Glossary
Summary
Canonical definitions for AtlasBurn terminology. These terms are used consistently throughout the documentation, dashboard, and SDK.
Observed Spend
Real-time capital outflow captured by the SDK. Calculated after the API call completes using actual token counts from the response.
Provider Cost
Cost calculated using exact rates from the PROVIDER_REGISTRY. Used when the model is recognized.
Estimated Cost
Fallback pricing used when model identity is ambiguous or unrecognized. Intentionally overestimates to protect against underreporting.
Burn Rate
The velocity of capital consumption, standardized to USD per minute ($/min).
p5 / p50 / p95
Stochastic cost percentiles emitted by the forecasting engine. p50 is the median, p95 captures most outliers. No p99 is published — use VaR95 / CVaR95 for tail risk.
VaR95 / CVaR95
Value-at-Risk and Conditional Value-at-Risk at the 95% level. VaR95 is the loss not exceeded 95% of the time; CVaR95 is the expected loss conditional on exceeding VaR95. Used as the canonical tail-risk metrics for guardrail thresholds.
Fallback Premium
The high-margin safety rate applied to unrecognized telemetry via FALLBACK_ECONOMICS. Intentionally conservative.
Auto-Detect
Zero-code interception of globalThis.fetch calls to known AI provider endpoints. No wrappers or middleware required.
Edge Proxy
AtlasBurn's language-agnostic Cloudflare Worker at proxy.atlasburn.com. The universal hard-enforcement path for any provider not auto-captured by the SDK.
Guardrails
Deterministic budget limits (daily/hourly) that trigger the Auto Kill protocol when breached. Currently organization-wide.
Auto Kill
The deterministic protocol that suspends AI API calls when a budget guardrail is breached. Requires manual resume in the dashboard.
Forensic Ledger
An append-only, server-authored audit trail of every AI API call. Each entry includes model, tokens, cost, latency, and attribution metadata. Clients can write only via the ingest API; entries are never updated client-side.
Reconciliation
The process of aligning forensic telemetry with actual provider invoices. Currently manual via dashboard; automated API is planned.
Normalization Engine
The component that converts provider-specific response fields into a unified deterministic schema.
Retry Cascade
A failure pattern where a failed API call triggers automatic retries, which also fail, multiplying costs 10–50× in minutes.
Fail-Silent
The SDK behavior where all telemetry errors are caught internally and never propagate to the host application.
Feature ID
A metadata tag (featureId) used to attribute AI costs to specific product features. Set on initAtlasBurnAuto or via request context.
Estimated Tokens (est / ~est)
When a response does not emit a usage frame, AtlasBurn estimates tokens from request size. The ledger flags such events with `est` or `~est` so they can be excluded from precise financial reporting.
HMAC-SHA-256
The hash algorithm used for API key verification. Provides deterministic, O(1) lookup without storing raw keys.