From a key to your first decision.

Keys are issued on request. Ask us for one. Your keys are listed and revocable on your API keys page.

Authorization: Bearer pt_live_<16 hex>_<43 base64url>

export BASE=https://your-engine.example
export PT_KEY=pt_live_...

BASE is the API’s address: the Playground in the app shows a request as code with it filled in.

curl -s "$BASE/v1/me" -H "Authorization: Bearer $PT_KEY"

What this key is, and what is left of its day. Any valid key reaches it.

POST /v1/systemone with "model": "helm-default" answers at once, with no model of your own. The key needs the spend scope: sealing an answer is a write.

curl -s -X POST "$BASE/v1/systemone" \
  -H "Authorization: Bearer $PT_KEY" -H 'content-type: application/json' \
  -d '{
  "model": "helm-default",
  "state": "I was charged twice for the same order last week and nobody has answered my emails. Please help.",
  "questions": {
    "billing": {
      "type": "noul",
      "instructions": "Is this about billing?"
    },
    "tone": {
      "type": "choice",
      "instructions": "What is the tone?",
      "criteria": {
        "calm": null,
        "frustrated": null,
        "angry": null
      }
    }
  }
}'

# 200: {"model": "helm-default",
#  "answers": {"billing": {"type": "noul", "noul": …},
#              "tone": {"type": "choice", "choice": "…", "confidence": …, "probabilities": {…}}},
#  "usage": {"input_tokens": …, "output_tokens": …},
#  "helm": {"forecast_ids": {…}, "sealed_at": "…", "chain_hash": "…", …}}

Each answer is sealed, one forecast per question, in your workspace’s general log. helm-default draws on your daily allowance rather than a per-token price.

• In the app, open Models, choose New model, then From your data. Choose or upload a CSV of past decisions: at least 1,000 rows, each with the facts known at the time and what happened.
• Helm proposes the outcome, the question and the role of every column on one screen. Confirm it and press Train: Helm runs its checks on the file, then trains the model.
• The model’s page shows its short name, which is what model takes from here on.

From code, once its checks have run: POST /v1/models/{model}/train. Train the Baseline from the model's checks, as the model page's Train does.

POST /v1/models/{model}/forecasts: Forecast 1–100 records with one of your models, each sealed with its reasons before it is returned.

export MODEL=your-model   # the model's short name

# What a record carries: its fields, and how each is read.
curl -s "$BASE/v1/models/$MODEL/schema" -H "Authorization: Bearer $PT_KEY"

# Forecast one record. "fields" holds the model's own fields, as the schema lists them.
curl -s -X POST "$BASE/v1/models/$MODEL/forecasts" \
  -H "Authorization: Bearer $PT_KEY" -H 'content-type: application/json' \
  -d '{
  "records": [
    {
      "ref": "r1",
      "fields": {
        "<field>": "<value>"
      }
    }
  ]
}'

# 200: {"model": "…", "version": "…", "arm": "…", "answers": [
#   {"ref": "r1", "forecast_id": "…", "probs": {…}, "reasons": […],
#    "sealed_at": "…", "chain_hash": "sha256:…", "input_tokens": …}]}
# One bad field refuses the whole call, names every problem in "rejected", and saves nothing.

The same model answers on POST /v1/systemone too, one record at a time: send the record as state and ask its own question.

POST /v1/models/{model}/outcomes: Record what happened, by outcome key, so forecasts can be scored.

curl -s -X POST "$BASE/v1/models/$MODEL/outcomes" \
  -H "Authorization: Bearer $PT_KEY" -H 'content-type: application/json' \
  -d '{
  "outcomes": [
    {
      "outcome_key": "<the outcome key of the record>",
      "outcome": "<an answer id>"
    }
  ]
}'

# "outcome" is one of the model's answer ids; the schema lists them.
# A changed outcome replaces the earlier one; the history is kept.

A forecast counts toward the model’s track record only when its outcome is recorded after the forecast was sealed.

GET /v1/models/{model}/track-record: A version's live record over your resolved forecasts, beside its card.