authority is not reputation
Level playing field
Admin and editor authority are outside the rating model. Authority contributes no mathematical advantage:
- admin bonus
- 0
- editor bonus
- 0
- authority bonus
- 0
Operational power, editorial power, contribution reputation, and curator trust are separate concepts.
source of truth
Raw ratings
One user may hold one rating per comment, per dimension. Updating a rating replaces it; it does not stack another vote. Raw ratings are the source of truth. Every derived aggregate can be deleted and rebuilt.
comment dimensions
What is measured
The active model uses a small public set of dimensions. Their contribution coefficients are part of the published model.
- Insightful
- Adds useful interpretation, reasoning, or non-obvious context. coefficient 1.00
- Informative
- Adds relevant facts, explanation, or concrete context. coefficient 0.80
- Well-sourced
- Supports claims with credible evidence or primary material. coefficient 0.80
- Combative
- Describes argumentative or confrontational style; not inherently bad. coefficient 0.00
comment and contribution score
Bayesian shrinkage
A comment with one 5-star rating should not automatically outrank a deeply sampled 4.2. Scores begin at the public site prior and move as evidence accumulates.
score = (C × m + Σ(weight × rating)) / (C + Σ weight)
- rating scale
- 1–5
- prior mean m
- 3.0
- prior strength C
- 15.0
anti-clique damping
Repeated relationships count less
The nth distinct comment one curator rates for the same author is damped. Multiple dimensions on the same comment share the same ordinal, so rating four dimensions does not pretend to be four independent relationships.
pair weight = 1 / sqrt(n)
First comment 1.000 · fourth 0.500 · ninth 0.333 · twenty-fifth 0.200
curator calibration
Judgment is measured
An earlier rating is compared with later independent consensus on the same comment and dimension. The curator's own rating is excluded. Trust weights are also excluded from this consensus, so the model cannot grade trusted curators using a consensus that their own trust already dominates.
At least 3 later independent raters are required before an earlier judgment becomes evaluable.
curator trust
Transparent trust formula
The raw curator-trust signal combines calibration, discrimination, and author diversity:
raw trust = 0.55 × calibration
+ 0.25 × discrimination
+ 0.20 × author diversity
The raw value is then shrunk toward neutral trust 0.5 according to the amount of evaluable evidence:
confidence = evaluated ratings / (evaluated ratings + 20.0) trust = 0.5 + confidence × (raw trust - 0.5)
Current rating influence is bounded:
weight = clip(2 * trust_score, 0.10, 2.00)
No curator can become infinitely powerful. A new confirmed curator begins at neutral trust 0.5 and weight 1.0.
public strata over continuous trust
Curators, trusted curators, elite, and outcasts
The mathematics remains continuous. The public strata are readable labels over the current trust score, and they are reversible when behavior changes.
| Trust score | Public tier |
|---|---|
| below 0.20 | outcast |
| below 0.40 | low-trust curator |
| below 0.65 | curator |
| below 0.82 | trusted curator |
| below 1.01 | elite curator |
Outcast means reduced influence on the signal system, not automatic loss of voice or an automatic ban.
important separation
Three independent concepts
- Contribution reputation
- How valuable are the user's comments, based on ratings received?
- Curator trust
- How well does the user's judgment survive later independent scrutiny?
- Authority
- What operational or editorial powers does the account have? This contributes zero mathematical bonus.
implementation note
Rebuildable, auditable, versioned
Model version 1.0 stores raw ratings separately from derived aggregates. Aggregates may be deleted and rebuilt from source ratings after a public model change.