Give Kāla a question, a date and some documents. It answers as of that date, says plainly when the answer is not yet knowable, shows which document it relied on, and is built so it can never peek at anything published later.
How it works, in 20 seconds
Watch one question go through Kāla, or click a step to jump to it.
“Will the monsoon reach Kerala before 1 June?”, as of 20 May.
Every document carries the date it became public.
Anything dated after 20 May is locked out inside the model, not just hidden from the prompt.
A calibrated probability, whether it is knowable yet, and the document relied on.
Will the monsoon reach Kerala before 1 June?
as of 20 May
The problem
Ask a model what was known on a past date and it quietly mixes in what it learned later. Ask about something still undecided and it often answers anyway. For research, audits, journalism, education and any system that replays history, that hindsight is a silent bug.
Evidence from after the question date changes the answer, even when the model is told to ignore it.
Models rarely say "this isn't decided yet". They guess, and the guess looks like knowledge.
You can't tell which document, from which date, an answer came from.
Move through time
Pick a question, then drag (or press play). Documents left of the line are the past and readable; those to the right are sealed. The answer changes only when the evidence does.
Illustrative scenarios and numbers, not live model output. For live output, see the live board.
The leak test
Same question, same evidence, as of 20 May. Then a report from 2 June is smuggled in. Press the button and watch who changes their answer.
Illustrative. Measured on Kāla-Bench: gpt-oss 120B changed its answer this way on 26.7% of questions, Gemma 4 31B on 23.8%, Kāla's precursor on 0%.
Try it yourself
Add dated events about anything, mark any that settle the question, then move the date. This sandbox shows Kāla's time rule at work: what may be read, what is sealed, and when an answer becomes knowable. It applies the rule only; it does not run the model in your browser.
Kāla will show what it may read at the date you choose.
Three kinds of answer
Alongside a probability, Kāla says whether the matter is settled, still open, or badly asked. Tap a card to flip it.
determinedtap to flipThe evidence by this date already fixes the answer. Kāla reads it and points to the document that settles it.
not_yettap to flipNo one could know yet. Kāla gives its best honest probability and says the matter is still open.
ill_posedtap to flipThe question has no clear answer as written: vague terms or a false premise. Kāla says so instead of inventing one.
Track record
Chatbots give an answer and move on. Ask Kāla saves every forecast before the outcome is known, scores it against what actually happened, and publishes the results, misses included. Every day it also forecasts about ten public events of its own, so the record grows whether or not anyone is asking.
Scored automatically as each question resolves.
Each day's forecasts are fingerprinted and timestamped in the Bitcoin blockchain (OpenTimestamps), so none can be edited or backdated.
Kāla now forecasts in the Metaculus AI tournament (as Kala-Bot) and enters ForecastBench from its 25 October round, where it is scored alongside other AI forecasters.
See the scoreboard → The first of Kāla's own forecasts resolve in late October; results appear there daily.
Live from the lab
Our models train in the open. Follow the run on the GPU right now: what stage it's in, how fast the gap to its teacher is closing, and whether it passes the targets we set before it started.
Kāla-Bench · v0.1
1,720 resolved questions closing on or after 1 July 2025, each with dated evidence. Every question is asked twice: with only past evidence, and with future documents slipped in as traps. Pick a measure.
Preliminary results on the 307-question lite split, October 2026. Tev R2 is Kāla's 0.6B precursor: it never leaks, but does not yet read evidence well. Closing that gap is what Kāla-1 is being trained to do.
Today Ask Kāla combines several AI models, which scores level with the best single model and slightly better than the main one on this hindsight-proof benchmark (not yet statistically certain). Kāla-1, our own model, is being trained to beat that combination. Details in the research notebook →
Confident calls
No model can predict everything with certainty: some things are simply not decided yet. Kāla's promise is narrower and checkable: it makes a confident call only when the evidence settles the question, and shows honest odds otherwise. Every call is committed to a public, tamper-evident ledger before the outcome is known.
confident calls made. Nobody can know next month's figure in advance, so Kāla declines to be sure, and gives honest odds instead.
confident calls, all correct: the answer is in the evidence, so Kāla reads it, says “settled”, and cites the release.
Kāla system on Kāla-Bench v0.2 (official US statistics, market-free), 17 October 2026. After-release answers use exact arithmetic on the published figure. Follow the running tally on the live board.
Developer preview
Kāla-1 will ship as open weights on Hugging Face, small enough for a laptop GPU, with a one-command demo. The interface below is planned.
# planned API — weights not yet released from kala import Kala model = Kala.from_pretrained("mastyf-ai/kala-1") answer = model.ask( "Has the RBI cut the repo rate this year?", as_of="2026-02-10", evidence=docs, # [{"date": "2026-02-07", "text": ...}, ...] ) answer.state # "determined" | "not_yet" | "ill_posed" answer.belief # 0.0 – 1.0, calibrated answer.receipts # the dated passages it relied on
Replay what was knowable on any day without hindsight contaminating results.
"What was known when this was said?" answered with dated receipts.
A reader for retrieval systems where documents have dates and order matters.
Reconstruct a decision against only the information that existed at the time.
Teach history and science as they unfolded, not as they are remembered.
Use Kāla-Bench to measure hindsight leakage in any language model.
Roadmap
Where things honestly stand.
The mark
The 24 carved wheels of the Sun Temple at Konark, Odisha, are working sundials: the shadow of the axle across the spokes tells the hour. Kāla (काल, "time") takes its mark from that wheel, because the model does what the wheel does: it reads the world as of one moment. Hover each colour.
Waitlist
Get one email when the open model and demo launch. No spam, no sharing, and you can ask to be removed anytime.
In plain words
No. Telling a model to ignore something does not work reliably: on our benchmark, large models still changed their answer because of future documents about one time in four. Kāla removes the later documents from what the model can attend to at all, so they cannot influence the answer.
No model can know the future. Kāla gives an honest probability based only on what was public at the date you choose, and says "not yet knowable" when the answer is not out yet. Its job is to be careful about time, not to be a fortune teller.
A way to grade probabilities. Say 80% and be right: small penalty. Say 80% and be wrong: big penalty. Always saying 50% scores 0.25; lower is better.
No model can, because many outcomes are not decided yet. Kāla is built to know the difference: it is confident when the evidence settles a question (and right when it is), and gives honest odds when it doesn't. That is a promise you can check on the live board.
Not yet. The precursor model works for time safety but is still learning to read evidence well. The live board shows its real results, good and bad, every day. Open weights come when it passes its quality gates.
Official US statistics from FRED (Federal Reserve Bank of St. Louis): unemployment, inflation, jobless claims and similar. No betting or market data.