Skip to main content

​
In the website

Open a completed Forecast or game snapshot. Three Scout cards cover Forecast, History and Evidence. Type your own question or explore quantile, range, data-quality and validation prompts. CSV previews also support questions about the selected observations before inference. Responses show numerical facts labeled observed, forecast, or metadata. Source details include the input cutoff, native frequency and completed model participants. Saved conversations appear in Profile → Requests, where Load reopens the conversation. Sign-in is required to ask and save. Chart annotations & notes can mark the observed and P50 high/low. Add a note at an actual observed or forecast point, then click Save notes. Notes are private to your account. Redrawing a chart or typing does not trigger a Firestore write.

​
Structured HTTP Q&A

Use an active pay-as-you-go or Pro (including an active 14-day trial), enterprise or administrator API credential with forecasts:read. The active 14-day trial includes programmatic API and MCP access.
Generate fresh UUIDv4 IDs for a new conversation and each new turn. On a retry, reuse both IDs and the same question. A completed retry returns the saved response without another model call. Reusing a turn ID for a different question is rejected. Context kinds: Clients cannot submit forecast values as authoritative context. Forecasts are loaded and permission-checked server-side. A game refreshed since the conversation opened requires a new conversation; saved game snapshots retain their original context. The response contains conversation_id, turn_id, the question, and a response object with:
  • model: "jev-1.13.0" and response_mode: "typed_routing_computed_facts";
  • category, topic, headline, explanatory answer and numerical facts;
  • context_hash, source references, warnings and next suggested questions.
GET /api/v1/jev/conversations/{id} reopens the owner’s conversation and rechecks current resource access. Each conversation retains up to 30 turns. Questions are limited to 1,500 characters. Usage controls include six new turns per user/minute, bounded daily user budgets and a global daily budget; repeated completed turns do not consume another model call.

​
Notes API

POST the same context to /api/v1/jev/annotations/read. To save, POST context and notes to /api/v1/jev/annotations/save:
Use history for an observed input point or the exact decimal quantile key for a forecast point. The timestamp must exist in that series. Values are derived server-side. Up to 20 notes of 240 characters are supported. Saving replaces that user’s notes for the exact context; it does not modify the forecast.

​
Interpretation and privacy

Scout selects a typed topic, quantile and aggregation. Quantura computes the figures from saved data: up to 500 input observations and 512 forecast steps. Ambiguous questions receive clarification; provider failure returns 503 JEV_UNAVAILABLE, without a substitute AI answer. These responses are structured calculations and authored explanations, not unrestricted generated market commentary. No news, executions, strategy win rates or returns are inferred from quantiles. Prediction-market price/probability forecasts are distinct from the probability that a game outcome wins. Historical economic cutoffs do not establish vintage availability for revised data. Historical metrics appear only when an actual saved validation report exists. The Scout provider receives the question, recent question topics and compact source/availability metadata, not raw CSV files or complete history arrays. The conversation and calculated response persist privately in Firestore. API keys and passwords are rejected from questions; do not submit sensitive information. The current Mintlify MCP allowlist exposes read-only discovery and capabilities. These HTTP chat/note mutations are not automatically available as MCP tools.