Skip to content

Latest commit

 

History

History
62 lines (49 loc) · 3.83 KB

RESPONSIBLE_AI.md

File metadata and controls

62 lines (49 loc) · 3.83 KB

Responsible AI at RAPP

Status: DRAFT · Owner: to be assigned (accountable RAI owner) · Last updated: 2026-06-27

RAPP is a platform for building and running AI agents. That power carries responsibility. This document is the front door to RAPP's Responsible AI program — how we align with the Microsoft Responsible AI Standard v2 (the six principles: Accountability, Transparency, Fairness, Reliability & Safety, Privacy & Security, Inclusiveness) and its generative-AI requirements.

Honest posture

RAPP is not RAI-signed-off today. We say so plainly. RAPP's design is, by intent, permissive: the brainstem executes arbitrary user-authored Python agents in-process, auto-installs their dependencies, can stream other people's agents in from the public Cave, and adds no content-safety layer of its own — it relies on the underlying GitHub Copilot model. Those are exactly the surfaces a Responsible AI review exists to gate.

The Responsible AI Roadmap is the documented, honest path from here to a defensible sign-off. Phase 0 removes every absolute blocker (the gating governance artifacts, containment of the code-execution and supply-chain surfaces, a safety layer at the /chat seam, and truthful data-flow disclosure); later phases harden, monitor, and polish.

The Phase-0 blockers, at a glance

Blocker Principle
No RAI Impact Assessment (the Standard's gating artifact) Accountability (A1)
No Sensitive Uses determination / per-agent screen Accountability (A2)
No content-safety / prompt-shield layer — safety delegated to Copilot Reliability & Safety (RS1)
Arbitrary code execution + auto-pip-install + 0.0.0.0/wildcard-CORS Privacy & Security (PS2)
Unsigned, git-invisible Cave-streamed third-party agents Privacy & Security (PS2)
No data-flow disclosure (conversation goes to GitHub/Microsoft Copilot) Privacy & Security / Transparency (PS1/T3)
No Transparency Note Transparency (T2)

What's here

How it connects to RAPP's governance

The Impact Assessment is wired as a release gate in CONSTITUTION.md: a VERSION bump may not ship until the assessment is current and its P0 exit criteria hold, and the accountable RAI owner signs off each release. RAPP's existing strengths — Article VII (privacy by default, no telemetry/accounts), the PUBLIC_BOUNDARY.md "bones not substance" scrub, rappid provenance/lineage, and the flight recorder — are the foundation this program builds on.

Reporting a concern

Found a safety, privacy, or misuse concern in RAPP or a RAPP-built agent? Open an issue on this repo, or contact the accountable RAI owner (above). Sensitive-Use concerns follow the escalation path in sensitive-uses.md.