API · Public archive and measurement
Read the public evidence behind Crene decision records as JSON: stored event forecasts, reported outcomes, model-family scores, numeric factors and scenario structures.
These public records are separate from authenticated private workspace research, assessments and team decisions. Public automatic updates have been retired; inspect each observation’s timestamp.
01 · Quick start
The GET endpoints in this reference require no API key and do not request new model forecasts. Read an inventory, follow the identifiers it returns, then inspect the dated record.
curl "https://api-get.crene.com/api/events/accuracy/"curl "https://api-get.crene.com/api/events/lang/en/?type=binary&limit=20&offset=0"curl "https://api-get.crene.com/api/clusters/?status=archived"curl "https://api-get.crene.com/api/factors/"curl "https://api-get.crene.com/api/scenarios/"02 · Endpoint reference
Examples show selected fields with illustrative values, not current forecasts or a complete schema. Preserve null values and the original payload when building an evaluation dataset.
4 endpoints · Events and scores
03 · Data conventions
A number is useful with its question, unit, model and evidence time. Collection shapes differ across these endpoints; validate each response before importing it.
Keep generated_at or polled_at with each observation. A summary computation, retrieval time or is_active flag does not date the underlying forecast.
Event estimates use 0–100. Summary accuracy can use 0–1. Numeric percentiles use the factor’s unit. Preserve null and missing values; do not convert unavailable estimates to zero.
Check HTTP status before parsing a result. Collection shapes differ; there is no single pagination envelope. Save the raw response and validate required fields before importing it.
POST /probability is an evaluation-service contract, separate from the public read base URL above. Agree the host, provider data handling, credentials and budget before sending new questions. Requests use the evaluation service’s X-API-Key credential.
Request: question (10–500 characters), optional horizon (up to 100), context (up to 2,000) and resolution_criteria (up to 1,000). Define the source, threshold and outcome horizon explicitly.
Response: request_id, generated_at, probability and disagreement on a 0–1 scale, a per-model breakdown and n_models for successful estimates. Model coverage can be partial. A model entry with error is unavailable, even if its probability field is zero. Confidence describes the range of successful model outputs; it can be high with only one successful model, so read it alongside n_models. This is not calibrated confidence. The response is marked calibrated: false. The service’s generated_at is UTC but currently omits an explicit timezone offset.
Errors: missing or invalid credentials return 401; a key’s configured hourly limit returns 429; unavailable authentication returns 503; all model calls failing returns 502. Validation can return 422. Requests can incur provider charges. A timeout does not establish that work was canceled; reconcile the outcome before repeating a paid evaluation.
04 · Public corpus
This public corpus population is separate from personal forecasts and model cohorts in a private workspace.
Unresolved is a lifecycle state, not a freshness guarantee.
The reported corpus population. Inspect dated definitions, source fields and per-model sample coverage before comparison.
Broad corpus mean on a 0–1 scale; lower is better. Historical scores do not establish a model ranking or investment return.
The summary timestamp dates this computation, not a new forecast run. Model-family aggregates can combine versions and have different eligible samples. Numeric factor and scenario structures do not enter binary Brier scoring. Inspect the measurement record.
05 · Integration access
Public archive reads are unauthenticated. Private workspace routes require an authenticated account and workspace membership. A public record identifier or developer key does not grant another team’s workspace access.
Read the public archive as JSON or export event records as CSV. Use dated examples to evaluate fit and validate your parsing before planning an integration.
View the endpoint reference ↑Bring one real research workflow. We can scope the questions, credentials, model access, expected volume and cost before a bounded evaluation. New model requests can incur charges; archive reads do not trigger those assessments.
Discuss a private review →A separately provisioned Probability API evaluates new binary questions. Its host, credentials, limits and cost need to be agreed for that integration; an account plan or workspace invitation alone does not establish access. How resolution works.