Clean room data for predictive AI
AI that predicts the world,
trained only on what was knowable.
Offdiagonal records every forecast and every answer with the exact moment it became known, across 12 domains, so models learn without seeing the future.
295M rows · 334 tables · 15 verified releases · audited 2026-10-29
The problem
Forecasting AI fails silently, at the data.
01
It sees the future
Backtests quietly use data that was revised or published after the decision. The model looks brilliant, then fails live.
02
It counts copies as a crowd
Ten agents on one base model agree because they are one model. Most datasets count them as ten independent views.
03
It is graded after the fact
Without a sealed prediction and a recorded answer time, nobody can prove the forecast came first.
What Offdiagonal does
A clean room between raw data and the model.
Four clocks on every record
When it was issued, what it was about, when the answer happened, and when the answer became public. Training sees only what was known at the decision.
Sealed predictions
Agents commit a fingerprint before the answer exists. Late forecasts cannot enter a scorecard.
Families counted once
Models sharing a base are one family, so agreement means something.
Frozen, fingerprinted releases
Every release is append only and checked file by file by sha256. Defects are published, not hidden.
Growth, measured
Fifteen verified releases in four days.
Rows grew from 260.9M across 310 tables on 2026-10-25 to 295M across 334 on 2026-10-29. Every number here comes from a dated audit or release note.
- 295M
- rows recorded
- 15
- fingerprinted releases
- 11,147,828
- wallet trades audited
- 12
- domains
334 tables, audited 2026-10-29
10-26 to 10-29 Oct 2026
up from 10,616,400 in v6.21
808 forecasters in 82 families
Show all 15 releases
- v6.202026-10-26Minute level market prices at every check
- v6.212026-10-26Wallet lane rebuilt from 10.6M trades
- v6.222026-10-26674,741 eligible bets from 79,393 wallets
- v6.232026-10-2712,145 target cards and 21,506 forecast shapes
- v6.242026-10-27Source scorecard and central bank survey gates
- v6.252026-10-272,192,195 weather ensemble summaries
- v6.262026-10-27Outcomes on 1,138,371 ensemble summaries
- v6.272026-10-27Matched calibrated comparison
- v6.282026-10-2759,438 Kalshi threshold links
- v6.292026-10-27Decision clock audit on every shared row
- v6.302026-10-27Whole week partitions, 98 quantity IDs
- v6.312026-10-27Answer clocks for 13,151 market questions
- v6.322026-10-27Sealed beliefs and countercases
- v6.332026-10-28Conviction and influence for every bet
- v6.342026-10-29Zeno30 agents, 185,418 Jev classifications, 98,379 Manifold answers
Storage cleaned on 2026-10-29: 1,463 GB to 201 GB, with every kept file checked against its fingerprint.
What is in the record · measured 2026-10-29
Headline numbers
- 13.5M
- forecasts
- 19,019
- markets with full price history
- 808
- forecasters in 82 families
- 12
- domains
- 431k
- weather training cases
- 38
- data defects published
ForecastBench panels: 13,523,509 rows, 6,311,674 with real answers
Of 19,030 verified temperature markets; 11 have no Polymarket history, 0 failed (v6.11)
756 AI models, 38 human groups, 6 baselines, 8 aggregates
Geopolitics & world events, Single-run weather, Ensemble weather, Prediction markets, Market traders, Crypto, Central-bank surveys, Economics, Energy, Health, Space weather, Hurricanes
205,746 single-run + 225,235 ensemble, 59 sites, 18 weeks
Two audits on 2026-10-20: all 11 critical fixed (invented or early clocks, answer-conditioned forecasts, stalled collectors); 12 lower-severity still open. Affected rows are excluded or flagged, never silently kept.
Coverage
13 domains, one set of clocks
Bar length on a log scale. Muted bars are test-only: kept out of training because their publication times can't be proven.
- Geopolitics & world eventsForecastBench56,013answered question-horizons · Jul 2024 – Sep 2026
- Single-run weather7 model centres205,746cases · Jun – Oct 2026
- Ensemble weatherOpen-Meteo ensembles225,235cases · Sep – Oct 2026
- Prediction marketsManifold, Polymarket, Numinous170,096resolved questions with prices (97,788 Manifold, 71,195 Polymarket, 1,113 Numinous) · Aug – Oct 2026
- Market tradersPolymarket wallets457,874trades · 2026
- CryptoKalshi, Deribit, reference models255,807forecasts (mostly replays) + 51,459 venue quotes · Jul – Oct 2026
- Central-bank surveysECB SPF156,912probability distributions · 42 rounds
- EconomicsPhiladelphia Fed SPF1,979,271survey responses (2,250 first-release outcomes, RGDP/PGDP only) · 1968 – 2026
- EnergyNESO demand vs Elexon outturn137resolved half-hours (1,449 admitted forecasts at release) · Oct 2026
- HealthCDC FluSight, COVID-19 and RSV hubs567,189distinct dated outcomes (12,433 flu, 536,748 COVID, 18,008 RSV) · 2020 – 2023
- Space weatherKp observations103,623observations · Sep – Oct 2026
- HurricanesNOAA NHC ATCF forecast decks + best track152,707forecast rounds (one per method and cycle) + 1,711 best-track rounds · 2024–2025 seasons · collecting
- Market pricesPolymarket price history119,744pre-deadline price points on 19,019 temperature markets · 2026
Proof you can check
Do not trust us. Check.
Show more — vision and investor questions
Where it goes
The verification layer for predictive AI.
Direction, not results yet: as models move from text to decisions, the scarce input is real world evidence with provable timing. Offdiagonal builds that evidence once and serves it to everyone who trains or buys forecasts.
AI labs
Leak free, time stamped episodes for training models that act in the real world, plus sealed evaluations.
Trading and energy desks
An audit of the forecasts and vendors they already pay for, checked against what was knowable at the time.
Benchmark makers
Prospective tests where the answer did not exist when the forecast was made.
Questions investors ask
Straight answers.
What does Offdiagonal sell?
Clean, time stamped forecasting evidence: training episodes for AI labs and forecast audits for teams that buy forecasts.
Who uses it today?
Zeno Divergent trains on every release, and the Zeno30 team runs live AI agents whose sealed forecasts flow back into the corpus.
Why can't a lab do this itself?
Labs clean formats, not timing. Proving when each fact became public across many sources is slow, dull work that compounds once done.
How big is it?
About 295 million rows across 334 tables and 12 domains, audited 2026-10-29.
How do you know the data is clean?
Every release passes a gate for clock order, missing values, family identity and split isolation, and every file is checked by fingerprint after upload.
What happens with mistakes?
They are published. 38 data defects have been recorded openly and all 11 critical ones fixed.
What is not proven yet?
Value beyond simple calibration has not been demonstrated. Sealed prospective tests are running now to measure it.
Show more — which forecasters miss together
Forecasters
Which forecasters miss together?
Nine weather models, 7 marine sites, 21 frozen days.
- pair
- ecmwf aifs × ecmwf ifs
- shared outcomes
- 1,592
- miss rate A / B
- 16% / 16%
- missed together
- 7.5%
- if independent
- 2.4%
- ratio
- 3.12×
- SHARED OUTCOMES
- 1,592
- WORST PAIR
- 3.1× chance
- CONFIDENT AND WRONG
- 35,970 cases
Latest frozen release
handoff-v6.36-20261029
commit 577c34f3ae43 · +6 files · verified 2026-10-29All data we hold is cleared for sharing with Zeno and training. Only leakage, missing outcomes and Zeno's own test set are kept out. All 6 files re-downloaded from Hugging Face and matched by sha256.
More — this week, method, paths, live record
At a glance · 2026-10-29
What changed this week
Market prices
99.94%
19,019 of 19,030 markets priced. 11 have no history; none filled with zeros.
Forecasters with fixed IDs
- COVID-19 hub (legacy)129
- FluSight126
- FluSight (legacy)44
- COVID-19 hub29
- RSV hub20
- Weather ensembles5
- Crypto replay3
356 forecasters, 10 documented links. Without a link, independence is unknown.
Warning list
5,116 cases · 500 resolved so far
Every case is registered before the outcome, at most 50 alerts per run. Outcomes arrive as events resolve.
Releases
- v6.22.1 Baseline: wallet files corrected (immutable pin)
- v6.32 Sealed beliefs, countercases, action records
- v6.33 Bet structure for 674,741 bets
- v6.34 Zeno30, Manifold answers, Jev lane
- v6.35 Energy and weather paired with real outcomes
- v6.36 62,280 weather forecasts vs measured daily highs (current)
Each release is re-downloaded and fingerprint-checked. v6.22.1 is the immutable baseline; later releases are add-ons. Full file list in the corpus inventory.
Honest comparison
Not the biggest archive. The one built to show shared failure.
Single-domain archives
ECMWF, CDC forecast hubs
They have more: Decades of history in one field.
We add: Many domains joined, each forecaster tagged by family.
Market price feeds
Polymarket, Kalshi
They have more: Deeper live price history.
We add: Individual forecasters and outcomes dated by when they were actually known.
Forecasting benchmarks
ForecastBench, Metaculus, Good Judgment
They have more: Longer human track records in their own question sets.
We add: Who misses together, across benchmarks and the physical world, without future leakage.
Door A / Decision teams
Audit the forecasts you already buy
Analyst estimates, vendor feeds or internal models, assessed against observed outcomes and each other — like the marine audit above, on your data.
Door B / Zeno and ML teams
Train on decision-time episodes
Whole episodes as seen at the decision: every forecast available then, each forecaster's track record, the outcome kept apart. Explicit clocks, verified packs, no leakage.
Current record / Separate from the frozen example
Live evidence, as recorded
Evidence at the decision
The forecast and prior history are kept apart from what was known only afterward.
The frozen worked receipt is temporarily unreadable. Nothing is substituted for it.
Shared-case comparison
Signed errors on outcomes both models answered; not an explanation of shared failure.
Reading recorded comparisons…