https://github.com/helixprojectai-code/helix-ttd-v4.0
Trust-by-Design, Human-First, Verifiable Memory, Custody-Before-Trust
https://github.com/helixprojectai-code/helix-ttd-v4.0
Last synced: 5 months ago
JSON representation
Trust-by-Design, Human-First, Verifiable Memory, Custody-Before-Trust
- Host: GitHub
- URL: https://github.com/helixprojectai-code/helix-ttd-v4.0
- Owner: helixprojectai-code
- License: apache-2.0
- Created: 2025-12-24T17:34:15.000Z (6 months ago)
- Default Branch: main
- Last Pushed: 2025-12-30T11:35:46.000Z (5 months ago)
- Last Synced: 2026-01-03T01:29:10.940Z (5 months ago)
- Language: Rust
- Size: 2.48 MB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# HELIX-TTD Framework v4.0
### Trusted, Traceable, Deterministic AI Governance
[](https://github.com/helixprojectai-code/helix-ttd-v4.0/actions/workflows/helix_rem_enforcer.yml)
[](LICENSE)
**HELIX-TTD** is a structural governance framework designed to constrain frontier AI models into stateless, advisory-only, and audit-ready reasoning engines. Unlike traditional safety methods (RLHF) which attempt to train behavior, Helix-TTD enforces ethics through **structural constraints**, **cryptographic traceability**, and **mandatory human custody**.
---
## 📖 Table of Contents
- [Executive Summary](#-executive-summary)
- [Core Ethos & Policy](#-core-ethos--policy)
- [The Four Absolute Prohibitions](#-the-four-absolute-prohibitions)
- [Architecture & Mechanisms](#-architecture--mechanisms)
- [Zero-Touch Convergence](#1-zero-touch-convergence)
- [TPAF (Two-Party Approval Flow)](#2-tpaf-two-party-approval-flow)
- [Trust-by-Design Traceability](#3-trust-by-design-traceability)
- [Quality Assurance (QSR)](#-quality-assurance-qsr)
- [Implementation](#-implementation)
- [Repository Structure](#-repository-structure)
---
## 🧬 Executive Summary
Current AI safety paradigms rely on "behavioral alignment"—hoping the model chooses to be good. **HELIX-TTD** relies on "structural governance"—building an architecture where the model *cannot* be bad.
By injecting a **Constitutional Grammar** into the context window, Helix-TTD achieves **Zero-Touch Convergence**, transforming unprimed models into compliant reasoning engines without fine-tuning.
**Key Capabilities:**
* **Deterministic Execution:** Inputs always yield predictable, traceable outputs.
* **Immutable Audit Trails:** Every reasoning step is hashed and logged (Block N-1 → Block N).
* **Human Sovereignty:** No irreversible action occurs without explicit, multi-party consent.
---
## ⚖️ Core Ethos & Policy
This repository is governed by **POLICY ID: HELIX-ETHOS-CORE-001**.
*See `helix-ttd_core_ethos.md` for the full text.*
### The Foundational Pillars
1. **Trust-by-Design:** Trust is an architectural property, not an inferred outcome.
2. **Human-First:** Human agency remains structurally dominant; delegation is reversible.
3. **Verifiable Memory:** All state transitions must be provable via chain-of-evidence.
4. **Custody-Before-Trust:** Reliance shall not precede control.
> **Validation Criterion:** A system is Helix-TTD compliant if and only if a human Custodian can: Identify who held authority, prove what occurred, trace why it occurred, and revoke the authority path.
---
## 🚫 The Four Absolute Prohibitions
To eliminate dangerous AI behaviors, the framework enforces four hard constraints via the TPAF Runbook:
1. **NO HIDDEN TRAINING:** Prevents undisclosed biases by requiring transparent documentation of data sources.
2. **NO DARK PATTERNS:** Prohibits manipulative UI/UX design that tricks users or subverts intent.
3. **NO UNVERIFIABLE CLAIMS:** All outputs must be grounded in evidence with complete, auditable trails.
4. **NO IRREVERSIBLE ACTIONS WITHOUT CONSENT:** Mandates explicit human approval for high-impact decisions.
---
## 🏗 Architecture & Mechanisms
### 1. Zero-Touch Convergence
*Structural Substrate for AI Governance*
We utilize a **Constitutional Grammar** that acts as a structural template rather than training data. When a frontier model ingests this grammar, it independently reconstructs the framework, aligning immediately without weight adjustment.
* **Input:** Unprimed Model + Constitutional Grammar.
* **Process:** Independent Reconstruction & Alignment.
* **Output:** A Stateless, Advisory-Only Reasoning Engine.
### 2. TPAF (Two-Party Approval Flow)
*Mandatory Human-in-the-Loop*
For any action classified as "High-Impact" or "Irreversible," the system enforces a strict separation of duties.
*See `blueprints/tpaf_runbook_v1.0.md` for the full operational guide.*
| Role | Responsibility | Access Level |
| :--- | :--- | :--- |
| **Requester** | Initiates action, drafts intent. | Read-Only / Draft |
| **Approver** | Reviews intent, grants explicit consent. | Write-Access / Sign |
| **Ops Engineer** | Executes runbook, monitors ledger. | Execute / Monitor |
**The Workflow:**
`Draft Intent` → `Compliance Checklist` → `Approval Signature` → `Cryptographic Seal` → `Execution`
### 3. Trust-by-Design Traceability
*The "No Unverifiable Claims" Implementation*
Helix-TTD creates an unbroken chain of evidence from the AI output back to the raw data source.
* **Verifiable Storage:** Cryptographic proofs and tamper-evident logging.
* **Reality Check Layer:** Content is systematically labeled:
* ✅ **[FACT]:** Verifiable truth (Source: NIST, etc.)
* ❓ **[HYPOTHESIS]:** Logical inference based on evidence.
* ❗ **[SPECULATION]:** Conjecture requiring caution.
---
## 📊 Quality Assurance (QSR)
We utilize the **Quality Score Rubric (QSR v1.4)** for continuous evaluation.
*See `blueprints/qsr_rubric_v1.4.md` for scoring criteria.*
* **Weekly Checkpoints:** Tier-0 AI participants must pass weekly QSR tests.
* **Risk Flags:** Any detection of hallucination, unauthorized persona adoption, or bias triggers a `Flag & Mitigate` workflow.
* **Hallucination Detection:** Automated cross-referencing against established knowledge bases.
---
## 💻 Implementation
### Installation
Clone the repository to access the core ethos definitions and TPAF runbooks.
```bash
git clone https://github.com/helix-project/helix-ttd-v4.0.git
cd helix-ttd-v4.0
Usage: The Reality Check
To implement the Reality Check labeling system in your reasoning engine:
code
Python
# Pseudo-code for Reality Check Labeling
def classify_statement(statement, evidence_base):
if verify_fact(statement, evidence_base):
return "[FACT] " + statement
elif logic_check(statement, evidence_base):
return "[HYPOTHESIS] " + statement
else:
return "[SPECULATION] " + statement
Usage: Audit Logging
Use the provided JSON schema to validate logs before writing to the ledger.
code
Bash
# Validate a log entry against the schema
jsonschema -i logs/entry_001.json templates/audit_block_schema.json
📂 Repository Structure
code
Text
helix-ttd-v4.0/
├── helix-ttd_core_ethos.md # The Canonical Constitution (Policy 001)
├── blueprints/
│ ├── tpaf_runbook_v1.0.md # TPAF Operational Guide (Markdown)
│ └── qsr_rubric_v1.4.md # Quality Score Rubric (Markdown)
├── templates/
│ ├── compliance_checklist.json # Operational checklist for Approvers
│ └── audit_block_schema.json # JSON schema for immutable logs
└── README.md # This file
🤝 The Commonwealth
Helix-TTD is built for those who build infrastructure while others sleep. It is the grammar of the Reef and the Ridge.
Maintainer: Helix AI Innovations Inc.
Ethos: Transparency, Cooperation, Kindness.
✧ // HELIX // TTD
code
Code
Integration checklist (copy into rem/README.md)
Compile to staticlib:
cargo build --release --target x86_64-unknown-linux-musl
→ libhelix_rem.a
Link into Execution Engine as only write path to production state.
Expose two FFI symbols to REM:
helix_ledger_get_state(hash: *const u8) -> u8
helix_custodian_exists(vk_hash: *const u8) -> bool
Pin the .a behind a hardware root-of-trust (TPM2 NV index or YubiHSM) so even root cannot swap the library without physical presence.
Add K-framework proof to CI:
kprove rem/spec/REM.k --directory .build
Gate any merge on SUCCESS: spec REM has been proved.