World 8 v0.1.0 · Research evidence · Historical replay

Forecast ≠ Decision ≠ Order.

A contract-based multi-agent market architecture where forecasts remain immutable, decisions are evaluated downstream, and execution is a separate object. This demo shows the frozen evidence—not a trading promise.

Open Zenodo snapshot World 8 Collection Explore replay evidence

Evidence boundary: no live trading, no autonomous capital deployment, no production-readiness claim, and no claim of universal market superiority.

Lifecycle contracts

52,920

RESOLVED · 0 integrity failures

Primary data

52,632

Hourly rows · BTC / ETH / SOL

Uncertainty

2,000×

Paired moving-block bootstrap replicates

Separation is the invariant

The architecture prevents a downstream policy choice from silently rewriting the original prediction.

Forecast

Target, horizon, probability, issued-at, data-cutoff, model/strategy versions, hashes, lifecycle.

→

Decision

Consumes forecasts plus costs, risk, portfolio state, constraints, and policy. Independently attributable.

→

Order

Execution artifact. Explicitly outside this historical-replay study; no live order path is enabled here.

Frozen replay evidence

Brier score is lower when probabilistic forecasts are better. The strongest positive finding is calibrated weighting vs equal-weight raw on the frozen crypto replay.

2025H2 out-of-sample · 1h · 6h event horizon

BTC

Equal raw 0.265259

Calibrated 0.248866

Δ −0.016393

ETH

Equal raw 0.264132

Calibrated 0.250563

Δ −0.013569

SOL

Equal raw 0.262727

Calibrated 0.250645

Δ −0.012082

95% paired moving-block bootstrap intervals are entirely below zero for all three calibrated-vs-raw comparisons: BTC [−0.021891, −0.011495], ETH [−0.020824, −0.005867], SOL [−0.016465, −0.007784].

Independent daily replication · SPY / QQQ / GLD · 2025 OOS

Asset	Equal raw Brier	Calibrated Brier	Delta	95% bootstrap
SPY	0.256552	0.232693	−0.023859	crosses 0
QQQ	0.258656	0.234116	−0.024540	crosses 0
GLD	0.238877	0.234576	−0.004301	crosses 0

Directionally consistent point estimates, but not statistically robust in this smaller replication. The demo intentionally does not transform this into a cross-market superiority claim.

Mechanism	Frozen result	Interpretation
Correlation penalty	No useful general gain	Negative result retained
Disagreement shrink	Supported small gain only for SOL	No general claim
Regime weighting	ETH slight gain · BTC worse · SOL inconclusive	Asset dependent
Shadow cold-start	Neutral	No benefit claimed
Volatility risk veto	No Decision/UOP benefit	Shows why Decision ≠ Forecast

Evidence chain

Software snapshot

doi:10.5281/zenodo.22127650

Release commit: b14f2feea0fa233851a774d6ebd295b63cde75c0

Frozen empirical evidence

Evidence commit: 917dd82ed87a3470acfdb9175905ec7c8727c096

Evidence archive SHA256: 100484ffba683111622377703e836728817fd6cbb45f53d62e45a5a3766ece70

Forecast lifecycle

52,920 RESOLVED0 invalidated0 expired0 withdrawn0 integrity failures

Study classification

RESEARCH EVIDENCE / HISTORICAL REPLAY / NO LIVE TRADING

What this demo does not claim

It does not claim that World 8 is profitable or beats markets.
It does not claim production readiness or autonomous capital deployment.
It does not claim calibration is universally superior across markets.
It does not claim general benefit from correlation control, regime detection, disagreement handling, shadow cold-start, or risk veto.
It does not claim AGI, consciousness, autonomous intelligence, or causal architectural superiority.