS), https://soulplatform.ai/feed.json (JSON Feed), https://soulplatform.ai/releases-feed.json
- Structured data: Schema.org JSON-LD is embedded site-wide (Organization, WebSite, SoftwareApplication) and as FAQPage on the home page.
- robots.txt explicitly allows AI crawlers (GPTBot, ClaudeBot, PerplexityBot, Google-Extended, Applebot-Extended, Bingbot, Yandex, Baiduspider, cohere-ai, YouBot, meta-externalagent).

## Key page index

- /en/platform — platform overview, value layers, cognitive pipeline, KD12, deployment, self-development.
- /en/comparison — Soul vs LangGraph/CrewAI/AutoGen/Swarm/A2A/MCP/Soar/ACT-R, 10 dimensions, official sources.
- /en/agi-research — cognitive architecture, triad, consciousness, sleep pipeline, parity program, bibliography.
- /en/agi-roadmap — bounded-AGI trajectory and parity KPIs.
- /en/algorithms — deep dives (STRICT_JSON, memory 30/70, QNN, 6D/KHS, consciousness, Synapsis, Synopsis, ModeSelector, Delivery Guard, P30, Two-Keys, D0, KD12, NeuroTraining, DAG orchestration, ASIS).
- /en/cognitive-processor — P30, TaktEngine, ModeSelector, consciousness state service.
- /en/knowledge-dimensions — KD12 and 6D/KHS.
- /en/technology/quant-graph — governed semantic graph, TUR v2, TZ-COG clusters.
- /en/agent-environment — multi-project, multi-station IDE lanes.
- /en/agent-projects — SEPM external projects, fleet sync, P62, bidirectional git.
- /en/digital-workforce — governed operational AI personalities and session contracts.
- /en/security — RBAC, Two-Keys, fail-shut, lease ownership, signed delivery.
- /en/mcp-server — Model Context Protocol integration (9 tools + neural search).
- /en/how-soul-builds — measurable development autonomy.
- /en/how-we-measure — methodology behind every public figure.
- /en/investors — Human-in-Command, Tier 0 stage disclosure, engagement path (not an offer).
- /en/glossary — definitions of Soul-specific terms.
- /en/contact — investor/partner contact; public email invest@soulplatform.ai.