https://github.com/sbay-dev/amad-proj
AMAD CNS Fintech Secure Lab - Arabic agreement and smart-contract comparison demo
https://github.com/sbay-dev/amad-proj
ai arabic cns fintech hackathon legaltech smart-contracts
Last synced: 15 days ago
JSON representation
AMAD CNS Fintech Secure Lab - Arabic agreement and smart-contract comparison demo
- Host: GitHub
- URL: https://github.com/sbay-dev/amad-proj
- Owner: sbay-dev
- Created: 2026-07-02T20:11:46.000Z (27 days ago)
- Default Branch: main
- Last Pushed: 2026-07-07T08:47:32.000Z (22 days ago)
- Last Synced: 2026-07-07T10:23:48.729Z (22 days ago)
- Topics: ai, arabic, cns, fintech, hackathon, legaltech, smart-contracts
- Size: 259 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- Support: docs/supporting-ecosystem.md
Awesome Lists containing this project
README
# CNS for AMAD 2026
## From native Arabic Linux text infrastructure to measured AI decisions
[](https://github.com/sbay-dev/amad-proj)
[](#production-experiences)
[](docs/README.md)
[](docs/claim-boundaries.md)
[](docs/rtl-vt.md)
**CNS — Cubic Network System** is an Arabic-first, non-Transformer research and
engineering program created and led by **Sultan AlYami**, a full-stack software
engineer with 15 years of development experience, five years of focused AI work,
and an entrepreneurial mission to make Arabic a first-class language across the
technology stack. It began with a practical fintech question: how can an
unresolved structural residual in a high-risk contract become an explicit route
and a reviewable decision instead of remaining hidden inside a similarity score?
The AMAD submission demonstrates that idea through three complementary
production experiences and one controlled contract-intelligence release:
```text
Arabic input
→ position and semantic surface
→ residual and uncertainty evidence
→ route selection
→ bounded decision
→ human review or verifiable action
```
> CNS is not presented as a generic language model or an unbounded “general AI”
> claim. It is a design-science framework and a family of measured systems for
> Arabic representation, routing, retrieval, and decision support.
## Founder and laboratory mission
**Sultan AlYami** is an entrepreneur and the owner of the **Sbay Supply
Platform** trademark, registered in Nice Classification Class 42, the class
associated with scientific and technological services, research and design,
industrial analysis and research, laboratory services, and software and
computer-system development.
This foundation is being developed into an independent Arabic technology
laboratory: a place where language intelligence, fintech verification, secure
software, and foundational Arabic computing infrastructure can be researched,
measured, and transferred into production systems. CNS is the laboratory's
intelligence nucleus, but the mission extends beyond AI models to the deeper
technical assumptions that have historically treated Arabic as an adaptation
layer rather than a native system language.
### Foundational infrastructure track — RTL-VT
The **RTL-VT** research line addresses a root-level limitation in the classic
Linux Virtual Terminal: its screen-cell and cursor model assumes left-to-right
presentation. Instead of hiding that limitation behind a graphical terminal or
an application-side BiDi overlay, RTL-VT introduces opt-in right-to-left line
semantics inside the VT path itself while preserving the default LTR behavior.
The kernel remains responsible for direction, logical cursor state, physical
cell placement, deletion, clearing, scrolling, and virtual-console behavior.
Arabic glyph shaping and the PSF2 font payload remain explicitly separated in a
small user-space loader, so the kernel is not turned into a Unicode or shaping
engine. QEMU evidence covers cursor origin, backspace, CSI-K, CSI-X, multiline
placement, modern connected Arabic rendering, and ncurses applications including
`vim`, `less`, `mc`, and `htop`.
**Read the evidence-bounded technical dossier:** [`docs/rtl-vt.md`](docs/rtl-vt.md)
[](https://kernelwave.sbay.com.sa)
## Production experiences
### 1. CNS Learning Arena — interactive fintech education
**Live:** [https://cns-sg.sbay.sa](https://cns-sg.sbay.sa/)
An Arabic-first educational platform that explains verification, ordering,
risk, and context through four connected interactive experiences: a guided lab,
a race, an arena, and a campaign. Instead of asking evaluators to decode an
architecture diagram, the experience lets them observe how a financial decision
changes as evidence, sequence, and uncertainty change.
**Engineering characteristics**
- bilingual Arabic/English interaction;
- WCAG 2.2 AA accessibility target;
- skippable guidance and progressive disclosure;
- explicit educational-simulation boundary;
- no connection to real customer, banking, or transaction data.
**CNS role:** the platform makes the progression from position and residual to
route and decision understandable through interaction. It is a learning and
demonstration surface, not evidence by itself of model accuracy.
---
### 2. SarmadAi — Arabic conversational intelligence and semantic memory
**Live:** [https://sarmadai.sbay.sa](https://sarmadai.sbay.sa/)
SarmadAi is the conversational and orchestration member of the CNS ecosystem.
It combines an Arabic-capable chat experience with session-aware retrieval,
document memory, tool execution, provider governance, and observable model
operations.
**Engineering characteristics**
- ASP.NET Core on .NET 10;
- encrypted CNS ONNX embeddings for semantic retrieval;
- AskRAG document and session memory;
- model/provider routing with measured versus estimated telemetry separated;
- MCP-compatible tools and controlled external-tool registration;
- server-side key handling and encrypted persistence boundaries;
- containerized production runtime.
**CNS role:** CNS supplies a semantic representation layer for retrieval and
memory. It helps SarmadAi locate relevant Arabic context and preserve model
family provenance; the surrounding chat, tools, authentication, telemetry, and
deployment services remain separately classified software components.
---
### 3. Newsboy — bounded Arabic financial-news intelligence
**Live:** [https://newsboy.sbay.sa](https://newsboy.sbay.sa/)
**Technical dossier:** [`docs/newsboy.md`](docs/newsboy.md)
Newsboy evaluates real Arabic financial-news candidates before selection. It
distinguishes financially material disclosures from general context and
out-of-domain content, estimates materiality, exposes confidence, and routes
uncertain cases to human review.
**Accepted release evidence**
| Evidence gate | Result |
| --- | ---: |
| C#/Python runtime parity | `30/30` |
| Transparent financial cases | `16/16` |
| Transparent out-of-domain cases | `21/21` |
| Sealed scope accuracy | `93.33%` |
| Sealed materiality accuracy | `90.00%` |
| Scope calibration error | `0.0615` |
| Reference/evaluation leakage | None detected |
An independent 400-record public-RSS evaluation reports **57.8005% clean
lexical out-of-domain recall** with **5.75% abstention**. This result is not
hidden or promoted into a release score: it is a material limitation showing
that real-world generalization remains incomplete.
**CNS role:** a frozen CNS semantic representation is combined with trained
Newsboy task heads for relevance, materiality, and out-of-domain routing. RSS
collection, de-duplication, UI rendering, telemetry, encryption, and deployment
are deterministic operational tools rather than additional neural intelligence.
Authorized evaluators can inspect the exact source, clean-clone reproduction
evidence, and encrypted release package through the private
[`sbay-dev/CNS-Model-Trade`](https://github.com/sbay-dev/CNS-Model-Trade)
repository. Keys, plaintext weights, private corpora, and internal training
optimization scripts remain outside GitHub.
## Core AMAD fintech capability
### Arabic agreement ↔ smart-contract review
The controlled CNS-Fintech release addresses the gap between Arabic legal text
and executable institutional logic:
1. accept an Arabic agreement or clause;
2. infer parties, obligation, beneficiary, condition, amount role, and route;
3. construct a canonical intermediate representation;
4. generate or inspect Hyperledger Fabric-compatible Go;
5. compare high-impact fields and cryptographic bindings;
6. expose matches, differences, correction proposals, confidence, and rejection
evidence for human review.
The current `v1-rc1` external development benchmark is deliberately reported
without metric inflation:
| Gate | Result | Required | Decision |
| --- | ---: | ---: | --- |
| Exact evaluation-contract matches in training | `0` | `0` | PASS |
| Maximum word-trigram similarity | `5.00%` | below registered limit | PASS |
| Contract route accuracy | `70.00%` | `80.00%` | **BLOCKED** |
| Executable-action accuracy | `85.71%` | `70.00%` | PASS |
| Go module tidy / test / vet | PASS | PASS | PASS |
The three contracts were excluded from training, but they influenced iterative
development and are therefore described as an **external development
benchmark**, not a pristine blind test. Promotion to `v2` remains blocked.
Authorized evaluators can access the encrypted reviewer theater and reproducible
source package through the private
[`sbay-dev/CNS-Model-Fintech`](https://github.com/sbay-dev/CNS-Model-Fintech)
repository. The public repository intentionally excludes model keys, plaintext
weights, private corpora, and internal training-optimization pipelines.
## System architecture
```mermaid
flowchart LR
A[Arabic text, document, disclosure, or agreement]
B[CNS representation surface]
C[Neural task heads]
D{Confidence and residual gate}
E[Route and bounded decision]
F[Human review]
G[Verifiable application action]
A --> B --> C --> D
D -->|uncertain, rejected, or incomplete| F
D -->|within registered gate| E
E --> G
```
| Layer | Responsibility | Representative technology |
| --- | --- | --- |
| Representation | Arabic semantic surface and fixed release contract | encrypted CNS ONNX, deterministic feature contracts |
| Neural decision | task-specific route, relevance, materiality, or contract-slot prediction | non-Transformer residual-gated and linear neural heads |
| Verification | hashes, confidence, rejection, parity, leakage, and release gates | SHA-256, Merkle manifests, C#/Python parity, fail-closed evaluators |
| Product runtime | secure user experience and operational integration | ASP.NET Core, .NET 10, ONNX Runtime, Cloudflare Containers |
| Human governance | review of uncertain or high-impact decisions | explicit abstention, claim boundaries, evidence dossiers |
## Why this submission matters
- **Arabic-first by design:** Arabic is treated as a structurally rich domain,
not as a translated afterthought.
- **Full-stack Arabic technology:** the program spans kernel-level text
direction, semantic representation, AI routing, secure runtimes, and live
product experiences.
- **From representation to decision:** the work connects language, semantic
memory, routing, and verifiable institutional action.
- **Production evidence, not mockups:** all three linked experiences are live.
- **Security-conscious delivery:** model artifacts are encrypted, runtime keys
are external secrets, and public documentation avoids operational secrets.
- **Fail-closed scientific governance:** a successful build or attractive UI
cannot override a failed semantic gate.
- **Honest limitations:** positive results and blocking evidence are presented
together.
## Scientific and reviewer material
- [CNS design-science foundation](docs/README.md)
- [Project overview](docs/project-overview.md)
- [Supporting ecosystem](docs/supporting-ecosystem.md)
- [Claim and use boundaries](docs/claim-boundaries.md)
- [RTL-VT Arabic Linux infrastructure dossier](docs/rtl-vt.md)
- [Newsboy technical presentation](docs/newsboy.md)
- [English Newsboy v4 scientific paper](docs/papers/CNS_NEWSBOY_V4_SCIENTIFIC_PAPER_EN_20260711.pdf)
- [Arabic Newsboy v4 scientific paper](docs/papers/CNS_NEWSBOY_V4_SCIENTIFIC_PAPER_AR_20260711.pdf)
- [Newsboy v4 public technical data](docs/papers/CNS_NEWSBOY_V4_TECHNICAL_DATA_20260711.md)
- [Arabic RTL-VT scientific paper](docs/papers/RTL_VT_NATIVE_RTL_LINUX_CONSOLE_PAPER_AR_20260711.pdf)
- [RTL-VT public technical evidence](docs/papers/RTL_VT_TECHNICAL_EVIDENCE_20260711.md)
- [RTL-VT publication integrity manifest](docs/papers/RTL_VT_PUBLICATION_MANIFEST_20260711.json)
- [English CNS-Fintech scientific paper](docs/papers/CNS_FINTECH_NON_TRANSFORMER_PAPER_EN_20260706.pdf)
- [CNS AskRAG Universal v3 paper](docs/papers/CNS_ASKRAG_UNIVERSAL_V3_EMBEDDING_PAPER_EN_20260707.pdf)
- [Public technical data](docs/papers/CNS_FINTECH_MODEL_TECHNICAL_DATA_EN_20260706.md)
- [Public presentation](presentation/AMAD-CNS-Fintech-Submission-public.pptx)
- [Sample agreement](demo/sample-agreement.md)
- [Sample smart-contract signature](demo/sample-smart-contract-signature.txt)
- [Sample comparison report](demo/sample-comparison-report.json)
## Team
| Member | Discipline | Role |
| --- | --- | --- |
| **Sultan AlYami** | 15 years full-stack software engineering; 5 years AI development | Founder, team lead, model architect, and scientific author |
| **Salha Sharahili** | Sharia and law | Legal-domain review |
| **Mona Sharahili** | Cybersecurity | Security review |
Contact details are maintained in the official submission platform rather than
published in this public repository.
## Claim boundary
The submission demonstrates bounded research and production engineering for
Arabic educational interaction, conversational retrieval, financial-news
triage, and contract-logic review. It is not a fatwa, legal advice, investment
advice, trading system, price predictor, article-truth oracle, banking
certification, regulatory endorsement, live smart-contract execution, proof of
profitability, general intelligence, or evidence of superiority to qualified
human experts.