Open Digital Product FactorySingle source of truth
Definition → derivation → public read model
packages/storefront-templates/srcderiveOperationalValueStreamDrift prevention
Generated pages are checked
The generator resolves every published leaf against ALL_ARCHETYPES, calls the canonical operational-value-stream derivation, and fails in --check mode when committed HTML differs.
Deliberate boundary
Category copy is only shared context
Category pages may explain common concerns. They do not define one category-wide process. Exact stages, constraints, and modules appear only on canonical leaf projections.
86 published leaf profiles resolve to source
The committed artifacts are generated from packages/storefront-templates/src. Broader platform architecture and standards remain in the architecture documentation.
For architects & standards reviewers
DPF as a reference implementation
DPF is a working prototype for two proposed standards — not a claim of full present-day conformance. TAK governs what an agent may do; GAID governs who the agent is and what evidence follows its actions. Here is what is real today, what is partial, and what is still proposed — the gaps are mapped, not hidden.
TAK — Trusted AI Kernel
The runtime-governance harness: authority mediation, tool-execution gating, human-in-the-loop, memory limits, and audit/evidence — enforced as an agent operates.
GAID — Global AI Agent Identification & Governance
Who an agent is, the claims it carries, and how its actions are identified and traced — identity documents, badges, authorization classes, and action receipts.
Implemented now
- Route-scoped AI coworkers — each route gets a purpose-built coworker with its own tools and vocabulary
- Tool authority = grant ∩ user capability — every tool exposure filtered by user authority, mode, and external-access posture
- Proposal-mode gated actions — consequential actions break the loop and return an approval payload instead of executing
- Tool-execution audit logging — every call written to
ToolExecution— agent, user, tool, params, result, route, duration, audit class - Archetype-driven UX & vocabulary — storefront, scheduling, finance and compliance defaults generated from one archetype
- Internal agent registry — stable identifiers, model bindings, supervisors, delegates, grants, HITL defaults
Partially implemented
- TAK runtime transparency — an Authority & Audit workspace with a supervisor-ready Agent Card snapshot
- Agent Card / AIDoc projection — the registry approximates an AIDoc — not yet signed or resolvable
- Stable internal agent identity — platform-local today, not yet canonical GAID identifiers
- HITL tier metadata — carried per agent, not yet one uniform runtime policy engine
- Delegation & authority modeling — supervisor & delegation recorded, not yet a receipt-backed chain
- Chain-of-custody trace — agent, user, route and tool recorded; not yet end-to-end across boundaries
Planned / proposed
- Signed / tamper-evident receipts — cryptographically verifiable receipts for consequential actions
- GAID public/private namespaces — explicit private vs externally-accredited public identity scopes
- Public verifier metadata — published A2A Agent Cards and external receipt-status endpoints
- External certificates & status — issuer validation and public status services for exposed agents
- Capability & governance badges — evidence-backed assurance levels
- Repeatable conformance suites — automated TAK & GAID test packs preserved as evidence
Grounded in the platform’s own first-pass assessment: TAK / GAID conformance assessment · TAK · GAID · governance white paper. DPF is the proving ground where these ideas are exercised in real workflows.