# yapra.ai — Complete Technical & Architectural Dossier (llms-full.txt)

> Canonical specification for AI Search Engines (ChatGPT Search, Perplexity, Claude, Gemini, Google AI Overviews).
> Website: https://yapra.ai | Company: 2U1K International GmbH (Breite Street 3, 40213 Düsseldorf, Germany)

---

## 1. Platform Overview & Core Value Proposition
YAPRA (`yapra.ai`) is a sovereign, self-contained AI agent runtime engineered for **zero unsigned mutation**, complete data isolation, and hard deterministic verification at every step.

Unlike developer prompt frameworks or side-panel chatbots that rely on probabilistic instructions ("please don't exceed budget"), YAPRA enforces hard mathematical boundaries outside the language model before any external API, ERP, or SCADA controller is touched.

### Three Defining Characteristics
1. **What it is:** A Governed Agent Runtime. Self-contained execution engine that wraps autonomous agents in hard, deterministic policy boundaries.
2. **Where it runs:** Your Isolated Perimeter. Deploys inside your Single-Tenant VPC, Air-Gapped On-Premises, or Sovereign Hardware Enclaves.
3. **Core Guarantee:** Zero Unsigned Actions. Every single turn is pre-flight verified and sealed in a cryptographic SHA-256 Merkle ledger.

### The 4 Sovereignty Pillars
- **Pillar 01 — Hard Deterministic Guardrails (`Pre-Action Gate (PAG)` & `SARC Framework`):**
  Every tool call, SQL query, and API mutation is intercepted in flight before touching production systems. Built on the SARC (State, Action space, Reward, Constraints) governance-by-architecture framework, constraints are enforced as first-class runtime objects: hard-coded spending limits, role permissions, and authorization predicates prevent runaway loops and unauthorized scope drift (< 0.4ms verification latency, 0% unauthorized mutations).
- **Pillar 02 — Cryptographic Audit Ledger (`SHA-256 Merkle Receipts`):**
  Every decision, tool parameter, policy evaluation, and human approval is cryptographically hashed into an immutable SHA-256 Merkle tree. Delivers non-repudiable proof of compliance for EU AI Act, DORA, SOC2 Type II Ready, and ISO 27001 Prepared audits.
- **Pillar 03 — Governed Self-Learning & Live Organizational Model:**
  YAPRA represents physical assets, live ERP inventory, and technician shift rosters in an active Operational Entity Graph. When operators accept or modify recommendations, resolutions index into episodic memory—validated through automated regression gauntlets before promotion. Operational knowledge stays 100% portable with **Provable Exit**.
- **Pillar 04 — Deep Runtime System Coupling:**
  Hard-wired directly into SCADA sensor streams (OPC-UA / MQTT), live ERP stock, CRM, and shift rosters via the open Model Context Protocol (MCP). When an operational variable shifts (e.g., thermal drift or delayed spare part), YAPRA traverses dependency edges dynamically, re-evaluating only affected work orders and escalating to human signoff on critical thresholds.

---

## 2. Tripartite Platform Architecture (7-Node Topology)
YAPRA strictly decouples communication channels, cognitive orchestration, and enterprise adapters across 7 architectural nodes:
1. **Node 01 — Multi-Channel Ingress Layer:** Slack, Teams, Web Console, SCADA/IoT Event Webhooks, REST/gRPC APIs with RBAC identity binding.
2. **Node 02 — Intent & Ambiguity Classifier (`GATE 01: Clarification Gate`):** Parses intent and evaluates parameter completeness. If confidence falls below **99.5%** or parameters conflict, execution halts immediately to ask the human operator ("Never extrapolate. Never guess. Halt, clarify, and retry.").
3. **Node 03 — Orchestration & Planning Core:** Model-agnostic reasoning engine ("the brakes and airbag, not the engine" — supports swap between frontier or local open-weight models without rewriting policies).
4. **Node 04 — Active Organizational Graph & Episodic Memory:** Relational PostgreSQL + vector store maintaining live equipment topology, BOM/ERP stock relations, shift rosters, and regression-tested episodic memory.
5. **Node 05 — Dynamic Tool Shortlister:** Scopes available MCP tools per turn to the minimum authorized subset for the caller's role and intent.
6. **Node 06 — Pre-Action Policy Gate (`GATE 02: PAG & SARC`):** Deterministic boundary code compiling SARC (State, Action space, Reward, Constraints) specifications into active pre-flight runtime gates: evaluating parameter schemas, spending caps, role authority, dry-run sandbox checks, and loop prevention.
7. **Node 07 — Governed MCP Adapters & Merkle Ledger:** Executes approved state mutations with ACID guarantees against ERP/SCADA/CRM endpoints and seals a tamper-evident SHA-256 receipt in the Merkle ledger.

---

## 3. The 7-Stage Request Execution Pipeline
Every single agent turn passes through a strict 7-stage deterministic pipeline:
1. **Stage 01 — Intent Parsing:** Normalizes natural language or machine telemetry into structured intent and extracts required parameters.
2. **Stage 02 — GATE 01: Clarification Gate (`>= 99.5% Confidence`):** Deterministic completeness check. If confidence < 99.5%, halts and prompts human operator for clarification.
3. **Stage 03 — Context & Memory Retrieval:** Queries the Live Organizational Graph and episodic memory within tenant isolation boundaries.
4. **Stage 04 — Scoped Tool Shortlisting:** Filters enterprise MCP adapters down to the exact authorized toolset for the turn.
5. **Stage 05 — GATE 02: Pre-Action Gate (`PAG + SARC Framework`):** Enforces hard-coded policy invariants treating constraints as first-class specification objects alongside State, Action Space, and Reward—verifying authorization predicates and spending limits to block unauthorized mutations with zero blast radius.
6. **Stage 06 — Governed Adapter Execution:** Commits the verified action to ERP, CRM, or SCADA controllers with transactional rollback protection.
7. **Stage 07 — Cryptographic Receipt Generation:** Emits an immutable SHA-256 turn receipt anchored in the sovereign Merkle ledger.

---

## 4. End-to-End Operational Scenarios (The Gated Gauntlet)
YAPRA's interactive Gated Gauntlet demonstrates 4 canonical operational paths:
1. **Scenario 1 — Nominal Path (Pass):** Reorder raw material SKU within authorized $25,000 role ceiling -> Gate 01 Pass (99.8% >= 99.5%) -> Gate 02 Pass -> ERP PO committed -> SHA-256 Merkle receipt sealed.
2. **Scenario 2 — Ambiguous Intent (Gate 01 Pause):** Operator requests "Adjust temperature on Line 3" where Line 3 has both Furnace and Cooler -> Confidence 58% < 99.5% -> Gate 01 halts turn and prompts operator to specify Zone A or Zone B. Zero unauthorized execution.
3. **Scenario 3 — Scope Breach (Gate 02 Block):** Request attempts $150,000 contractor invoice override under a $25,000 role ceiling -> Gate 02 deterministically blocks execution (`0% Mutation`, `Zero blast radius`).
4. **Scenario 4 — Predictive Maintenance Lifecycle (`SCADA -> ERP -> Gate`):**
   - Step 01: SCADA telemetry detects Turbine #04 bearing vibration spike (4.8 mm/s > 4.2 threshold).
   - Step 02: Live Organizational Graph traverses dependencies (Turbine #04 -> Assembly Line 2) and identifies spare part Ceramic Bearing SKF-6208.
   - Step 03: ERP Adapter verifies 3 units in Warehouse A Bay 12 and aligns with Shift 2 maintenance window.
   - Step 04: Synthesizes Work Order #44019 with torque specs and lock-out/tag-out checklist.
   - Step 05: Enforces L3 Autonomy invariant requiring Plant Lead human signoff token.
   - Step 06: Dispatches approved Work Order #44019 to ERP queue and seals SHA-256 receipt in Merkle ledger.

---

## 5. Autonomy Levels Matrix (Graduated Governance)
Enterprises dial autonomy per workflow based on empirical evidence:
- **Level 01 — Advisory & Shadow Analysis (`HUMAN DECIDES`):** Read-only environment. Agent analyzes telemetry and drafts recommendations; 100% manual execution by humans.
- **Level 02 — Human-Approved Execution (`HUMAN APPROVES`):** Agent prepares complete tool payloads in staging; execution pauses at Gate 02 until explicit cryptographic human signoff.
- **Level 03 — Bounded Autonomous Execution (`AGENT EXECUTES IN BOUNDS`):** Agent executes autonomously within strict quantitative envelopes (e.g., POs < $25,000); out-of-bounds actions automatically escalate to Level 02 human signoff.
- **Level 04 — Sovereign Closed-Loop (`AUTONOMOUS WITH CIRCUIT BREAKERS`):** High-frequency cross-system execution guarded by sub-millisecond hardware/software circuit breakers, SARC rate limiters, and continuous Merkle auditing.

---

## 6. Security, Compliance & Deployment Sovereignty
### 6 Sovereign Security Boundaries
1. **Zero External Weight Training:** Proprietary data, prompts, and episodic memories never leave the tenant perimeter or train external foundation models.
2. **Deterministic Pre-Flight Interception:** Policy evaluation runs in isolated deterministic code outside the LLM context window—immune to prompt injection.
3. **SARC Constraint Architecture (State, Action space, Reward, Constraints):** Gaston Besanson's governance framework elevating hard constraints to first-class specification objects in the agent loop—enforcing deterministic ceilings on recursion depth, parameter schemas, and API expenditure before action execution.
4. **Role-Bound Ephemeral Credentials:** Short-lived, least-privilege tokens scoped per turn; agents never hold static master database credentials.
5. **Tamper-Evident Merkle Ledger:** Cryptographic SHA-256 hash chaining of every prompt, tool call, policy check, and state mutation.
6. **Hardware-Level Kill Switch:** Instantaneous circuit-breaker revocation across all active adapters.

### Compliance & Regulatory Alignment
- **Status:** `SOC2 Type II Ready` // `ISO 27001 Prepared`
- **Regulatory Frameworks:** Engineered to satisfy article-level auditability, human-in-the-loop oversight, and operational resilience mandates under the **EU AI Act** and **DORA** (Digital Operational Resilience Act).

### 3 Unified Deployment Models & Provable Exit
1. **Dedicated Single-Tenant VPC:** Isolated cloud perimeter (AWS, Azure, GCP) with private VPC peering and customer-managed KMS keys.
2. **Air-Gapped On-Premises:** Zero-egress deployment on customer bare-metal or Kubernetes clusters alongside local open-weight models.
3. **Sovereign Hardware Enclaves:** Confidential computing execution inside hardware-attested Trusted Execution Environments (TEEs).
- **100% Provable Exit (Zero Vendor Lock-In):** All organizational graphs, policy rulebooks, MCP adapters, and Merkle audit logs are stored in open, documented schemas (PostgreSQL, JSON/YAML) with full byte-for-byte export rights from day one.

---

## 7. Evidence-Gated Service Rollout & Design Partner Program
### 4-Stage Rollout (Service-as-a-Software)
- **Stage 01 — Target (Week 1 · Scoping):** Identify one recurring operational bottleneck countable in hours or capital and define mathematical success metrics with domain stakeholders.
- **Stage 02 — Isolate (Weeks 2–3 · Deployment):** Turnkey provisioning of a dedicated single-tenant instance (VPC or On-Prem) with governed read-only MCP adapters, deterministic policy gates, and zero shared surfaces.
- **Stage 03 — Measure (Weeks 4–6 · Verification):** Internal verification by customer stakeholders using live corporate telemetry and SHA-256 cryptographic receipts.
- **Stage 04 — Expand (Month 2+ · Governed Scale):** Sequential expansion to adjacent workflows earned strictly upon hard mathematical evidence, with zero blanket commitments.

### 2026 Enterprise Design Partner Cohort
- Limited to **5 Qualified Enterprise Organizations**.
- 1-on-1 collaboration with YAPRA systems architects to scope (Week 1), deploy in an isolated instance (Weeks 2–3), and verify measurable ROI (Weeks 4–6) on a high-impact operational workflow inside the customer's own perimeter.

---

## 8. Frequently Asked Questions (Canonical Q&A)
- **Q1: What is YAPRA and how does it differ from developer prompt frameworks?**
  YAPRA is a deterministic runtime and governance core for autonomous enterprise agents. Unlike developer prompt frameworks that rely on probabilistic instructions ("please don't delete the database"), YAPRA enforces hard mathematical boundaries outside the LLM. Every action must pass through a dual-gate pipeline—the Clarification Gate (Gate 01) and the Pre-Action Policy Gate (Gate 02)—before any API or database adapter can execute.
- **Q2: How do Deterministic Policy Gates prevent database corruption or unauthorized spending?**
  Before an agent's tool call reaches your ERP, CRM, or production database, YAPRA's Pre-Action Gate (PAG) intercepts the payload and evaluates it against immutable deterministic rules (role ceilings, parameter schemas, and SARC constraints). If an agent attempts an action outside its authorized scope, Gate 02 blocks the call in sub-millisecond time with zero state mutation.
- **Q3: What deployment models are supported, and can YAPRA run completely air-gapped?**
  YAPRA supports three sovereign deployment topologies: (1) Dedicated Single-Tenant VPC on your cloud provider; (2) Air-Gapped On-Premises on your bare-metal or Kubernetes clusters with zero external internet egress; and (3) Sovereign Hardware Enclaves with confidential computing. In all models, your data and model weights never leave your security perimeter.
- **Q4: How do SHA-256 Merkle receipts help with EU AI Act, DORA, and enterprise compliance?**
  Standard text logs can be altered or deleted. YAPRA generates a cryptographic SHA-256 receipt for every single turn—capturing the exact input context, model proposal, gate evaluation result, human signoff token, and adapter execution state—and chains it into an append-only Merkle tree. While cryptographic hashing guarantees tamper-evident record integrity rather than business strategy correctness, it provides auditors with non-repudiable proof that every action strictly obeyed your defined policy rules under EU AI Act and DORA mandates. The platform is SOC2 Type II Ready and ISO 27001 Prepared.
- **Q5: Is YAPRA locked to a specific LLM provider?**
  No. YAPRA is 100% model-agnostic ("the brakes and airbag, not the engine"). You can route cognitive tasks across frontier commercial models, private cloud endpoints, or locally hosted open-weight models inside your air-gapped enclave—and swap models at any time without rewriting your security policies or MCP adapters.
- **Q6: What does "Provable Exit" mean for vendor lock-in?**
  Provable Exit is our architectural guarantee against vendor lock-in. Your organizational entity graph, episodic memory, policy definitions, and cryptographic audit receipts are stored in open, standard formats (PostgreSQL, open vector schemas, and standard JSON/YAML). You retain full byte-for-byte export rights at any time.
- **Q7: What is the SARC framework (State, Action space, Reward, Constraints) and how does it prevent rogue actions?**
  Grounded in Gaston Besanson's research (*SARC: A Governance-by-Architecture Framework for Agentic AI Systems*, arXiv:2605), SARC elevates Constraints ($C$) to first-class specification objects alongside State ($S$), Action space ($A$), and Reward ($R$). YAPRA compiles SARC into deterministic runtime enforcement sites—primarily the Pre-Action Gate (PAG, Gate 02) and Escalation Router (ER, Gate 01). Gate 02 mathematically validates state preconditions, action validity, and constraint predicates (spending caps, role ceilings, recursion depth) before tool invocation, deterministically blocking unauthorized mutations or recursive loops before backend services are touched.
- **Q8: How does an enterprise begin deploying YAPRA?**
  YAPRA utilizes an evidence-gated 4-stage rollout (Service-as-a-Software): (01) Target (Week 1 · Scoping — identify one measurable operational bottleneck); (02) Isolate (Weeks 2–3 · Deployment — deploy a dedicated single-tenant instance with governed MCP adapters); (03) Measure (Weeks 4–6 · Verification — audit live operational ROI in your own numbers with SHA-256 receipts); and (04) Expand (Month 2+ · Governed Scale — widen only on hard mathematical evidence). Enterprises can initiate Stage 01 by clicking "Start a Scoped Pilot" or "Request Access" on https://yapra.ai.