Policies, standards, research, or technical records are registered with their source.
INDEPENDENT AI GOVERNANCE BUILD
A safety and review layer for AI-assisted work.
I built ChangeFront as an independent local application that sits between an AI assistant and the recommendations or document changes it proposes. It gives the AI approved evidence, keeps every supporting source visible, and requires a person to make the final decision.
The AI proposes a recommendation or document change and cites the exact evidence used.
A human reviewer approves or rejects the work. The AI cannot approve itself.
Evidence changes, drafts, decisions, and failures remain available for later review.
RUN THE SYSTEM
From download to review in one local command.
The package uses synthetic data, Python 3.12, SQLite, and no runtime package dependencies. Extract it, run the launcher from PowerShell, and inspect the seeded decision trail on loopback only.
.\launch-demo.ps1The launcher creates an ignored local database, binds to 127.0.0.1:8765, waits for health, and opens the reviewer interface.
TRACE ONE DECISION
The system makes the path from source to authority explicit.
HTTPS provenance and applicability define what a source is allowed to support.
Every accepted change becomes a new content-hashed record with append-only history.
The recommendation records the exact evidence versions supporting its reasoning.
AI may draft and submit. Only an explicitly confirmed human reviewer can decide.
A SHA-256 hash chain exposes missing, reordered, or modified audit events.
REPRODUCIBLE EVALUATION
Exact results from one machine-readable report.
The verifier runs the entire suite three consecutive times with a temporary SQLite database, fixed clock, sequential identifiers, disabled network access, and normalized output comparison.
.\verify.ps1SERVICE + STORAGE30 tests
Governance rules, immutability, approval boundaries, invalidation, and audit integrity.
REAL HTTP + API14 tests
Loopback server behavior, request contracts, review states, and error handling.
MCP SUBPROCESS19 tests
Discovery, request metadata, tool catalog, framing, results, and protocol errors.
DETERMINISMIdentical hash
4b6320aee0b39af1faa9b337d779ddc46655a9a0d5148f458e3f2fc9ff74436b
MCP INTERFACE
Agent access is useful because it is bounded.
The stateless stdio server follows MCP 2026-07-28 with per-request metadata, server discovery, deterministic tool listing, structured results, and standard protocol errors. Every adapter calls the same policy-enforcing service.
Read the official MCP specificationEXPOSEDEvidence, drafting, submission, staleness, summaries, and audit verification
INTENTIONALLY ABSENTApproval, decision, document apply, shell, and external network access
The interface can accelerate analysis without inheriting the authority to approve or apply its own work.
TRUST BOUNDARIES + NON-CLAIMS
What this reference does not prove.
This is portfolio work using synthetic organizations, sources, controls, records, and outcomes. It contains no employer code or data.
It is a single-user loopback reference implementation, not a hosted deployment or evidence of organizational adoption.
Reviewer headers demonstrate the service boundary but are not a complete authentication or identity system.
Approved proposals remain recorded proposals. The system does not write target documents or external systems.
The hash chain detects local audit tampering but is not anchored to an external transparency or signing service.
Staleness evaluation is invoked by the reviewer or tool interface; no unattended scheduler or source connector is claimed.
INSPECT THE IMPLEMENTATION
Review code, controls, and results.
Public GitHub repositorySource, tests, documentation, verification, and commit history->README + quick startPlain-English scope, commands, repository map, and standards references->ArchitectureAdapters, service boundary, state model, and failure paths->Trust modelAuthority, threat assumptions, limitations, and non-claims->Reviewer guideA short, reproducible technical inspection path->MCP referenceProtocol metadata, discovery, tools, results, and omissions->Synthetic fixtureRepresentative inputs with no private or employer data->
ROLE FIT
AI solutions leadership with implementation depth.
This case is one proof point in a broader portfolio spanning regulated AI architecture, governance and enablement, MCP integration, software delivery, and physical AI foundations.