Accuracy & confidence

The platform's core methodological commitment is simple: no false precision, ever. A wide range with caveats is better than a sharp number that's wrong. So every result carries an accuracy score — a number from 0 to 100 — that any audience can read at a glance, and every input underneath it carries the three component scores it is built from.

This is the piece classical LCA usually treats as a bolt-on uncertainty analysis. Here it is a first-class, per-flow, propagated property of every number the platform produces.

Three components

Every input to an activity is scored on three independent components, each from 0 to 1. Each measures how well the input fits the activity it serves, not how good its data is in isolation:

Component	What it measures	Best (1.0)	Worst
Source	The standing of the organisation that published the input	Verified by the framework (1.0); peer-reviewed (0.85)	Neither (0.4)
Geography	How far the input's geography is from the activity's	The same place	A global average (0.1)
Temporal	How far the input's year is from the activity's year	The same year	More than ten years apart (0.1)

The framework fixes these at exactly three in the current version. Two further components — technological fit and measurement basis — are reserved and follow the same architecture, so they can be added later without changing the model.

The score is the mean of the three components. Where a component cannot be measured at all — no geography declared, no year, no producer behind the input — it scores 0.5: never zero, which would claim a mismatch nobody measured, and never one. Ignorance is shown, not hidden.

The score, not a grade

There are no letter grades. Bucketing a continuous score into five boxes invites people to compare the boxes rather than the numbers, at thresholds that are still provisional. The score is the whole of the claim, shown as a number from 0 to 100, and each component carries the reason it scored as it did — a parent geography, a five-year gap, an unreviewed source — so two inputs with the same number can be told apart.

How it propagates

Accuracy doesn't stop at a single input. An activity's accuracy is the mean over its inputs, and each input's score is blended with the score of the activity that produced it — so a well-matched input drawn from a badly sourced upstream cannot score as if that upstream were sound. The weakness follows the chain, and you can trace it to the input that introduced it. Every input counts the same, whatever its share of the carbon: the score says how much the model knows, not how much it matters.

Worked illustration for the bread example (see Core concepts):

Input	source	geography	temporal	score
Flour (European avg)	0.85	0.80 (one level up: BE → Europe)	0.90	0.85
Tap water (Belgium)	0.85	1.00 (exact)	0.90	0.92
Electricity (Belgium)	0.85	1.00 (exact)	1.00	0.95
Yeast (global avg)	0.85	0.10 (global)	0.70	0.55

The yeast pulls the average down because a global average is a poor geographic match for a Belgian bakery, and the reason behind its geography score says so:

(0.85 + 0.92 + 0.95 + 0.55) / 4  =  0.82  →  shown as 82

This is the whole point. A missing or generic input never disappears behind a clean number; it shows up as a lower score and a named weakness the user can decide to improve.

Why substitution isn't failure

When the platform can't find an exact-match supplier, it substitutes the closest one it has — the European flour activity for a Belgian bakery, a global yeast average when no regional one exists. That substitution is tracked and scored, never silent. The geographic distance (Belgium → Europe → global) directly sets the geography component.

The result: a model built from imperfect data is still honest. It tells you not just what the number is but how much to trust it, and exactly where the trust runs thin.

How accuracy improves over time

Accuracy is not fixed at import. It rises as the data behind a result gets better:

Better producers supersede weaker ones. The same thing (say, rice) can be produced by many activities — a coarse global import, then a regional dataset, then a peer-reviewed, location-specific one. A consumer wired to "rice" resolves to the best available producer, and the score climbs.
Review raises the source component. Scores are computed, never entered. What review changes is the standing of the organisation whose data it is — peer-reviewed, or verified by the framework — and every input it publishes inherits that standing. See Community review.
Data ages. The temporal component measures the gap between an input's year and the activity's year, so an old input scores lower the further the activities that use it move on, rather than quietly presenting itself as current.