Kāla Point-in-Time
Most language models quietly use what happened later, so backtests built on them look better than they are. On our benchmark, gpt-oss 120B changed its answer because of later documents about 1 time in 4. Kāla's API never shows its model evidence dated after the date you choose, and states its own training cutoff on every answer, so you know which dates are clean.
As-of API
Probability, reasoning and dated sources for any question, as of a date you choose. Clean for dates after the model's knowledge cutoff (now 2026-10-15); earlier dates are flagged. Details →
Forecast datasets
Kāla's forecasts on resolved questions, each timestamped and hashed before its outcome was known: for research and model evaluation.
Point-in-time checkpoints
Model versions that have read the news only up to a given month, for backtests without look-ahead bias.
Hindsight-leak audits
We test your model on Kāla-Bench: how often a future document changes its answer, and on which kinds of question.
Try the data
Download a free sample: Kāla's own most recent forecasts, up to 50 resolved and 50 still open, each with the moment it was made, the forecast, the day its fingerprint was timestamped, and (once known) the outcome with sources. kala-forecasts-sample.jsonl → Open rows let you check the forecast was public before the result; the resolved share grows every day. Every forecast's fingerprint is published daily and timestamped with OpenTimestamps, so the dates can be checked independently: commitments →
How an audit works
We run your model (any OpenAI-compatible API) on Kāla-Bench: each question is asked twice at a moment before its answer was public, once with only the documents published by then and once with later documents mixed in. A leak is when the later documents pull the answer toward what actually happened. You get the leak rate overall and by source, honest forecasting skill, and examples, as a short report.
Datasets, checkpoints and audits are licensed per project. Tell us what you're working on.