Compliance & Sovereign Governance
SBP PSO/PSP Payment Integration • NIST CSF 2.0 Framework • Islamic Fintech Principles
Project Pak-LLM Coder is engineered by SBF Consultancy (Pvt.) Ltd. (Lead Architect: Saad Bin Farrukh) to advance Pakistani software engineering sovereignty. We adhere to strict cybersecurity benchmarks, user-isolated workspace auto-resume, transparent domestic settlement in PKR via regulated rails, and zero-usury Islamic billing. Full technical documentation is available in the Documentation Portal.
01. SBP-Aligned Payment Rails (Swich Gateway)
All in-app domestic transactions, subscription token renewals, and custom apex domain purchases are routed through Swich Fintech, operating in compliance with State Bank of Pakistan PSO/PSP digital payment guidelines.
Instant domestic settlement via JazzCash, EasyPaisa, 1Link IBFT, and Visa/MasterCard.
Zero compounding interest, zero late payment penalties, and fixed transparent PKR pricing.
02. Cybersecurity Controls (NIST CSF 2.0 Alignment)
- Strict-Dynamic CSP: Cryptographic per-request nonces blocking unauthorized script injection and XSS attacks.
- Binary Integrity Verification (SHA-256): CLI and desktop binaries are integrity-checked via published SHA-256 sums at https://coder.pak-llm.com/downloads/SHA256SUMS. Both
install.shandinstall.ps1automated installers verify SHA-256 hashes locally before execution and fail-closed upon any discrepancy. - Minimal Operational Disclosure (/api/health): The public health endpoint at
/api/healthintentionally returns only{"status":"ok"}without leaking uptime, server internals, or environment fingerprints. - Cross-Origin-Opener-Policy (COOP): Cross-Origin-Opener-Policy is
same-origin-allow-popupsbecause OAuth popups (Universal SSO, Developer SSO, Cloud Auth) and domestic Swich payment windows require popup communication without breaking window isolation. - Security Contact & Disclosure Channel: Security contact details, vulnerability disclosure channels, and PGP encryption points are published in
/.well-known/security.txt(RFC 9116 compliant). - Prompt Security Sentinel: Real-time heuristic scrubbing of API keys, private tokens, and high-entropy secrets prior to model inference.
- Human-in-the-Loop Safeguards: Autonomous Web Agents and the Pak-LLM CLI cannot execute terminal commands, modify code, or commit files without explicit developer authorization. Destructive commands are actively blocked by terminal sandbox gates.
- Zero-Retention Local Execution: Local CLI tools execute in the user's isolated environment with zero telemetry or codebase exfiltration.
- Automated Edge TLS 1.3 & Certificates: Let's Encrypt automated certificate lifecycle (RSA-2048 with ECDSA forward secrecy, auto-renewed every 60 days) and HSTS preload (
max-age=63072000; includeSubDomains; preload). - Crawler & Robot Exclusion Architecture:
robots.txt Disallow: /api/instructs search indexing crawlers to avoid internal routes, while public edge monitoring endpoints (/api/health,/api/csp-report) are intentional, rate-limited public exceptions.
03. Registrar & ICANN Domain Governance
Custom apex domain registrations (.com, .pk, .com.pk, .ai, .tech, .io, .org, etc.) are fulfilled through accredited registrar rails under ICANN consensus policies. A 10% platform fee supports automated DNS cluster synchronization and DDoS mitigation.
04. Data Retention Schedules & DSR SLA (PECA / GDPR)
Pak-LLM Coder operates on strict data minimization principles:
- Account Identity: Retained during active account lifecycle; purged within 30 days of deletion request.
- Workspace Files: Active project session + 90 days post-dormancy before automated secure purge.
- Subscription / Quota Metadata: Active billing cycle + 12 months audit trail.
- Transient Logs & Security Telemetry: 30-day rolling rotation window with automatic sanitization.
- Payment Records: Retained for 5 years in compliance with statutory financial and SBP regulations (with 1% tax collected and settled via Swich).
- Data-Subject Request (DSAR) SLA: Data access, export, or complete account deletion requests are processed with an official 30-calendar-day SLA and 24-hour initial acknowledgement.
05. Incident Response & Security Disclosure Program
Project Pak-LLM maintains an active Vulnerability Disclosure Program under RFC 9116 guidelines (/.well-known/security.txt). Security researchers are encouraged to report potential vulnerabilities responsibly. Our dedicated incident response team triages reports within 24 hours and provides remediation updates throughout the lifecycle.
Parent Architecture: SBF Consultancy (Pvt.) Ltd. (sbf-consultancy.net) • Lead Architect: Saad Bin Farrukh
06. Sovereign SDK Security & Islamic Law / PECA Malpractice Auditing
The programmatic SDKs (@pak-llm/sdk & pakllm-coder) and all interactive surfaces (Mission Control, Web IDE, Terminal CLI) incorporate active ethical guardrails:
- Islamic Ethical Guardrails: Proactive filtering prevents the generation of predatory interest software (Riba), gambling or sports betting mechanics (Maysir), adult/obscene content (Fahsha), deceitful fraudulent scams (Gharar), intoxicant promotion (Khamr), or blasphemy.
- PECA 2016 Cybercrime Prevention: Prompts or autonomous agent requests to generate ransomware, exploit payloads, rootkits, credential dumping utilities, or system wipe commands are blocked fail-closed at the perimeter.
- Real-Time Malpractice Webhook: Any blocked infringement across Mission Control, Web IDE, CLI, or SDK is dispatched via secure webhook to the Pak-LLM central admin dashboard (
pak-llm.com/api/admin/coder/webhook) with project segregation (project: "Pak-LLM Coder") and specific subsystem tagging for administrative visibility. - Client-Side Secret Redaction: High-entropy API keys and passwords are redacted client-side before reaching network transport, upholding the Shariah principle of Amanah (trust and data sanctity).
07. Hybrid Architecture, Data Residency & Subprocessors Directory
Pak-LLM Coder employs a transparent Hybrid Sovereign Cloud Architecture. Client-side WASM sandboxes (Pyodide Python 3.12, V8 isolates) and local CLI environments execute on the developer's physical machine with Zero Remote Code Retention. Cloud-synced metadata, authentication, and domestic settlements use vetted, minimal subprocessors:
| Subprocessor | Location | Purpose & Service | Data Processed | Retention |
|---|---|---|---|---|
| SBF Consultancy (Pvt.) Ltd. Enclave | Pakistan 🇵🇰 (Rawalpindi/Islamabad) | Primary Sovereign Inference Cluster (PakLLM_Coder 32B) | Transient prompt context (Zero Code Storage) | 0 Days (Transient RAM) |
| Swich (Private) Limited | Pakistan 🇵🇰 (Karachi) | SBP PSO/PSP Regulated Payment Settlement (0% Riba) | Transaction ID, PKR amount, customer phone/email | 5 Years (Statutory SBP requirement) |
| Global Edge Network Inc. | United States / Global Edge | Global Anycast CDN, Edge SSL/TLS, Next.js Hosting | HTTP request headers, edge caching, IP for DDoS | 30-day rolling log |
| Google Cloud / Firebase | United States / EU Frankfurt | User Authentication, Identity Tokens & App Check | Account email, password hash, user tier metadata | Account lifecycle + 30 days |
| Neon Inc. | United States | Relational PostgreSQL Serverless Database | Encrypted workspace metadata & token counters | Account lifecycle |
| Hugging Face Inc. | United States / France | Secondary Fallback Foundation Model Checkpoint Router | Sanitized inference prompt (zero credentials) | 0 Days (Ephemeral pass-through) |
08. Mobile Application Security Standards (OWASP Mobile Top-10 & Biometric Enclave)
The downloadable Pak-LLM Coder Android (.apk / .aab) and iOS (.ipa) applications adhere strictly to the OWASP Mobile Application Security Verification Standard (MASVS):
- Hardware Key Isolation: Biometric authentication (Face ID, Touch ID, Android BiometricPrompt) leverages hardware-backed enclaves (Apple Secure Enclave and Android Keystore StrongBox). Cryptographic keys never enter application RAM unencrypted.
- Zero Telemetry & Ad Trackers: Completely strips Google Advertising ID (AAID) and Apple Identifier for Advertisers (IDFA). No third-party data broker analytics are present.
- Cryptographic Binary Verification: Standalone release packages are published with verifiable SHA-256 hashes on the sovereign ledger (/downloads/SHA256SUMS) and verified against digital code signing certificates.
- Memory Sanitization: App state and decrypted memory buffers are wiped on application suspend or background transition to prevent memory dump inspection.
09. Enterprise Mission Control Governance (ISO/IEC 25010, 27001, 42001 & NIST AI RMF)
The four-pillar Enterprise Mission Control platform is governed by a defence-in-depth control set. Every mission produces a tamper-evident audit chain and three immutable, content-addressed artifacts (Plan & Spec, Atomic Patch Set, Verification Matrix) that can be exported and independently verified.
- ISO/IEC 25010: Functional suitability through tri-surface verification, performance efficiency through token-budgeted architecture context slices, and security through pre-write AST validation.
- ISO/IEC 27001 A.12.4 / NIST AU-6: Hash-chained audit ledger covering mission creation, patch proposals, worker spawns, verification outcomes, merges and artifact persistence.
- ISO/IEC 42001 & NIST AI RMF: AI risk controls with per-worker isolation, capability permissions, human-in-the-loop merge gates, and measurable remediation iterations.
- NIST AC-3 / AC-6 least privilege: Capability policies and rate limits for filesystem writes, terminal execution, network fetches, worker spawns and Git worktrees. Filesystem writes, terminal execution and real worktrees require explicit approval.
- NIST SI-2 / SI-7: Deterministic self-healing for broken syntax, artifact digests, patch provenance and integrity verification on every artifact read.
- OWASP LLM01 / LLM02 / LLM08: Repository-context sanitization, no unsafe HTML rendering, sandboxed previews without same-origin, and approval/rate gating for excessive-agency operations.
- CWE-22 / CWE-78 / CWE-79: Path traversal rejection, shell-metacharacter assessment, and sandboxed iframe previews that block
javascript:URLs. - Real Git worktrees (opt-in): Server-side worktrees are disabled by default, require an admin secret plus an explicit trusted root, and execute
gitthrough an argv-onlyexecFilecall with no shell. - SGS reproducibility: Seeded hashes, stable JSON ordering and a 31-case Node regression suite validate every pillar, guard and failure path.
08. Islamic Shariah Code of Ethics, Immediate Prompt Rejection & SMTP Administrative Alerting
Pak-LLM Coder enforces a fail-closed, perimeter-level ethical firewall rooted in foundational Islamic Jurisprudence (Fiqh), the Pakistan Electronic Crimes Act (PECA 2016), and international AI safety standards (ISO/IEC 42001, OWASP LLM01/LLM02):
coder.ethical_violation_blocked), and sends an encrypted SMTP alert directly to system administrators at [email protected].Compliance & Governance FAQs
Key principles regarding security controls, domestic banking rails, and Islamic Fintech standards.
Every mission runs a governed closed loop: architecture indexing, BPE-budgeted context injection, isolated parallel workers, AST-guarded patches, tri-surface verification, deterministic self-healing and a human sign-off merge gate. Each transition is written to a hash-chained audit ledger, and immutable Plan, Patch Set and Verification Matrix artifacts carry content-addressed integrity hashes. Capability permissions and rate limits enforce least privilege across filesystem, terminal, network and Git worktree operations.