Read the change.
Gold shows relative posterior mass at each observation. The white path traces the most likely value. Pause and use the slider to inspect a stable moment.
Watch belief change with evidence.
Observe how Solana market data changes a probability model. Follow the distribution, inspect each update, and keep the evidence behind every revision.
PROJECT CA
Loading…Belief field
relative mass / observation model modeUSD · observations →
Current distribution
Where probability sits now.
Information per update
How much each observation changes the distribution.
Concentration
A shape measure, not an accuracy score.
Source: deterministic preview · generated locally
Keep the record.
After four initial observations, a KL change of at least 0.1 nats enters the revision log. Export the window, source and model values to reproduce it.
Inspect revisions →Read the observation. Then the model.
Compare incoming values with the model’s evolving mode and probability interval. Supporting views explain how the distribution changes.
Observed value & model mode
95% posterior interval
Inside the distribution
Current distribution
Information per update
Distribution entropy
Distribution concentration
Measure behavior. Keep the limits visible.
These checks describe the current observation window. Concentration and historical interval coverage do not establish future accuracy.
Largest distribution change
—
nats · after four initial observations
Posterior displacement
—
total variation from uniform
Previous-band coverage
—
same-window descriptive fraction
Revisions observed
—
within the active window
Distribution change intensity
Each bar measures a distribution change. The dashed line marks 0.1 nats; the first four observations initialize the model and are excluded from revision entries.
Displacement from uniform
Distance from equal weight grows as probability concentrates. Displacement measures a distribution’s shape, not whether its mode is correct.
Historical interval coverage
Fraction of observations inside the previous posterior’s 95% interval. The grid uses the whole active window; this is a same-window description, not a leakage-free backtest or next-price guarantee.
Distribution & latest observation
Gold is the current posterior; the white marker is the latest observed value. They can differ substantially. This is model state, not a price forecast.
An observation changes a distribution.
Start with a concrete update from the active window, then inspect the rule behind it.
BEFORE / MODE—
→NEW OBSERVATION—
→AFTER / MODE—
DISCOUNTED BAYESIAN UPDATE
P(H | E) ∝ P(E | H) · P(H)0.96
01
256 explicit hypotheses.
The price grid spans 85% of the minimum to 115% of the maximum in the active window. Each bin starts with equal probability. The grid depends on the whole window.
02
Discount, observe, normalize.
Multiply previous normalized log weights by 0.96, add a Gaussian observation likelihood, then normalize with log-sum-exp. The Gaussian scale is 3.5% of the grid range.
03
Read probability honestly.
Entropy measures spread in bits. The equal-tailed 95% interval covers grid probability, not 95% of future observations. Concentration is 1 − H / H₀, not accuracy.
When an update becomes a revision.
The first four observations initialize the model. Subsequent updates enter the log when KL(current || previous) ≥ 0.1 nats. A revision can have positive or negative entropy change.
log wᵢ ← 0.96 · log wᵢ − ½((observation − hypothesisᵢ) / σ)² pᵢ ← exp(log wᵢ − logsumexp(log w)) KL ← Σ pᵢ · log(pᵢ / previousᵢ)
Every significant change has a record.
A revision is a measurable distribution change. It can move the mode, narrow the distribution, or widen it. Select an entry to inspect the corresponding observation.
Revisions in window
—
KL ≥ 0.1 nats · after initialization
Largest revision
—
nats · among recorded revisions
Average revision gap
—
observations · current window
Change is broader than uncertainty.
Each entry records the observed value, previous and updated modes, KL change in nats, and the signed entropy change in bits. Negative ΔH means the distribution narrowed; it can still represent a significant revision.
The latest 25 entries belong to the active input window, not an all-time archive. They are local model calculations, not blockchain events.
Data → model → inspectable results.
The entire inference path runs in your browser. Public providers supply observations; the local model supplies the distribution.
Observe
Deterministic preview
or public Solana closes
Update
256 hypotheses
0.96 evidence retention
Inspect
Distribution, interval,
revisions and export
Implemented capabilities
Probability engine & diagnosticsLOCAL · IMPLEMENTED
Default observation streamSIMULATED · DETERMINISTIC
Solana market adapterPUBLIC HISTORICAL DATA · OPTIONAL
Wallet connectionPUBLIC ADDRESS ONLY
On-chain program / token / tradingNOT IMPLEMENTED OR DEPLOYED
Source failures stay visible.
The adapter selects a Solana pool by reported liquidity, then requests one-minute historical closes. A failed request keeps the last available snapshot and shows UNAVAILABLE. No synthetic history is substituted for a failed public source.
There is no private backend, transaction signer or private-key storage. The network mark identifies the adapter’s network; it does not imply Solana Foundation affiliation.
The exact model behind the charts.
The code below is imported directly from the running module. Copy it or reproduce an exported window with the same parameters.
MODEL 2.0.0BINS 256RETENTION 0.96REVISION KL ≥ 0.1 nats
Input → output
const result = infer([1, 1.02, 0.99, 1.04, 1.01], {
bins: 256,
retention: 0.96
});
// result.posterior: normalized Float64Array(256)
// result.klHistory: information change in nats per observation
// result.revisions: significant changes after initialization
// infer() returns null for fewer than 5 positive finite inputsSnapshot contract
Schema version 2 includes model parameters, the exact revision rule, source and fetch metadata, chronological observations, KL and entropy histories, and the revision entries. Observation indices are one-based in exports; the model’s t field is zero-based.
The infer(input, {bins, retention}) signature and veris:update browser event are unchanged. Consumers read revisions.
Evidence deserves a visible path.
An independent research interface for examining how observations change a model. The probability creature represents an evolving distribution; its gold core represents incoming evidence.
VERIS / ORIGINAL IDENTITY
Keep uncertainty in the picture.
The charcoal lobes hold multiple possibilities around one evidence core. The character is separate from the official Solana network mark.
Download transparent character ↗
Principles
Show the input, the update rule and the resulting distribution together. Distinguish simulated observations, public market history and unavailable sources. Record significant change whether entropy rises or falls.
Implemented
Eight linked views, source configuration, a local probability model, observation inspection, significant-revision history, descriptive diagnostics, versioned exports and optional wallet address connection.
Current boundaries
No trading, token issuance, on-chain program or predictive guarantee. The model reprocesses a rolling window, so its grid and local revision history can change when the window changes.
On your device
The chosen public address and FIELD preference stay in browser storage when available. Preview mode makes no market requests. Configured mode sends the public token address to DexScreener and GeckoTerminal. Exports download locally.