Intent · CHG-2026-014907

change_id: CHG-2026-014907
risk_class: R2
autonomy_tier: A2
controls: ['TPR-05', 'TRC-01', 'CHG-04']
data_classification: public
originator: olaf@freundcloud.com
agent_identity: svc-agent-platform
model_route: gateway/tier-balanced
supersedes: None

Problem

Everything this repository argues for is verified by running it, and the running currently depends on whatever happens to be installed on the presenter’s laptop. The README says python -m venv .venv and hopes. There is no pinned toolchain, no single command that performs the demonstration, and no way to find out that an act is broken except by reaching it in front of the audience.

That is an odd gap in a repository whose thesis is that a control you cannot reproduce is a control you cannot evidence. A demo whose result depends on the machine is the same category of claim as a control whose result depends on the reviewer being awake.

There is a second, sharper problem. The demo has an order. The gates mean little before you have seen the loop they guard; the Substitution Test means nothing before you have seen what would otherwise be lost. Reconstructing that order from memory each time is how the strongest act — make negative, the gates refusing — gets skipped when the meeting runs late.

Who is affected

• Whoever presents this, who currently rehearses by remembering.
• Anyone cloning it, whose first impression is a dependency error rather than an argument.
• The maintainer, for whom “does the demo still work” is currently answerable only by performing the whole demo.

Success criteria

1. One command runs the entire demonstration, in order, narrated.
2. It exits non-zero if any act fails, so a broken demo fails in rehearsal rather than in front of the audience.
3. The toolchain is pinned and declared, so the demo behaves identically on a machine that has never seen this repository.
4. Nothing in the demo is staged, mocked, or replayed from a previous run. Every number on screen is computed while the audience watches.
5. The existing Makefile remains the single definition of the loop and the gates. CI and AGENTS.md both point at it, and a second definition would drift.

Out of scope

• Recording or asciinema capture. Separate concern, separate change.
• Any change to what the gates check, what the evals assert, or what the workflows do. This change makes the existing behaviour reproducible; it does not add behaviour.
• Packaging the service for distribution. The payload is deliberately a demo service.

Constraints

• No new runtime dependency for the repository itself: someone without Nix must still be able to python -m venv .venv && make.
• The demo must run offline. The act that talks to GitHub is opt-in and degrades to a printed note when gh is not authenticated.