Know what is answering.

BrokenGPT is the product and API layer. broken-one is the stable public alias, not a claim that BrokenGPT trained the underlying model. The active third-party upstream can vary by request route and is disclosed below.

Current configuration

Public alias: broken-one
Display name: broken-one
Primary upstream: deepseek-v4-flash
Less-restrictive route upstream: protoLabsAI/Qwen3.6-35B-A3B-uncensored-heretic-FP8
Advertised context cap: 1,048,576 tokens
Runtime: operator-configured inference

The gateway selects a route using the published service boundary. If only one upstream is configured, requests fall back to that model. The assistant is instructed to identify both the public alias and the active upstream when asked.

What “highly steerable” means

The product aims to reduce unnecessary refusals on lawful requests. It does not promise zero moderation or guaranteed compliance. Platform-level hard boundaries remain in force even when a model would otherwise answer.

Accuracy and limitations

Model output can be false, biased, incomplete, outdated, or insecure. Do not rely on it as the sole basis for medical, legal, financial, safety-critical, or high-impact decisions. Verify code and factual claims independently.

Data flow

Web-chat messages are stored in Postgres to provide history. Requests sent directly to /v1/chat/completions are not stored as conversation content by this application; usage metadata such as token counts, model, latency, key ID, and request status is recorded. The configured inference provider may have its own retention terms, which must be disclosed here before launch.

License and provenance

A model identifier alone does not prove that a checkpoint is open source or open weight. Operators must verify commercial rights for each exact model, derivative, and quantization; preserve required notices; pin revisions; and publish any license restrictions that affect customers.