📘 mova-twin — prospective cost predictions, measured runs, online calibration

mova-twin predicts a workload's wall time from structural features, runs the real command, records the prospective error and updates its coefficients by recursive least squares (ridge around an explicit prior). It models cost only — not answers, correctness or learning. The model and its update equations: Training Mova.

Run

cargo build --release -p mova-twin
# copy tools/twin/example.json and fill in absolute paths
target/release/mova-twin predict SPEC.json [PARENT_RUN]
target/release/mova-twin run SPEC.json NEW_RUN [PARENT_RUN]
target/release/mova-twin report RUN...

• predict prints the forecast (seconds, observation count, whether the features lie outside the observed box) from the prior or from a parent run's updated state.
• run creates NEW_RUN (it must be new), copies and hashes the executable, runs it with the spec's arguments, environment and timeout (no shell), checks the declared outputs, and writes forecast.json, observation.json, stdout.txt, stderr.txt and the updated state.
• report prints one JSON row per run: features, forecast, actual seconds, absolute relative error, accepted, errors; it refuses a run whose evidence hashes changed.

The spec (schema: 1) names the family, declared context (hardware, revision, warm/cold), executable, cwd, args, env, the env axes allowed to vary in a cohort, inputs, feature_source, features, prior, ridge, timeout and output checks.

Limits

No forgetting factor and no transfer across changed binaries, data or hardware. The covariance is inverse information, not a calibrated interval. Feature choice is an explicit hypothesis of each experiment.

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/tools/twin/