"qualification_result_policy":"record exact observed values in the generated run result, not the source tree, without requiring literal model or version equality; every supported component must satisfy the release matrix"
"median_definition":"sort numeric values ascending and select value at zero-based index floor(count/2); every measured count is odd",
"within_run_relative_mad_formula":"median(abs(sample - median(samples))) / median(samples), evaluated separately for direct elapsed times, routed elapsed times, and paired routed/direct ratios in each workload, preparation state, and independent run",
"between_run_median_spread_formula":"(max(run_medians) - min(run_medians)) / median(run_medians), evaluated across the three independent-run medians separately for direct elapsed time, routed elapsed time, and paired routed/direct ratio in each workload and preparation state",
"outlier_policy":"retain every completed measured sample",
"sample_timeout":"four times the workload absolute budget",
"within_run_relative_mad_limit":0.10,
"between_run_median_spread_limit":0.10,
"budget_tolerance":0.0
},
"preparation_states":{
"unprimed":"fresh uniquely named materialization, empty workload output directories, and no fixture prewalk; operating-system cache dropping is not required",
"primed":"fresh materialization followed by one untimed complete workload invocation before measured samples"
"environment":"reject a run that violates any qualification-environment constraint or direct-placement calibration limit; record every exact observed qualification-result field",
"background_load":"reject a sample when another process consumes more than five percent of one reference core or benchmark memory pressure causes reclaim",
"pass_rule":"every independent run and preparation state meets the direct-placement calibration, routed absolute median, paired median-ratio budget, and both variance limits"
"fixture":"100 directories each containing 100 regular files of 64 bytes; half host-backed and half ignored isolated",
"fixture_algorithm":"directories d000 through d099; files f000 through f099; content is the 32-byte SHA-256 of fixture_seed, slash, and byte path repeated twice; even directories are host-backed and odd directories are ignored isolated",
"operation":"bytewise walk, lstat, open, read 64 bytes, close, and statx each file",
"fixture_algorithm":"tracked/t0000 through t0999 are committed; ordinary/u0000 through u0499 are untracked; ignored/i0000 through i0499 match one root ignore rule and have distinct host sentinel and isolated content derived from fixture_seed and byte path",
"operation":"porcelain-v2 status, add the ordinary untracked set, quiet mixed reset, and porcelain-v2 status",
"result_oracle":"both status byte streams equal the locked direct placement streams, the index returns to its initial identity, and every ignored host sentinel is unchanged",
"fixture":"one root wheel and 31 deterministic local dependency wheels with no build scripts or network access",
"fixture_algorithm":"packages lazaret-bench-pkg-00 through lazaret-bench-pkg-31 at version 1.0.0; package 00 depends exactly on packages 01 through 31; each pure-Python wheel contains one module whose VALUE is the SHA-256 of fixture_seed and normalized package name",
"operation":"uv pip install --offline --no-index into a fresh ignored virtual environment",
"fixture":"one binary and 31 path-library crates with deterministic generated Rust sources and an exact lockfile",
"fixture_algorithm":"workspace packages lazaret-bench-crate-00 through lazaret-bench-crate-31 at version 1.0.0; package 00 is a binary with exact path dependencies on libraries 01 through 31; each library returns the first eight bytes of SHA-256 of fixture_seed and package name as a u64 constant",
"operation":"cargo check --offline --locked --workspace with a fresh ignored target directory",
