Report outcomes without reporting user data

After an operation, report whether it succeeded and which implementations were called. The dashboard uses many installations’ reports to compare code that should behave the same.

Report one execution

use polyform_runtime::CallTelemetry;

let calls = vec![
    CallTelemetry::new(
        "decode_filename",
        selected,
        if result.is_ok() { "ok" } else { "error" },
    ),
];

client.report_execution(
    "extract_archive",
    result.is_ok(),
    elapsed_ms,
    result.as_ref().err().map(|_| "invalid_filename"),
    &calls,
)?;

Send only what comparison needs

• A random ID for this installation.
• The current list version and implementation IDs being used.
• A short operation name, such as extract_archive.
• Success or failure, duration, and a small error category.
• The specified functions called during that operation and their outcomes.

Never send user content

• No filenames, file contents, database records, prompts, or user documents.
• No full error messages that may contain private values.
• No names, emails, IP addresses, or device fingerprints.

What happens when you quarantine an implementation

1. The dashboard compares its failure rate with other implementations of the same function.
2. A high relative risk tells you where to investigate; it is not proof by itself.
3. Open the linked public source and review the suspicious implementation.
4. Choose Quarantine to stop distributing it to new installations immediately.
5. Existing installations receive another already-shipped implementation when they next call refresh.
6. A new stage appears on the health graph so you can see whether successful executions improve after the change.

Restore makes the implementation available for distribution again. Earlier reports stay attached to the exact code combinations that produced them.