REMORSEKnow whether a solver change is worth shipping.
Compare candidates on the same instances. Inspect quality, reported validity and runtime before promoting a change.
1. Load evaluation results
Bring paired baseline and candidate measurements. CSV stays in this browser.
CSV format and comparison rules
Use the eight columns in the downloaded template. Keep one unique instance_id per paired result. Set split to development, holdout or stress.
baseline_objective and candidate_objective must be finite numbers in the same scoring units. Select Minimize when smaller is better, or Maximize when larger is better. baseline_ms and candidate_ms are runtimes in milliseconds, zero or greater. Use the same measurement environment.
baseline_valid and candidate_valid accept true/false or 1/0. These flags are your reported validity; Evaluate does not verify them. Use up to 2,000 data rows and 1 MiB per CSV.
Positive objective change means improvement. Aggregate objective change sums comparable pairs, where both validity flags are true. Its percentage uses the sum of absolute baseline objectives as the denominator. JSON stores relative values as fractions: 0.1 means 10%. The regression input uses percentage units: 10 means 10%. Runtime is a candidate/baseline ratio; 2 means twice the baseline runtime. A zero baseline runtime has no ratio and allows only a zero candidate runtime.
Meeting selected rules requires holdout and stress results. Reported invalid flags, regressions beyond your allowance or runtime violations block the decision. Reports are for human review.
2. Set promotion rules
3. Review the decision
Load a CSV to begin.
Compare paired baseline and candidate results.
No data loaded.
Uses the measurements and validity flags in your file. Evaluate does not execute or independently validate your solver. Summaries include pairs with both validity flags true. Positive objective change means improvement. A pass qualifies the candidate for human review.