Skip to content

Guardrails

VERIFIED PromptDesk docs/guardrails.md is the baseline for Argus and Quizzeira. Align contracts; do not weaken without an explicit task.

Enforcement model

RuleDetail
Registry SoTPrompts in versioned files (PromptDesk: apps/chat-worker/prompts/registry.yml) — not inline in code
Pre-LLM screenDeterministic patterns before any model call
Trust boundaryUntrusted history/guidelines only inside fenced user-message context — never into system instructions
Fail closedUnknown verdicts / contract violations are errors
LoggingIds, statuses, models — never message/guideline bodies
AgencyText-only LLM; no tools / no writes from the model

Provider policy

  • Gemini-first; OpenAI optional failover
  • Honor LLM_PROVIDER_ORDER (unknown names ignored; keyless providers skipped)
  • Cross-provider failover only after all attempts on current provider fail

OWASP LLM Top-10 (2026) highlights

RiskEnforcement note
LLM01 Prompt InjectionFenced context; guideline quarantine + validator
LLM02 Sensitive DisclosureNo secrets in prompts; PARTIAL PromptDesk Redis chat:events carries reply text
LLM03 Excessive AgencyText-only; no tools
LLM06 Unbounded ConsumptionHTTP caps + worker minute/daily budgets
LLM09 Vector weaknessesN/A for PromptDesk (no vector store); see product pages for Argus/Quizzeira

Product mapping

ProductGuardrails home
PromptDeskdocs/guardrails.md (reference)
Argusdocs/ai-engineering.md, apps/api/docs/guardrails.md
Quizzeiradocs/guardrails.md