StretchGPT is secure by default. Every potentially sensitive action passes through an execution gate, and until an operator explicitly activates the runtime, the platform runs in a safe, non-live mode. This article explains the safety model so operators know exactly what's gated, why, and how activation works.
Why a safety envelope exists
An AI control plane can do powerful things — call external models, run GPU workloads, write training data, mutate models. StretchGPT treats all of these as privileged and refuses to do any of them implicitly. The safety envelope is the set of guarantees that hold at all times; the execution gate is the check that enforces them on each request.
What the gate hard-blocks
The execution gate blocks these live actions unless the runtime is explicitly activated for them:
live_provider_chat— actually calling a model provider for a live answer.worker_node_execution— running work on a GPU node.search_bridge_execution— executing the search bridge.embedding_generation— generating embeddings.training_job— running training.dataset_mutation— changing datasets.model_write— writing model files.auto_promotion— promoting a model automatically.crawler_trigger— triggering a crawl.
When the gate blocks a request it still records an audit decision — the mode (dry_run or blocked), the reasons, and the source — so there's a trail of what was attempted and why it was denied. Importantly, raw prompts from the public site are not stored; the gate keeps only a hash and metadata.
Dry-run by default
Until activation, requests resolve to dry-run or blocked. In dry-run mode, StretchGPT exercises the full path — validates the request, checks readiness, prepares the audit event — but does not perform the live action. This lets you verify that everything is wired correctly before you flip the switch, with zero risk of an unintended model call, training run, or write.
Standing guarantees in the safety envelope
Regardless of activation, these always hold:
- Keys are encrypted at rest and never returned. Provider secrets never reach the browser and are never exposed by the API.
- No direct browser-to-provider calls. The browser never calls a model provider directly; requests go through the governed server path.
- Retrieval is evidence-only from the index. StretchGPT never fetches or scrapes the live web; it reads only what StretchSearch has indexed.
- Raw public prompts aren't persisted. Only a hash and metadata are kept for public chat.
- Sensitive writes need step-up authorization. Connecting providers, approving training data, and promoting models each require an admin action token.
What activation unlocks
Activation is the deliberate, operator-driven step that turns dry-run into live. To get live chat, the runtime must be activated and the prerequisites met: a model provider connected, a model configured, and live execution enabled. When those line up, the gate allows live_provider_chat and users get real answers. The same explicit-activation principle governs training and worker execution — they run only when their runtime is turned on. See Activating Live Chat for the chat-specific steps.
A realistic example
You've just deployed StretchGPT. A test message comes back blocked, listing reasons like provider not connected and chat execution disabled. That's the safety envelope doing its job. You connect a provider (step-up authorized), configure its default model, and enable live execution. You send the test message again: this time the gate reports allowed / live_provider and a real answer comes back. Nothing about the platform changed except the explicit activation you performed — and every step is in the audit trail.
Tips
- Read the blocked_reasons. They tell you exactly which prerequisite is missing.
- Use dry-run to validate wiring before activating, especially for training and worker lanes.
- Treat step-up prompts as intended friction. They exist so a sensitive write is always a deliberate act.
- Remember retrieval is index-only. If you need broader evidence, index it in StretchSearch — StretchGPT won't reach out to the web.
FAQ
Why is my chat blocked even though I typed a valid message? The live runtime isn't activated, or a provider/model prerequisite is missing. The response lists the exact reasons.
Does StretchGPT store what users type? For public chat, no — only a hash and metadata. Authenticated conversations are stored so you have history, always scoped to the owning user and organization.
Can anything be promoted or trained automatically? No. Auto-promotion and training are hard-blocked actions that require explicit activation and, for promotion, an eval and human sign-off.
Was this helpful?
Help us improve this article
Use these controls to share whether this answer solved the issue. Feedback helps prioritize updates to StretchSuite Support.