shrug.ai API
shrug.ai is stupid, fast. It answers decision and judgment requests with a principled random draw instead of a model, giving you a chance-level baseline to benchmark your real endpoints against. If your model can’t beat a shrug, the model isn’t the thing adding value.
It speaks two request/response shapes so it can slot in wherever you already call a judgment API:
POST /v1/decisions— the OpenAI Decisions shape (input+ an array ofquestions).POST /v1/systemone— the TypeSafe System One shape (state+ a keyedquestionsobject).
Both endpoints return calibrated probabilities for every question (choice, predicate, and score) so the output is directly comparable to a real model’s.
Quickstart
Authenticate with a Bearer API key (create one in your dashboard), then POST a request to either endpoint:
curl https://api.shrug.ai/v1/decisions \
-H "Authorization: Bearer shrug_sk_your_key_here" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-6-luna",
"input": "the sky at noon",
"questions": [
{ "type": "predicate", "name": "is_daytime", "instructions": "is it day?" }
]
}'
Where to go next
- Authentication — Bearer keys, and the 401 vs 402 divergence (bad key vs over quota).
- POST /v1/decisions — the OpenAI Decisions shape, end to end.
- POST /v1/systemone — the TypeSafe System One shape.
- Density extension — bias the draw with
shrug.distributions. - Errors & limits — the full status-code table.