QuantData · Client demo
Twenty-three scenes carry one request from Level 0 (watch everything) to Level 3 (a permission slip per action) — the same system under progressively more governance, not four unrelated products.
“Hedge all our accounts back to USD.”
Watch everything, gate nothing — and the permanent record, the ledger, fills anyway.
Check what a request means before it runs — semantic governance.
Keep each registered agent inside its written scope, and check pairs of agents — not just one at a time.
Almost no standing permissions: each action gets a short-lived permission slip — an operational grant — and proof it ran unchanged.
What the demo must prove
Every scene is traceable to at least one claim. Each card links to the scenes that prove it.
Every governed action is a transaction; every quantitative control is a certified expression over the ledger plus the incoming request.
Degradation is honest, never silent.
The Arc
Level 0 → 3 on one story. The first two scenes cost the prospect nothing to say yes to.
Audience routing: UC-16 and UC-13 are the CISO's meeting · UC-15 and UC-22 the controller's · UC-08 and UC-20 the platform lead's · UC-19 closes an audit buyer.
Group A
Nothing is gated here — these scenes are conversations, not dashboards; each opens on a call and its answer while the record quietly fills beside it.
“We changed nothing about how your agent behaves. It is not slower, not blocked, not asking permission. And you now know exactly what it did, in business-process terms.”
“Coverage does not depend on the agent's cooperation. Registration is the entry ticket to Level 2, not to being seen.”
“The agent was flagged and never interrupted — and that flag is exactly the evidence you would want before turning Level 2 on for this scope.”
“QuantData records its environment baseline; it does not infer it. That is the only reason these two cases separate at all.”
Group B
“The agent guessed. It guessed plausibly. It guessed wrong. And it had no way to know.”
“Structural ambiguity stays explicit. The alternative is a system that quietly edits the question behind the caller's back.”
“Pushing route-version diagnosis into the AI would undermine the whole point of centralizing meaning.”
“Every other tool in your stack returns one word for this request. Which word would have been correct?”
“The computation was allowed. The output was not.”
“The enterprise publishes a narrow certified action instead of granting a broad technical surface that looks like trade FX.”
Group C
“It ships with a manifest. We tell you what it can do here, today, and exactly what is missing.”
“The meaning was governable. The actor was out of its lane. And the denial was constructive — it told the agent how to finish.”
“Each agent is individually compliant. The pair is not. No IAM or IGA product in your stack evaluates the pair.”
“Every rung is proportionate, recorded, and reversible — up to the point where the authority plane takes over, and then there is a change set with someone's name on it.”
Group D
“Your agent may submit five hundred thousand dollars of adjustments per day. Your controller chose that number. Its utilisation is a report, not a mystery — and two requests cannot both slip under it.”
“Approved and executed are two different requests in every system you own. Here they are provably the same request.”
“The same semilattice accounting, segregation-of-duties, certification, and evidence machinery applies to grants unchanged. There is nothing new to trust.”
“Nothing to revoke. Nobody else's work stops.”
Group E
Optional scenes — each complete on its own; none is required for the arc.
“A silent fallback is a hidden failure. A recorded one is an operable event. This is the difference between a control model and control theatre.”
“The runtime never silently changes the governing model. It observes friction, groups it, explains it, and turns it into explicit improvement work.”
“A limit decision is audit evidence, not a transient check.”
“The object that caps an external agent's payments is the object that enforces your internal control.”
“Which control level applies where is assigned to each scope, not written into code — and strictest-wins resolves the overlaps live.”
Honesty about liveness: this backend hosts no language model. The agent persona is scripted and labeled as such; the QuantData side is genuinely live wherever the wiring badge says LIVE, and recorded objects are always labeled REPLAY — with a note saying why.