Glossary

API Platform (Console) Safeguard's fifth product: an OpenAI-compatible inference API, plus the Console, its pages in the Safeguard web app.

API key The credential your code sends to the inference API: sg-api- followed by 64 lowercase hexadecimal characters, created in the Console and shown once. See Authentication.

Available credit Your workspace's credit balance minus what requests in flight have reserved.

Balance The workspace's prepaid credit in US dollars, before reservations.

Base URL The address the OpenAI SDKs are pointed at: https://api.safeguard.sh/gpt/v1.

Catalog The list of models Safeguard knows, with prices, context windows, data policies and status. It is published at safeguard.sh/models, and the gateway reads the same catalog.

Credit ledger The record of every change to your credit: credit granted by Safeguard, card top-ups and usage, each with the balance after it.

Eagle Safeguard's discovery and triage model, safeguard/eagle: 70 billion parameters and a context window of 1 million tokens, listed at $0.00001 per million input tokens and $0.00001 per million output tokens. Not callable through the API yet. See Safeguard models.

Estimated usage Token counts the gateway worked out from the text because the provider reported none. Marked "estimated": true in usage, and est. in the Console's request log.

Fallback A different model the gateway may try when the requested model's provider fails. Only with your permission, and only within the same model family. See Reliability and fallbacks.

Gateway The Safeguard service behind the inference API. It checks the key, the model and your credit, calls the provider, and records the cost.

Gateway settings Workspace settings in the Console's Usage page: the monthly spend limit, prompt storage, retries and fallbacks, and the zero data retention requirement. Only tenant admins and super admins can change them.

Griffin Safeguard's remediation and reasoning model, safeguard/griffin: 750 billion parameters and a context window of 1.1 million tokens, listed at $0.0001 per million input tokens and $0.0001 per million output tokens. It is one model, with no Lite, S, M, L or XL variants. Not callable through the API yet. See Safeguard models.

Inference scope The scope every Console-created key carries, inference, and the only one it carries. It allows chat completions, listing models and GET /key, and nothing else.

Launch checks What a catalog model must pass before it can be called through the API: available in the catalog with prices, a configured provider account, the provider's launch checklist for third parties, and a successful end-to-end test. See Models.

Lion Safeguard's compliance and narrative model, safeguard/lion: 15 billion parameters and a context window of 128K tokens, listed at $0.000001 per million input tokens and $0.000001 per million output tokens. The AI Models page calls it Lino. Not callable through the API yet. See Safeguard models.

Masked key How the Console shows a key without its secret: its start and end, such as sg-api-1a2b3c4d…9f8e.

Markup An amount added on top of a provider's price. The API Platform adds none to third-party model usage.

Model id The provider/model string you pass as model, such as openai/gpt-6-sol.

Monthly spend limit A cap on what a key, or the whole workspace, may spend in a calendar month in UTC. Requests that would pass it are refused with 402.

Prompt storage An opt-in workspace setting that stores the full text of prompts and outputs. Off by default.

Provider The company whose model answers a request. The response's provider field names the route used, such as openai.

Request id The id of one request: gen- followed by 32 hexadecimal characters. It is the response's id, and every chat response, streamed or not, also sends it in the X-Generation-Id header.

Request log The Console's list of requests, with model, key, tokens, cost, latency and status. Metadata only.

Reservation The amount held from your credit while a request runs: the most it could cost. It shows as Reserved in the Console and is released when the request settles, or after about 30 minutes if it never does.

Settlement The step after a request finishes, when the gateway charges its actual cost, releases the reservation and writes the request log entry.

Streaming Receiving the answer as server-sent events while it is generated, with "stream": true. See Streaming.

Top-up Buying credit by card in the Console, where Safeguard has switched card top-up on. The card fee is passed on at cost.

Usage chunk The last event of a stream when you set stream_options.include_usage: no choices, just usage with tokens and cost.

Workbench The Console page for trying models without code. It uses your sign-in instead of a key. See Workbench.

Zero Safeguard's no-weights model, safeguard/zero. It uses no model tokens and is priced at $0 per million tokens. See Safeguard models.

Zero data retention (ZDR) Routing only to models whose provider offers zero data retention, per request or for the whole workspace (the setting Only use models whose provider offers zero data retention). A model without it is never used, not even as a fallback. See Zero data retention.