AtlasBurn
DocsPlatformIntegrationsOperationsReference

Advanced Concepts

Summary

This page covers advanced topics, planned features, and architectural decisions that affect how AtlasBurn operates at scale.

Monte Carlo Burn Forecasting

AtlasBurn uses stochastic simulation (Monte Carlo) to project future AI spend. The model takes historical usage patterns and generates probabilistic forecasts at three confidence levels:

  • P50 — median expected spend. Use for capacity planning.
  • P95 — captures 95% of scenarios. Use for budgeting.
  • VaR95 / CVaR95 — Value-at-Risk and Conditional VaR at 95%. Use for alerting and guardrail thresholds.

The engine reports p5 / p50 / p95 percentiles plus VaR95 and CVaR95. There is no P99 output — earlier docs cited it in error.

System maturity: Stable

Monte Carlo forecasting is production-ready.

Guardrails and Auto Kill Protocol

The legacy SDK-level Auto Kill (deterministic daily/hourly financial caps) has been superseded by Autonomous Runtime Guardrails — a server-side, 5-layer defense engine with Cloudflare KV edge enforcement. The simple flow is still:

  1. The breach is recorded in the Breaches Ledger
  2. The organization enters a Suspended state
  3. Further AI API calls are rejected at the edge until TTL expiry or manual resume in the dashboard

Organization-wide only

Feature-specific hard stops are not yet available. Guardrails currently apply at the organization level only. (Planned)

Retry Cascade Detection

A retry cascade occurs when a failed AI call triggers automatic retries, which also fail. This can multiply costs 10–50× in minutes. AtlasBurn detects cascades by tracking retryCount on each event and flagging sessions with abnormal retry rates.

FALLBACK_ECONOMICS

When the cost engine encounters an unrecognized model, it applies rates from FALLBACK_ECONOMICS instead of PROVIDER_REGISTRY. These are intentionally conservative (high-margin safety rates) to prevent underreporting.

If you see "Fallback" tags in your Forensic Ledger, it means the model wasn't matched. Update your PROVIDER_REGISTRY or contact support.

Planned Features

FeatureStatusNotes
Multi-Org MembershipPlannedNot yet finalized in the UI
Automated Reconciliation APIPlannedCurrently manual via dashboard
Feature-Specific Hard StopsPlannedGuardrails are currently org-wide only
Vector Database CostsPlannedNot yet in normalization engine

What This Means in Practice

AtlasBurn's advanced features — Monte Carlo forecasting, Auto Kill, and retry detection — are designed for teams running AI at scale where silent cost explosions are an existential risk. These features work together: forecasting predicts the burn, guardrails enforce the budget, and retry detection catches the cascades.

Next Steps