@bycelium/evidences-diagrams

Combine evidence, measure dissonance, and explore how prior beliefs change the result.

Open the live demo · Download the standalone demo

Add evidence

Change the prior

Adjust evidence weights

The demo supports English and French, with editable categories and illustrative weights. The downloaded HTML works offline.

Demo walkthrough

No evidence and the color legend

With no evidence, the result is none (0 dBan). The prior band is visible and the probability remains indeterminate. The color legend places dissonance at the top, evidence against at the left, evidence for at the right, and no evidence at the center.

Adding and combining evidence

Choose a category, set its supporting or opposing weight, and add it to the history. Editing a category control updates its existing evidence entries of the same category and direction. The history supports selection and removal.

Weak evidence can remain gray when the posterior interval crosses 50%.

Evidence for	Evidence against
+10 +10 +10 = +30 dBan	−10 −10 −10 = −30 dBan

These walkthrough examples use adjusted weights. The resultant starts at the conservative prior endpoint and ends at the posterior odds used for the displayed probability.

Conflicting evidence and prior bounds

The ribbon retains the cumulative path, so opposing evidence remains visible.

Yellow identifies dissonance: equal and opposing weights cancel in the net weight while remaining in the absolute weight.

Prior interval	Exact prior

Move the prior handles or use the keyboard to change each bound. Equal bounds form an exact prior and expose one reusable handle.

Model

For independent signed evidence weights wᵢ, the package calculates:

• S = Σ wᵢ, the net weight.
• A = Σ |wᵢ|, the absolute weight.
• D = 1 − |S| / A, the dissonance, with D = 0 when A = 0.

For example, +9 dBan and −1 dBan produce S = +8 dBan, A = 10 dBan, and D = 20%. Decibans are log odds: +10 dBan multiplies the odds by ten. For probability p:

d = 10 × log10(p / (1 − p))
p = 1 / (1 + 10^(−d / 10))

A prior interval [min, max] is expressed in decibans. Net evidence translates it to [min + S, max + S], then each endpoint is converted to probability. The conservative probability uses the lower endpoint for positive evidence and the upper endpoint for negative evidence. Equal bounds produce an exact probability.

Install

pnpm add @bycelium/evidences-diagrams

import {
  aggregateEvidence,
  buildEvidenceDiagram,
  boundProbability,
} from '@bycelium/evidences-diagrams';

const evidence = [
  { id: 'report', weightDecibans: 12 },
  { id: 'counter-analysis', weightDecibans: -5 },
];

const summary = aggregateEvidence(evidence);
summary.netWeightDecibans; // 7
summary.absoluteWeightDecibans; // 17
summary.dissonance; // 10 / 17

const diagram = buildEvidenceDiagram(evidence);
diagram.steps[1].end; // { netWeightDecibans: 7, absoluteWeightDecibans: 17 }
diagram.resultant.jeffreysBand; // 'moderate'
diagram.probability.text; // 'Probability ≥ 13 %'

boundProbability(summary).direction; // 'lower'

pnpm install
pnpm build:demo
pnpm check

Apache-2.0. See LICENSE and NOTICE.