Digital Twin Universe

Where AI-built code faces reality

The Gap

"All tests pass" only ever means it passed in the environment the AI built it in

By default AI-generated software is verified in the same environment and context that it was built — which frequently leads to agents claiming it worked due to context poisoning. A real user starting from zero still hits packaging, install, and integration failures the AI never saw.

That leaves a gap between "tests pass" and "actually works." This isn't a hunch — the repo says so itself.

Stated Verbatim

The repo names the gap in its own README

The problem statement isn't outside framing — the Digital Twin Universe README declares the failure mode and the mission to fix it in plain text.

So the fix is purpose-built for exactly this gap. What is it?

# README.md, lines 3–6
"By default AI generated software is
 verified in the environment and context
 that it was built."

"This frequently leads to agents claiming
 it worked due to context poisoning."

"This leaves a gap between 'tests pass'
 and 'this actually meets all the
 requirements and works in a
 real-world scenario.'"

The Answer

DTU answers one question: "What will reality be like if this were actually deployed?"

The Amplifier Digital Twin Universe is a complete, isolated environment, stood up on demand from a single declarative YAML profile, that simulates the world the code will live in — so consumers can clone, install, run, and experience it like a real user.

The next question is how a profile becomes a running, real-user environment.

How It Runs

Each profile spins up an Incus container that runs the code like a real user would

The engine wraps the Incus container runtime to clone, install, provision, and run the code from zero — driven by 9 lifecycle commands from launch to destroy.

But isolation only works if the code inside still sees the real world.

launch exec status list check-readiness update file-push file-pull destroy

Faithful, Yet Isolated

Real APIs pass through; everything else is mocked behind real hostnames

Real stateless APIs pass through untouched, while external dependencies are mocked as Docker sidecars — with mitmproxy transparently rewriting DNS and URLs (including WebSocket upgrades) so the code inside sees real-world hostnames.

One profile puts the whole mechanism to the test.

Pass through — real stateless APIs (LLM, package registries) go straight to the source
Mock — external dependencies run as Docker sidecars inside the environment
Rewrite — mitmproxy transparently redirects DNS and URLs to the mock

The Proof

amplifier-user-sim catches the install failures a build-only check never sees

The amplifier-user-sim profile simulates a real Amplifier user end to end — catching exactly the packaging and install failures a build-environment-only check would miss.

And this discipline is written into how the repo works.

Rebuild amplifier-core from a freshly built git wheel via pypi_overrides
Serve the anthropic provider module from an external Gitea via url_rewrites
Pass through the real Anthropic API
Install the Amplifier CLI from upstream GitHub
Validate with amplifier --version

The Pattern

DTU turns "verify like a real user" into a repeatable discipline

The repo institutionalizes the practice: its own AGENTS.md and PR template demand verification evidence in the PR body — "not just tests pass."

Close the gap between "tests pass" and "actually works" — on purpose, every time.

# AGENTS.md, L36
"The reviewer expects evidence in the
 PR body, not just 'tests pass.'"

# PULL_REQUEST_TEMPLATE.md, L11
"PR body includes verification evidence,
 not just 'tests pass'"

Sources

Sources & Research Methodology

Primary source: robotdad/amplifier-bundle-digital-twin-universe (fork tracking upstream microsoft/amplifier-bundle-digital-twin-universe), cloned to /tmp/dtu-clone. 62 commits on main; git tag v0.3.1 present.

Feature status: Development Status :: 3 — Alpha (pyproject version 0.3.0; bundle.md declares 0.3.1). Main branch history spans 2026-03-25 to 2026-05-22.

Research performed:

git remote -v — confirmed origin=robotdad, upstream=microsoft
git log --reverse --format='%ad|%an|%h|%s' --date=short — 62 commits, timeline & milestones
cat -n /tmp/dtu-clone/README.md — problem statement (L3–6), DTU definition (L14–18), amplifier-chat (L47–58), mock_services (L160–162)
grep -nE '^def |@main.command' src/…/cli.py — confirmed 9 CLI commands
wc -l src/…/*.py — core library ~4,397 lines (engine.py 1784, incus.py 869)
grep -c incus incus.py (65) & grep -c mitmproxy engine.py (13)
cat profiles/tests/amplifier-user-sim.yaml — pypi_overrides, url_rewrites, passthrough, validation
grep -rin 'tests pass' --include='*.md' — AGENTS.md L36, PULL_REQUEST_TEMPLATE.md L11
git log --format='%an' | sort | uniq -c | sort -rn — contributors

Primary contributors: David Koleczek (45 commits on main), Brian Krabach (12). Authored by the Microsoft MADE:Explorations Team.

Gaps: Only the default unit + fast suite was executed (256 passed, 77 skipped); integration/e2e tiers require live Incus, Docker, and API keys and were not run. PR discussion dates were not separately fetched. The "Jenga tower" phrase is external framing and does not appear in the repo.