Language evidence

Clear words. Measured quality.

Compare how plainly, warmly, and directly models write in German and English. Choose the lowest measured cost that clears your quality threshold.

Loading dated language evidence…

—models in the catalogue

—measured observations

—public studies

Public studies inform the method. A dash means no comparable measurement for this model, language, and thinking level. Provisional research records cannot establish a model ranking.

Model matrix · Cost and quality · Study inventory · Dated evidence

Model and language matrix

Six independent dimensions from 0 to 1; higher is better. Overall is their equally weighted mean. Thinking levels and languages are never pooled.

Language Thinking level

Latest row of the strongest evidence status for each model, language, and thinking level
Model	Language / effort	Readability	Warmth	Directness	No AI mannerisms	Tone adherence	Factual restraint	Overall	USD / sample	Evidence

What clears your quality threshold?

Sample costs come from evidence runs, not token list prices. Compare costs within the same language, effort, prompt set, and rating method. USD per quality point is sample cost divided by overall quality; zero quality has no ratio.

Minimum overall Include provisional measurements

Passing measurements, cheapest sample first
Model	Language / effort	Overall	USD / sample	USD / quality point	Status

Text routing priors

Copy, documentation, operator messages, and replies use the existing DocEdit task class. Their prior requires overall ≥ 0.75, readability and absence of AI mannerisms ≥ 0.75, and factual restraint ≥ 0.80 at medium effort, with observed evidence.

Language quality adds a constraint. Concrete-task correctness floors still apply.

Study inventory

Primary publications and one-line takeaways. Applying their methods to these models remains provisional; the papers do not supply this catalogue’s six scores.

Evidence, with dates

All retained records, including earlier observations. Study dates show source verification; imported benchmark timestamps use the result date at UTC midnight, with day precision.

Use the capability

var requirement = new HumanFriendlyLanguageRequirement
{
    Language = "de",
    MinimumOverall = 0.75m,
    ThinkingLevel = EffortLevel.Medium,
    MinimumScores = new() { FactualRestraint = 0.80m },
};
var candidates = ModelEfficiencyMatrix.Default.SuggestModel(
    TaskClass.DocEdit, BudgetPressure.Tight,
    [Cli.Codex, Cli.ClaudeCode], DateTime.UtcNow, requirement);

An empty list means no retained evidence clears the constraint. Voice Lint runs the study; Token Economy imports the shared result file and publishes the evidence.

Method and import contract · Typed API · Download JSON · Task studies