Visualization

The schema-2 reporting artifact is AnalysisReport. RetroCast keeps plotting downstream from evaluation: the core produces stable metric data, then Python, Rust, a notebook, or another application chooses how to render it.

For managed CLI workflows, start with the generated report.md. Use the structured report when you need custom figures or cross-model comparisons.

Prepare Plot Data

The same metric names and strata are available from both library interfaces.

Python 0.8.xRust 0.8.xPython 0.7.1

metric = "solv_0[buyables]_rate"
strata = sorted(report["by_stratum"])
summaries = [report["by_stratum"][name][metric] for name in strata]

values = [summary["value"] for summary in summaries]
ci_low = [summary.get("ci_low") for summary in summaries]
ci_high = [summary.get("ci_high") for summary in summaries]

let metric = "solv_0[buyables]_rate";
let plot_rows = report
    .by_stratum
    .iter()
    .filter_map(|(stratum, metrics)| {
        metrics.get(metric).map(|summary| {
            (
                stratum.clone(),
                summary.value,
                summary.ci_low,
                summary.ci_high,
            )
        })
    })
    .collect::<Vec<_>>();

metric = "solv_0[buyables]_rate"
strata = sorted(report.by_stratum)
summaries = [report.by_stratum[name][metric] for name in strata]

values = [summary.value for summary in summaries]
ci_low = [summary.ci_low for summary in summaries]
ci_high = [summary.ci_high for summary in summaries]

Plot With Python

Plotly is not bundled with RetroCast. Install it in the environment that renders the figure:

pip install plotly

import plotly.graph_objects as go

fig = go.Figure()
fig.add_trace(
    go.Bar(
        x=strata,
        y=values,
        error_y={
            "type": "data",
            "symmetric": False,
            "array": [
                hi - value if hi is not None else 0
                for hi, value in zip(ci_high, values, strict=True)
            ],
            "arrayminus": [
                value - lo if lo is not None else 0
                for value, lo in zip(values, ci_low, strict=True)
            ],
        },
        name="Solv-0[buyables]",
    )
)
fig.show()

Plot With Rust

AnalysisReport is a serializable Rust type. Feed plot_rows into the plotting or application framework that owns your presentation layer. For example, a service can return the complete report as JSON:

let json = serde_json::to_string_pretty(&report)?;
std::fs::write("analysis.json", json)?;

RetroCast does not require a Rust plotting crate and does not impose a chart style on consumers.

Compare Reports

Use the exact serialized metric key when comparing models. Align reports by benchmark, stock label, match level, and bootstrap settings before plotting them together.

Python 0.8.xRust 0.8.xPython 0.7.1

metric = "solv_0[buyables]_rate"
comparison = {
    name: model_report["metrics"][metric]["value"]
    for name, model_report in reports.items()
}

let comparison = reports
    .iter()
    .map(|(name, report)| (name, report.metrics[metric].value))
    .collect::<Vec<_>>();

metric = "solv_0[buyables]_rate"
comparison = {
    name: model_report.metrics[metric].value
    for name, model_report in reports.items()
}

Confidence intervals communicate uncertainty within a model. Use paired target-level analysis when making inferential claims about the difference between two models.