Service

Decision model API

Typed questions about a text, answered with probabilities.

This is a small classifier, not a chat model. You give it a text and up to 8 questions, each one of three kinds: a predicate (how likely is this statement to hold), a choice (which option fits, with a probability for each) or a score (where on an ordered scale). It answers in well under a second.

From code

request

curl -s https://slop.her.st/api/decide -H 'Content-Type: application/json' -d '{
  "input": "Subject: Your invoice is 30 days overdue. Please pay within 7 days.",
  "questions": [
    {"name": "spam", "type": "predicate",
     "instructions": "This message is unsolicited bulk or fraudulent mail."},
    {"name": "topic", "type": "choice", "instructions": "What is this message about?",
     "choices": {"billing": "payments and invoices", "personal": "private correspondence",
                 "marketing": "promotions"}},
    {"name": "urgency", "type": "score", "instructions": "How urgent is this message?",
     "levels": ["low: no action needed", "medium: within weeks", "high: within days",
                "critical: immediately"]}
  ]
}' | jq .

response

{
  "answers": {
    "spam": {"type": "predicate", "probability": 0.32},
    "topic": {
      "type": "choice", "choice": "billing", "confidence": 0.68,
      "probabilities": {"billing": 0.78, "personal": 0.15, "marketing": 0.06}
    },
    "urgency": {
      "type": "score", "score": 1.81, "confidence": 0.5,
      "legend": {"0": "low: no action needed", "1": "medium: within weeks",
                 "2": "high: within days", "3": "critical: immediately"},
      "probabilities": {"0": 0.01, "1": 0.36, "2": 0.5, "3": 0.13}
    }
  },
  "usage": {"input_tokens": 253}
}

Tip: phrase predicates as statements ("The author is angry."), not questions. This small model answers statements more reliably, and its probabilities are cautious: a clearly true statement may come back around 0.6, a clearly false one near 0. For a compact view, pipe through jq:

one line per answer

curl -s https://slop.her.st/api/decide -H 'Content-Type: application/json' -d @request.json \
  | jq -r '.answers | to_entries[] | "\(.key)\t\(.value.probability // .value.score // .value.choice
      | if type == "number" then (. * 100 | round / 100) else . end)"'
# spam    0.32
# topic   billing
# urgency 1.81

Reference

Field	Type	Notes
`input`	string	The text to judge, up to 16,000 characters.
`questions`	list	1 to 8 questions.
`name`	string	Unique per request: lowercase letters, digits, underscores; at most 32.
`type`	string	`predicate`, `choice` or `score`.
`instructions`	string	The statement or question, up to 500 characters.
`choices`	object	Choice only: option to description, 2 to 10 pairs.
`levels`	list	Score only: 2 to 10 strings, lowest first.

Question type	Answer fields
predicate	`probability`, 0 to 1: how likely the statement holds.
choice	`choice` (the winner), `probabilities` (every option), `confidence`.
score	`score` (0 up to the top level index), `legend`, `probabilities` per level, `confidence`.