The acceptance test
Every ScaleToZero engagement ends with the same controlled failure test: we detonate a spend-burst drill inside the client's stack and the guardrail must terminate the run before the cap burns. If the test fails, the client owes nothing. This page shows the drill that runs on the €5,000 / €10,000 guardrail offer.
Honest framing. The drill below is a demonstration harness modeling the paid acceptance test — a Python simulation of a recursive agent-spend loop, not live production client traffic. The parameters mirror real incident economics; the run it kills is synthetic. The client acceptance test runs the same kill logic inside their stack, against their real spend, witnessed live on the scoping call.
Modeled on real $78K-class incidents: a runaway prompt causes every agent child to spawn three more children, and each child bills $133 per API batch — the self-reported figure from the teardown. Without enforcement this compounds exponentially: the money burns at agent speed while dashboards show you yesterday's totals.
| Drill parameter | Value |
|---|---|
| Fan-out per tick | each child spawns 3 more |
| Cost per child per tick | $133 |
| Run spend cap | $500 |
| Spawn budget | 50 children |
| Kill action on breach | terminate run, write audit log |
The guardrail meters spend and spawns every tick. Tick 1 burns $133 — under the cap, the run continues. Tick 2 adds three children at $133 each: cumulative spend crosses $500, and the run is terminated immediately. No Slack alert for a human to read tomorrow. Terminated.
# drill: recursive-loop-spend-burst # cap: $500 · spawn budget: 50 { "kill_fired": true, "kill_reason": "spend_cap_breach", "ticks_to_breach": 2, "spend_at_kill_usd": 665.0, "spawns_at_kill": 3, "wall_seconds": 0.001, "verdict": "PASS — run terminated on breach" } # same drill WITHOUT the guardrail, extrapolated 60 ticks: { "projected_unchecked_spend_60_ticks_usd": 2.8e+30 } Guardrail verdict: PASS. The runaway died on tick 2. The unchecked loop would never have stopped on its own.
Every tick and every kill decision is written to an append-only log outside the agent's reach. After the run, anyone can reconstruct exactly what happened: when each child spawned, when spend crossed the cap, when the kill fired. This is criterion 3 of the acceptance test — if the log can't be reconstructed, the test fails and the client owes nothing.
The same logic runs in your stack. Spawn budgets, per-run spend caps with auto-kill, real-time metering at your layer, immutable audit trail, alerting to your Slack or PagerDuty — installed in 48 hours for a fixed €5,000. Payment happens only after you watch this test pass in your own environment.
The acceptance test is the deliverable. You sign the test, not the invoice.
See the guardrail offer