Roadmap Paper Docs Demo

Roadmap

CAFE is open-source and under active development. This page tracks where we are taking it — the capabilities we are building next across statistics, judging, experimental design, and the platform.

Statistics & attribution

Deepening the part that makes CAFE more than a leaderboard: honest, scale-correct variance attribution.

Planned

Order-independent deviance partition

A fair credit-split across factors that works for the ordinal model and for unbalanced or fractional designs — exactly where a single-model partial η² is weakest.

Exploring

Hierarchical rater modeling

Treat each human rater as a random effect, then nest raters within their group or organization — revealing how different expert pools systematically score the same system, and separating that rater-side variation from the factor effect you are actually measuring.

Exploring

Bayesian mixed models

Full posterior effect sizes and credible intervals, not just point estimates and p-values.

Exploring

Prediction-powered inference

Use a handful of human labels to debias the LLM judge's effect estimates, with statistical guarantees.

Planned

Power & sample-size advice

Turn "not significant" into "underpowered — add N questions to detect a gap this size."

Judging

More ways to turn answers into trustworthy scores.

Planned

Pairwise / comparative judging

"Is A better than B?" per item (Bradley–Terry / Elo), alongside the existing absolute-rubric mode.

Planned

Multi-criteria scoring

Several named scores per answer — e.g. correctness, faithfulness, tone — each with its own attribution.

Exploring

Multi-judge ensembles

Run a panel of judges and aggregate their verdicts to reduce single-judge bias.

Exploring

Judge-bias audits

Surface position, verbosity, and self-preference checks as a first-class result, not just a reliability number.

Designs & efficiency

Making large factorial studies affordable — and even smarter.

Planned

Computation reuse & caching

Memoize shared sub-paths across configurations so a factorial only recomputes what changed — the lever that makes big studies financially survivable.

Exploring

Split-plot & advanced designs

Cost-aware designs that exploit hard-to-change versus cheap-to-change factors, plus D-optimal and response-surface methods for continuous knobs.

Exploring

Optimize mode

Bayesian optimization and bandits to reach the best configuration fastest — complementing the "which factor matters?" spine rather than replacing it.

Platform & workflow

Lowering the barrier from "write Python" to "click and read the verdict."

Planned

"Does capability X actually help?"

An add/remove-a-component wizard that returns a one-line verdict: "web search: +0.4, d = 0.6, p = 0.03 — worth it."

Planned

System-vs-system comparison

A/B two structurally different pipelines head-to-head, not just levels within one system.

Planned

One-click report export

The effects table, forest plot, and comparison table as paper-ready LaTeX and figures.

Exploring

Bundled datasets & trace ingestion

Ship public benchmarks out of the box, and evaluate an untouched production system straight from its OpenTelemetry traces.