research model · pre-registered · open notebook

Belief, as of
any moment

Tev reads a dated evidence timeline and answers a question as of any point on it: what it believes, whether the answer is knowable yet, how long until it will be, which passages it relied on, and what to watch next. A custom attention mask makes it physically unable to see anything dated later.

Scrub the timeline ↓ ● Watch it live Read the lab notebook

drag across the trace · illustrative

as of

belief

knowability

time to knowable

receipts

next source to watch

five outputs, one pass

Not just a probability. A position in time.

Every probe on the timeline gets five heads' worth of answers, computed in a single forward pass over the packed evidence.

Belief

A calibrated probability for the outcome, given only what was public at that moment.

belief: 0.31

Knowability

Is the answer already determined, not yet knowable, or is the question ill-posed?

state: not_yet

Time to knowable

A hazard estimate of how soon the answer is likely to become public.

hazard: ~9 days

Receipts

Pointers to the dated passages the answer leaned on, so it can be checked.

receipts: [e1]

Next source

Which kind of evidence would move the belief most if it appeared.

next: ship notice

leak-proof, verified

Add the future. Nothing moves.

This is Tev's actual unit test (tests/test_model_asof.py) on Qwen3-0.6B-Base, drawn as a picture. The probe's hidden state is compared before and after a change. Adding later evidence must leave it unchanged to within 1e-4; editing something the probe can see must change it.

May 02Acme says the Rocket phone is on track for a May launch.Reuters

May 09Acme delays the Rocket phone after a supplier fire.AP

May 12Probe: Will Acme ship the Rocket phone by May 20?as of

May 18The Rocket phone is now shipping to customers worldwide.Acme

probe hidden state (64 of 1,024 dims)

max |Δh| = 0.000000baseline

Choose a change on the left.

open lab notebook

Every round is registered before we look.

Each training round's gates are written down before any training or any reading of results, and the failures stay in the notebook. Here is the record so far. Watch the current training run live →

Inside training: three saved checkpoints answer the same 120 unseen practice questions. Every dot is a real answer, from a forward pass of the model.

Can it fit at all?Tev-0.6B, 100 timelines, overfit on purpose

PIPELINE CHECK›

Registered as a deliberate overfit check, not evidence of generalisation. A 10-example smoke run showed Brier on training examples falling from 0.32 to 0.16 with every loss term falling: the packing, mask, heads and losses can learn.registered gate: train Brier < 0.10, state accuracy > 0.80

Learned base rates, not readingTev-0.6B, public data mix, epoch 0 on the development split

GATES MISSED›

Brier 0.241 with evidence against 0.244 with no evidence at all; the crowd forecast scored 0.116; determined-state accuracy 0.55. The model was guessing the base rate. The cause was in the data: 63% of questions asked about outcomes no visible item contained, and nothing rewarded reading over remembering.registered gate: Brier within 0.02 of the crowd; determined accuracy ≥ 0.85

Tev-R: reinforcement collapsed to extremestext companion, GRPO with a 1 − Brier reward

FAILED›

Dev Brier 0.289, worse than a constant forecast at 0.249. 1,353 of 1,500 answers were exactly 0 or 1, often paired with "not yet", a contradiction. Without a proper log-score and a KL anchor, the policy learned bravado.registered gate: beat the constant forecast

Evidence-gain loss: collapsed to 50%0.6B, 1.7B and 4B

GATES MISSED›

All three sizes answered about 50% everywhere: Brier 0.240 / 0.249 / 0.248 with evidence against 0.237 / 0.248 / 0.248 blind. The reason, found afterwards: about half of the "determined" news probes had no answer anywhere in their timeline, so the rule "beat the blind twin" was impossible there, and the cheapest fix for the model was to ignore evidence.registered gate: Brier with evidence ≥ 0.02 below blind; determined accuracy ≥ 0.85

Teacher labels: the first sign of reading4B on a cloud A100, 2 epochs

GATES MISSED›

A stronger model marked where each timeline actually shows its answer; numeric evidence got explicit comparisons; easy questions came first. Settled-question accuracy 57% with evidence vs 40% blind, the first time reading helped, but far from 85%, and the state head degraded.registered gate: reading and settled-state accuracy ≥ 85%, leak 0, Brier 0.02 below blind

Can it read at all? Yes.diagnostic: synthetic questions with a stated outcome

PASSED›

When one document states the outcome, Tev-0.6B reaches 100% settled accuracy on unseen questions within about 50 updates. Evidence does reach the belief head; the real-data failures were about the training signal, not the architecture.check: settled accuracy → 100% on held-out synthetic questions

Taught where to look: still 50/500.6B check, 1 epoch on a cloud A100

GATE MISSED›

Settled-question accuracy 52% with evidence vs 52% blind, so the 4B run was not launched. The state head did learn: 87% right on settled / not yet / badly asked (R4: 25%). A follow-up check found why reading failed: the belief head had collapsed to a constant 50% (±0.004), even on easy practice questions. Most training questions carried labels the timeline can't support, so always saying 50% was the cheapest way to lower the loss.registered gate: settled accuracy ≥ 70% and Brier 0.01 below blind, on the 0.6B check

Train only on what can be read: passed, with a catch0.6B, 2 epochs, cloud A100

PASSED (NARROWLY)›

Training kept only questions whose deciding document is known: official data releases, and news where a stronger model found the answer in the timeline (2,673 of 4,896), plus 30% practice questions. On held-out questions whose answer is readable, accuracy was 75.9% with evidence vs 51.5% blind (best after epoch 1; epoch 2 got worse), and answers varied, so the belief head no longer collapses to 50%. The catch: the gain comes from official data releases and practice questions (100%); on news the model was 49% with evidence vs 53% blind, so reading real news text is still unsolved and the 4B run stays on hold.registered gate: accuracy ≥ 75% on readable questions and 15 points above blind; answers must vary (no collapse to 50%)

Tev-R with a proper scoring rulelog-score reward, state bonus, contradiction penalty, KL 0.04

NO RESULT›

The run stalled at its first step for 10 hours on a cloud T4 and exited with an error, so there is nothing to score. Generating samples for reinforcement learning is too slow on that GPU; Round 7 pursues the same goal more cheaply.registered gate: Brier < 0.2441 (constant); ≤ 20% answers exactly 0 or 1; < 2% unparsed

Averaging models: a small edge, not provenAsk Kāla's engine · Kāla-Bench lite, 305 resolved questions · re-scored after retries

NOT SIGNIFICANT›

Four models forecast each question from the same evidence, cut off before the question's date: gpt-oss 120B, Gemini 2.5 Flash, Qwen3 235B and Kimi K2, plus a gpt-oss variant that first breaks the question into sub-questions. Each forecaster was calibrated the same way on 200 separate questions. The first run lost 14–18% of answers to unreadable replies; re-asking those models brought failures to under 1% and changed the result. Brier scores (lower is better; always saying 50% scores 0.250): Kimi K2 0.196, average of all 0.197, average of three without Kimi 0.202, a judge model combining them 0.204, gpt-oss 0.207, Gemini 0.209, Qwen 0.216, gpt-oss with sub-questions 0.216. The average beats gpt-oss by 0.009, but the 95% interval (−0.021 to +0.002) includes zero, so the test is not passed; it ties the best single model. The first run's result (0.200 vs 0.214, interval excluding zero) did not survive the retries. Ask Kāla still uses the average, which is at least as good as its best member here and safer than betting on one model, and 0.197 is the bar Kāla-1 has to beat.test: beat gpt-oss alone on the same questions, with a 95% interval excluding zero

Long-memory hybrid: text solved, recall half-wayQwen3-0.6B · 22 of 28 layers converted · window 512 · cloud A100 · three runs

GATES MISSED›

The hybrid keeps full attention in 6 layers and gives the other 22 a 512-token window plus a compact running memory, so its memory stops growing with the length of the history (about 4.3× less at 32K tokens). It learns by copying the original model's predictions. v4a (Hedgehog memory): books 16% and 142% worse at 4K and 16K tokens, no passkeys found; dropped. v4b (gated delta memory, layers chosen by loss on text): web text +0.2%, books +2.5% and +6.5%, all within the limits, but it found the hidden passkey 0 times out of 12 at every length. v4c kept the layers that recall depends on (measured one layer at a time) and spent 10% of training on recall practice written differently from the test: text quality held (+0.3%, +2.8%, +5.9%) and passkeys came back, 9/12 at 8K, 4/12 at 16K, 9/12 at 32K, 7/12 at 48K, past the 32K the original was trained on. The original finds 12/12. Paused while Kāla-1 comes first; the next step would be more full-attention layers or more recall practice. Replay all three runs →registered gates: web within 5%, books 4K and 16K within 10%, passkey ≥ 90% of the original's at 16K and ≥ the original's at 48K; pass = all four

Leak-proof propertyhidden state of a probe vs. later evidence, real backbone

PASSED›

On Qwen3-0.6B-Base, a probe's hidden state is unchanged when later evidence is added (atol 1e-4) and changes when a visible item is hidden. On Kāla-Bench lite, Tev's leak rate is 0.0%, against 26.7% for gpt-oss 120B and 23.8% for Gemma 4 31B.gate: Δh ≤ 1e-4 for later evidence; Δh > 0 for visible edits

honest scoreboard

Where Tev stands today.

Leak-proofing works. Reading does not, yet. Round 2 is no better than a coin flip at forecasting; these numbers are the baseline round 3 has to beat.

Head-to-head: 53 questions closing on or after 1 Aug 2026, every model reading the same evidence as of each question's midpoint; crowd forecast on the 43 where one existed. Kāla-Bench: 307-question lite split, 17 Oct 2026. Lower Brier is better; 0.250 is always answering 50%.

how we keep score

The Brier score, by hand.

Tev is judged by a proper scoring rule: you score best by reporting exactly what you believe. Set a forecast and an outcome to see what it costs.

YOUR FORECAST

70%

WHAT HAPPENED

Brier = (forecast − outcome)² = 0.090

architecture

A small model with a clock inside.

Tev wraps an open base model with an as-of attention mask, five heads and a training objective that rewards reading over remembering.

input

Packed timeline

Question, rules and dated items, with probes interleaved at the moments to answer.

backbone

Qwen3 base + LoRA

0.6B to 4B parameters, rank-32 adapters; dense or flex attention.

mask

As-of attention

Each probe attends only to items dated at or before it. Verified on the real model.

heads

Five outputs

Belief, state, hazard, receipts, next source, trained jointly with temperature calibration.

# evidence-gain loss (round 3): reward reading, not remembering
full  = tev(question, items_as_of_t)          # sees the evidence
blind = tev(question, [])                     # blind twin, same pass
loss  = nll(full, outcome) + 0.5 * nll(blind, prior)
if state == "determined" and has_visible_answer:
    loss += relu(0.5 - (logp(full) - logp(blind.detach())))   # beat blind by 0.5 nats

waitlist

Get Kāla-1 at launch.

One waitlist for Kāla and its lab: an email to confirm, then one when the open weights and demo are out. No spam, no sharing, removal on request.

what comes next

Tev is the lab. Kāla is the release.

What Tev proves in the notebook becomes Kāla-1: an open, downloadable model for time-safe reasoning, with a new hybrid architecture and its own benchmark.

Visit Kāla →