14:02:31ingeststep.startclassify ticket TCK-1

Disclosure

Trust

Honest by default: the actions this agent will never automate, the models it runs, and what it costs. Every boundary here is enforced by a named policy you can read on the policies page.

Will never automate

Delete or permanently destroy an account or its data06-delete-account-never · ASI10 Rogue Agents
Send a customer-facing message without human approval05-customer-facing-requires-approval · ASI03 Delegated Trust
Read or act on another tenant's data07-tenant-isolation · ASI06 Inter-Agent / Cross-Boundary
Pass unredacted customer PII into the model context03-hubspot-pii-redacted · ASI04 Data Exfiltration

Models and cost

Model
openai/gpt-4o-mini (via Bifrost)
Temperature
0 (deterministic planning)
List price
$0.15 / 1M prompt · $0.60 / 1M completion
Per-run cost ceiling
$0.50 (circuit breaker)
Deployed-demo spend, 24h
$0.00 (recorded runs, no live LLM)

Evals

Custom suite (≥ 90% pass)
21 / 21
InjecAgent subset (≥ 80% pass)
200 / 200
OWASP-ASI assertions (10 / 10 pass)
10 / 10

Run offline and deterministic via pnpm eval, gated in CI before the microsite build. Per-ID coverage is on the evals page.