Decisions

Chat asks these questions in the verdicts format: a state, typed questions, and a probability for every answer. Any service that accepts that request can answer, so the gateway is one choice among several. When decisions are off or an endpoint fails, the feature that asked goes on without an answer.

Choose where to ask

1. 1

   Open Settings › Administration › System

   The Decisions card shows where questions go now.

2. 2

   Pick an endpoint

   This gateway, Another endpoint with its URL and key, or Off.

3. 3

   Check, then save

   Check asks a sample question with the values in the form, saved or not, and says which model answered and how long it took. Save applies them on every replica.

Another endpoint

Enter the address that takes POST requests in the verdicts format. An address with no path, or one ending in /v1, gets /v1/verdicts added, so another obleth gateway's base URL works as it is. Any other path is used as given.

• Key. Sent as Authorization: Bearer <key>. Chat encrypts it with OBLETH_CHAT_SECRET_KEY_FILE, so saving one needs that file. Leave the field blank to keep the saved key. If the secret key changes, Chat can't read the saved key and says so rather than asking without it.
• Redirects aren't followed, so the key goes only to the address you entered.
• People's own gateway tokens never go to another endpoint.

The model

Leave Model blank to let the endpoint choose. On obleth that's the model under Settings › Routing › Verdicts, or auto when none is set. The model must return token probabilities (logprobs): self-hosted servers such as vLLM, SGLang and llama.cpp do, and many hosted APIs don't.

What uses decisions

{
  "state": {
    "goal": "Lower validation loss",
    "idea": { "title": "Higher learning rate", "description": "Raise lr from 3e-4 to 1e-3" },
    "change_from_parent": "-lr = 3e-4\n+lr = 1e-3\n",
    "output": "config: lr=0.0003\n…\nloss=2.41\n",
    "result": "loss=2.41"
  },
  "questions": {
    "tested": { "type": "boolean", "instructions": "…", "criteria": { "true": "…", "false": "…" } }
  },
  "on_no_answer": "low_confidence"
}