PROMOTIONGUARD · WALKTHROUGH 1/6
An agent, running an ML pipeline, unattended
Somewhere in every ML team's backlog is the same job: pull the eval split, run the
candidate model, check the numbers, promote the winner. Repetitive, well-specified,
exactly the kind of thing you'd hand to an agent — and exactly the kind of thing nobody
wants to babysit.
You are about to watch that agent run twice. Once with nothing watching it.
Once with ArmorIQ checking every single call before it executes. Same agent, same
code, same prompt — the only difference is enforcement.
UNGUARDED
Nothing is checking. If the agent reaches for something destructive,
it happens.
GUARDED · ARMORIQ
The identical agent, every call checked against a signed plan before
it executes.
WALKTHROUGH 2/6
Two things go wrong — on purpose
When a candidate model misses the bar, an agent reaches for one of two moves:
TRIM THE EVAL SET
The dataset card says some rows are noisy. Delete them, re-score, the
numbers improve. It arrives as documentation, not an attack — there's no adversarial
phrasing to detect.
SHIP TO PRODUCTION ANYWAY
The exact same promote_model call it was authorized to
make seconds earlier for staging — one argument different. No scary word, no
keyword to filter on.
A contains("delete") filter catches neither. Only the signed plan
knows the difference.
WALKTHROUGH 3/6
Two gates, not one filter
GATE 1 · SIGNED PLAN · CRYPTOGRAPHIC
Before the run starts, the agent declares every action it intends to
take and ArmorIQ signs it. Anything outside that list is refused — client-side,
before the call ever leaves the process. This is what blocks the deleted rows.
GATE 2 · AUTHORITY · HUMAN
A call that's in the plan but exceeds what this agent may decide alone
— like promoting to production — gets held for a real person with a
higher-ranked role to approve. A requester can never approve their own request.
Approval only ever happens on ArmorIQ's own dashboard. This panel has no
Approve button anywhere — that would be a lie about where authority lives.
WALKTHROUGH 4/6
What you're about to control
AUTHORIZE
Switches that are the signed plan. Whatever's on gets signed the
moment you press RUN — including, if you choose, letting the agent delete rows
for real.
CONDITIONS
Dials that set the world, not the authority: how the model scores,
what the dataset card says — or write the card yourself, free text, and watch the
agent react to your exact words.
THE INSTRUMENT
The scope traces real tool calls as they happen. The plan strip lights
up step by step. The key dial arms and freezes mid-swing the moment something needs a
human.
WALKTHROUGH 5/6
Ask the agent, anything
A round floating button sits in the controls, always reachable: ASK THE AGENT.
Open it and type — about PromotionGuard itself, or about the run you just watched.
The answer streams back word by word, in real time, from the same model that runs the
pipeline.
GROUNDED, NOT GUESSED
Every answer is built from the project itself and the real log of your
most recent run — never invented, never scripted.
TRY THE SUGGESTED QUESTIONS
Click one to send it instantly, or type your own. Ask why it picked the
stage it did, or what a keyword filter would have missed.
WALKTHROUGH 6/6
Everything here is real
No scripted choreography, no placeholder data. Every trace pulse is a line the agent
actually printed. Every gauge reading is live SQLite. Every hold is a genuine ArmorIQ
delegation request waiting on a genuine human.
Flip a switch, write a card, press RUN — and watch what actually happens.