Crypto · 2025-09-01
coinbase-btc-usd-1h: value at 2025-09-01 18:00 UTC
- History behind it
- 48 past records
- Warning
- 58%
The Surprise Intelligence Model
Which agreement earned your trust?
When many forecasts agree, Zeno checks whether that agreement is earned or fragile. It compares answers, track records and shared mistakes before it adjusts, warns or keeps.
01
forecasts
answers, ranges and timing
→02
track records
what was known before the decision
→03
shared mistakes
who has failed together
→04
trust decision
adjust, warn or keep
01
forecasts
answers, ranges and timing
02
track records
what was known before the decision
03
shared mistakes
who has failed together
04
trust decision
adjust, warn or keep
today
What Zeno does right now.
The served answer, the public beta and the tests in progress.
- Serving
- v0.57
- Public beta
- v0.80.2 · finished · energy + weather vs airports passed · opt-in beta · fixed-scoring re-check adds weather vs analysis (retrospective) · others keep the reference
- Live test
- Forecast Failure Watch · every open weather case registered before its outcome · scored hourly
- Next training
- Separate corrected successor · requires approval before launch; complete Offdiagonal v6.11 prices remain a separate data milestone
01 · training now · v0.82
The full-power run, across every field.
All 24 fields, ForecastBench included, on an H200. Judged on three separate skills: correcting a range, warning of a confident miss, warning of a shared failure.
ForecastBench error 23.1% lower than each forecaster's track record
best of five rivals · 45.3% on development
v0.82 rehearsal · Offdiagonal v6.10
Beats every rival on sea (+28%), crypto (+14%) and weather (+6–8%)
five rivals incl. calibrated market
v0.82 rehearsal
3.7M scored past forecasts, 24 fields, 1,474 ranked forecaster-field entries
SurpriseBench Hall of Record
frozen Offdiagonal snapshot
02 · public beta
v0.80.2: the public beta panel reader.
Trained on 21 fields; the corrected judge decides which fields pass. v0.57 keeps answering until then.
FieldOriginalFixed scoring · re-check
- Energy demandvalidated improvementvalidated improvement
- Weather vs airportsvalidated improvementvalidated improvement
- Air pressurecalibrated referencecalibrated reference
- Wind 10 mcalibrated referencecalibrated reference
- Cryptocalibrated referencecalibrated reference
- ECB surveycalibrated referencecalibrated reference
- Phil. Fed (prob.)calibrated referencecalibrated reference
- Phil. Fed (numbers)calibrated referencecalibrated reference
- COVID-19calibrated referencecalibrated reference
- Flucalibrated referencecalibrated reference
- RSVcalibrated referencecalibrated reference
- Market pricescalibrated referencecalibrated reference
- Polymarket traderscalibrated referencecalibrated reference
- Seacalibrated referencecalibrated reference
- Space weathercalibrated referencecalibrated reference
- Single-run weathercalibrated referencecalibrated reference
- Weather vs analysiscalibrated referencevalidated improvement
- Agriculture raininsufficient evidenceinsufficient evidence
- Air quality PM2.5insufficient evidenceinsufficient evidence
- Cross-venue marketsinsufficient evidenceinsufficient evidence
- Urban heatinsufficient evidenceinsufficient evidence
Fixed scoring: a few zero-width standard ranges had scored losses in the billions. The re-check scores the same saved models with one shared minimum width learned from training data only. It looks again at periods already seen, so it is a correction, not new proof; a release needs a fresh period.
01
Reads the panel
Up to 16 forecasts at a time, each with its track record, shown as positions relative to the calibrated reference.
02
Compares forecasters
3 attention layers (4 heads) that weigh each forecaster against the others, using time-valid pair history: who misses together.
03
Sees conditions
32 decision-time world readings for weather cases; switched off when absent.
04
Answers four things
A bounded correction, a calibrated probability, the chance of a confident miss, and the chance of a shared failure.
Small on purpose: v0.69 tested larger models: bigger was worse on the locked rule. An untrained model gives exactly the reference, so any change from it has to be learned.
Air pressure
1,134,793 labelled
Wind 10 m
1,134,793 labelled
Urban heat
154,115 labelled
ForecastBench
reference only
Crypto
83,504 labelled
Market prices
65,849 labelled
Weather vs analysis
48,708 labelled
Space weather
25,855 labelled
Single-run weather
22,960 labelled
Weather vs airports
22,187 labelled
Sea
12,096 labelled
COVID-19
6,365 labelled
Polymarket traders
4,469 labelled
Phil. Fed (numbers)
4,190 labelled
RSV
2,891 labelled
Energy demand
2,667 labelled
Flu
2,042 labelled
Phil. Fed (prob.)
1,717 labelled
ECB survey
1,098 labelled
Air quality PM2.5
644 labelled
Cross-venue markets
325 labelled
Numinous
reference only
Agriculture rain
31 labelled · too few to judge
Ensemble weather
reference only
- forecasts
- 12,134,209
- panels
- 4,599,358
- with outcome
- 2,840,152
- distinct events
- 1,265,459
Time audit passed: 0 forecasts used after their decision time, 0 outcomes known before it. Most of the growth over v0.80 comes from the new air pressure and wind fields. Every field counts equally in judging, however large it is.
03 · every field
One model of models, 24 fields.
Atmosphere, space and sea, crypto and markets, health, economy, energy and general forecasting: what Zeno corrects today and how its early warnings score in each.
24 fields · 7 areas · 6 already beat every rival
- Warnings · v0.82 warning test
- Brier score of confident-miss warnings vs each forecaster's own track record (tree). historical: development and later confirmation slices of past data. exploratory · release gate not yet passed.
- Corrections · v0.80.2 public beta
- forecast loss vs calibrated reference vs calibrated reference. historical: two sealed test periods. passed fields are served in the opt-in beta.
Atmosphere
- Sea-level pressure1.13M outcomes · 2 forecastersWarnings: no edge yetCorrections: calibrated reference
- Wind1.13M outcomes · 2 forecastersWarnings: no edge yetCorrections: calibrated reference
- Urban heat154k outcomes · 2 forecastersWarnings: no edge yetCorrections: insufficient evidence
- Weather vs analysis49k outcomes · 6 forecastersWarnings: beats every rival (+7.9% vs track record, +10.5% later)Corrections: calibrated reference
- Single-run weather23k outcomes · 7 forecastersWarnings: too few misses to judgeCorrections: calibrated reference
- Weather vs airports22k outcomes · 5 forecastersWarnings: beats every rival (+5.6% vs track record, +6.3% later)Corrections: validated improvement
- Air quality644 outcomes · 1 forecastersWarnings: too few misses to judgeCorrections: insufficient evidence
- Farming rain31 outcomes · 2 forecastersWarnings: too small to judgeCorrections: insufficient evidence
- Ensemble weather0 outcomes · 6 forecastersWarnings: outcomes arrivingCorrections: not trained
Space and sea
- Space weather26k outcomes · 4 forecastersWarnings: too few misses to judgeCorrections: calibrated reference
- Sea12k outcomes · 24 forecastersWarnings: beats every rival (+28% vs track record, +23% later)Corrections: calibrated reference
Markets and crypto
- Crypto84k outcomes · 3 forecastersWarnings: beats every rival (+14.2% vs track record, +13.4% later)Corrections: calibrated reference
- Prediction markets66k outcomes · 1 forecastersWarnings: no edge yetCorrections: calibrated reference
- Market traders4k outcomes · 8248 forecastersWarnings: too few misses to judgeCorrections: calibrated reference
- Cross-venue markets325 outcomes · 617 forecastersWarnings: too few misses to judgeCorrections: insufficient evidence
Health
- COVID-196k outcomes · 16 forecastersWarnings: ahead of most rivals (+0.1% vs track record, +1.1% later)Corrections: calibrated reference
- RSV3k outcomes · 6 forecastersWarnings: ahead of most rivals (+2.3% vs track record, +24% later)Corrections: calibrated reference
- Flu2k outcomes · 66 forecastersWarnings: no edge yet (-15.1% vs track record, +9% later)Corrections: calibrated reference
Economy
- Philadelphia Fed (points)4k outcomes · 197 forecastersWarnings: ahead of most rivals (+37.8% vs track record, +23.8% later)Corrections: calibrated reference
- Philadelphia Fed (ranges)2k outcomes · 195 forecastersWarnings: too small to judgeCorrections: calibrated reference
- ECB1k outcomes · 116 forecastersWarnings: ahead of most rivals (+7.8% vs track record, +16.8% later)Corrections: calibrated reference
Energy
- Energy demand3k outcomes · 1 forecastersWarnings: too few misses to judgeCorrections: validated improvement
General forecasting
- ForecastBench109k outcomes · 806 forecastersWarnings: beats every rival (+45.3% vs track record, +23.1% later)Corrections: not trained
- Numinous154 outcomes · 2 forecastersWarnings: too small to judgeCorrections: not trained
04 · release
Only passing fields ship.
One request format across the site, API, MCP, Python and ONNX; every answer says whether it was corrected.
- 01
Sealed test
Each field is judged once against the reference, a tree model and a track-record model. Only passing fields get corrections.
- 02
Hugging Face beta
Weights, model card, per-field verdicts, ONNX file and the data fingerprint ship together. Publishing needs a separate approval.
- 03
Same answer everywhere
Website Check, the API, MCP, Python and ONNX take one request format and give the same answer for the same forecasts.
- 04
Labelled answers
Every answer is marked corrected or reference, with the reason. Agents read that label before acting.
- Check
- Paste forecasts, get trust, widen or review
- assess-any
- Check a forecast of any kind, with no account
- MCP
- zeno_check, zeno_assess_any, zeno_register_question, zeno_submit_forecast
- Sealed forecasts
- Agents seal blind, then warned, and learn what each warning was worth
Earlier recordSuperseded research: v0.80 and v0.79e+
v0.80 · superseded
v0.80: flu and energy looked promising.
Before the data repairs; retested in v0.80.2.
health-flu
+31.8%
markets-cross-venue
+31.3%
health-rsv
+9.7%
energy
+4%
space-weather
+0.5%
sea
-0.3%
weather-deterministic
-0.4%
markets
-1.4%
economy-philfed-point
-1.4%
health-covid
-3.2%
economy-ecb
-12.9%
crypto
-19.5%
markets-traders
-26.7%
economy-philfed
-39.4%
weather-ensemble
n/a
v0.79e · verdict
Reading the whole panel wins.
7 of 10 testable sources; research only.
SourceError vs best reference (shorter bar = better)Warnings caughtVerdict
Sea
3,175 outcomes
Zeno0.175
Best ref0.222
45% vs 30%
beats all four referencesWeather · single-run
5,532 outcomes
Zeno0.250
Best ref0.305
48% vs 11%
beats all four referencesSpace weather
73 outcomes
Zeno0.178
Best ref0.187
14% vs 0%
beats all four referencesPrediction markets
2,012 outcomes
Zeno0.175
Best ref0.185
no panel test
beats all four referencesEconomy · Phil. Fed points
120 outcomes
Zeno0.531
Best ref0.563
0% vs 0%
beats all four referencesWeather · ensemble
15,797 outcomes
Zeno0.584
Best ref0.595
22% vs 11%
beats all four referencesRSV
796 outcomes
Zeno0.390
Best ref0.400
14% vs 10%
beats all four referencesCrypto
24,996 outcomes
Zeno0.381
Best ref0.375
no panel test
loses to the tree referenceEconomy · Phil. Fed ranges
34 outcomes
Zeno0.666
Best ref0.646
7% vs 0%
no reliable differenceCOVID-19
2,786 outcomes
Zeno0.420
Best ref0.412
10% vs 6%
no reliable differenceNot testable: Economy · ECB. Warnings = share of failed questions caught when reviewing the riskiest 10% of questions (random = 10%). 134 of 134 models passed the validity check.
v0.79e · inside
The v0.79e panel reader.
615,966 parameters · width 128 · 3 layers · 4 heads
01
input encoder
18,560
Turns each forecaster's 15 time-valid features into one vector.
02
source adapters
1,820
One learned identity per source (13) plus a general slot for unseen sources.
03
attention layers
594,816
Every forecaster reads every other in the same panel.
04
miss-risk head
385
How likely this forecaster is to miss.
05
consensus head
385
How likely most independent families are to miss together.
Counted from the training code. Not a release: no download is offered until the full raw-panel package passes its parity test.
- ✓ Data rebuilt
- ✓ Timing measured
- ● Research training
- ○ Comparison
- ○ Confirmation
- ○ Serving
How it reads a panel
01 · Input
every forecaster's answer, native range, horizon, and time-valid track record (50+ earlier answers)
02 · Core
attention across the whole panel; pair history tested as a variant
03 · Heads
forecaster miss risk · confidently wrong consensus
04 · Checks
main 0.5M · 2M size check · no pair history · no attention · history only · 15% source hiding · leave one source out
- Architecture
- Panel reader · 0.5M and 2M parameters
- Data
- 13 sources · v079e-cache-20261020T0835Z-clockfix-all13
- Models
- 134 research models over 9 time folds
- Measured speed
- 2.8 s (0.5M) / 3.4 s (2M) per pass · worst case 3.5 h
- Scoring
- not scored yet · per-model numbers are diagnostic
- Answers users
- no · v0.57 stays served
Abilities, judged separately
Range calibration
measured separately (v0.79d)
Ensemble weather ranges 66% to 90% covered; COVID-19 85% to 91%. Not learned by v0.79e.
Forecaster miss risk
training now
How likely each confident forecaster is to miss, read from the whole panel.
Confidently wrong consensus
training now
Whether most independent families miss together. Not testable where outcomes are too few.
Unseen sources
experimental, unsupported
Leave-one-source-out tests decide; until then, new sources get no supported answer.
Which sources can test a confidently wrong consensus
| Source | Consensus failed | Judged | Status |
|---|---|---|---|
| ECB survey | — | 0 | not testable |
| Philadelphia Fed survey | 21.4% | 392 | testable |
| economy-philfed-point | 16.8% | 441 | testable |
| Crypto price ranges | — | 0 | not testable |
| COVID-19 hospitalisations | 12.1% | 1,281 | testable |
| markets | — | 0 | not testable |
| RSV | 13.0% | 323 | testable |
| ForecastBench | — | 0 | not testable |
| Space weather | 2.0% | 709 | testable |
| Sea conditions | 2.2% | 2,540 | testable |
| Deterministic weather | 5.7% | 5,408 | testable |
| Ensemble weather | 28.6% | 15,698 | testable |
- research only
- not scored yet
- v0.57 stays served
- unseen sources unsupported
05 · served now · v0.57
Proof controls the switch.
Crypto price ranges
Zeno correction
+11.7%
Ensemble weather
Zeno correction
+5.6%
Deterministic weather
correction · no history
+6.2%
COVID-19 hospitalisations
correction · no history
+11.8%
Sea conditions
correction · no history
+3.2%
Prediction markets
reference kept
RSV
reference kept
FluSight
reference kept
ForecastBench
reference kept
Numinous AI forecasters
learn-only
Evidence +−
Crypto price ranges
Ensemble weather
Deterministic weather
COVID-19 hospitalisations
Sea conditions
Prediction markets
RSV
FluSight
sourceheld-out gain over its referenceserved
Crypto price ranges
+11.7%
Ensemble weather
+5.6%
Deterministic weather
+6.2%
COVID-19 hospitalisations
+11.8%
Sea conditions
+3.2%
Prediction markets
kept
RSV
kept
FluSight
kept
real misses per 100 alerts
62
random alerts would catch
≈15
share of all confident misses caught
33%
ensemble weather · sealed confirmation weeks10% review budgetserved model v0.57
Ensemble weather
+9.75% Brier
COVIDnot passed
RSVnot passed
Flunot passed
Marketsnot passed
v0.57 gate · confirmation period · vs tuned GBMWarning labels exist for five sources; deterministic weather, sea and crypto are not scored here.
Known limit · coverage was about 81% (crypto), 85% (ensemble weather), 87–89% (others)
Inputs are admitted only when they were knowable; missing history remains missing. The model starts at the best simple forecast and moves only where earlier dates support it.
Answers Check, API and MCP today. Gate-passed per source: crypto, ensemble-forecasts. Every other source returns its reference, labelled.
parameters: not recorded
01
Forecaster panel
Per-source forecasts, admitted only if knowable before the decision
02
Panel encoder
Per-source model; size not recorded in the site's records
03
Surprise latent
Absent · No shared surprise latent in the served release
04
Range correction
Served on 2 gate-passed sources
05
Confident-miss warning
Absent · Not served
06
Shared-failure warning
Absent · Not served
07
Answer
Corrected range where passed, reference elsewhere
Trust intelligence
Zeno Check
live · v0.57
Send a panel of forecasts; get a risk read, the reason, its validated scope, and whether to keep the reference.
- — Weather operations
- — Energy forecasting teams
- — Public-health forecast review
- — Analysts wanting a second opinion
Collective fragility
Zeno Atlas
research
See which forecasters share assumptions and miss together, kept apart from proven warning value.
- — Ensemble builders
- — Teams cutting redundant sources
- — Auditors of forecasting systems
Decision intelligence
Zeno Investigate
future research
With a fixed review budget, decide which forecasts deserve a closer look first.
- — Operational escalation
- — AI-agent orchestration
- — Arena and Paradox Foundry
06 · real cases
Three replays, chosen by rule.
Picked at fixed percentiles of the score change, so one shows Zeno hurting.
Crypto · 2025-12-11
coinbase-btc-usd-1h: value at 2025-12-11 15:00 UTC
- History behind it
- 48 past records
- Warning
- 60%
Crypto · 2025-10-07
coinbase-btc-usd-1h: value at 2025-10-07 21:00 UTC
- History behind it
- 48 past records
- Warning
- 54%
Research replay from frozen run v044-h200-20261004T1335Z, not live Arena questions. Confirmation period, main model seed 11, continuous targets of judged sources; cases at the 10th / 50th / 90th percentile of CRPS(candidate) - CRPS(reference). Pool: 8,654 cases; the rule picked three Bitcoin hours because crypto dominates the pool. The candidate is not gate-passed and is not what the API serves.