PHAT v2 result data
These files support the current PHAT comparison, completed 2026-10-04 for phat-v2-2026-10-04. Its single scored population contains 48 medium/large inputs, with four families, four variants per family and three sizes. All requests completed correctly. No pilot or earlier-study measurements enter these results.
| File | Contents |
|---|---|
results.json | Scope, family and phase aggregates, verification counts, elapsed stages, host observations and memory |
observations.json | All 144 case/phase rows, including five paired process medians and available allocation/retention counters |
timings.csv | Compact version of all 144 rows |
provenance.json | Machine, versions, source/input/adapter identities and raw-evidence hashes |
family_ratios.svg | Individual combined-phase ratios and family aggregates with intervals |
graph_scaling.svg, simplicial_scaling.svg, cubical_scaling.svg, algebraic_scaling.svg | Three sizes of each of the sixteen variants, with both tools shown |
performance_profile.svg | Fraction of the fixed cases completed within a given factor of the faster tool |
SHA256SUMS | Digests of these files, excluding itself |
PHAT v2 units and row meanings
Times are seconds; memory and allocations are bytes. phase is construction, query or directly timed combined. cells, nonzeros and bars describe the complete input and output. size_level is the position in a variant's ladder, not a shared geometric dimension; parameter records its generator setting. mathematical_check states the independent evidence available for that input.
Each tool's case time is the median of five process medians, each based on three accepted fresh measurements after two warmups. tools retains those medians and available counters. phat_over_tamerop is the geometric mean of the five paired ratios; it can differ from dividing the displayed medians. Intervals are 95% t intervals on five process-block log ratios, with four degrees of freedom. Family and whole-suite intervals aggregate within paired blocks. All families, variants and sizes have equal declared weight at their respective level.
The frozen practical classification uses win above 1.10, loss below 1/1.10 and near_tie between, with ratios always PHAT/TamerOp. Near ties are described as similar performance in the report; these labels do not establish statistical equivalence or imply that a full interval lies in one region.
Julia allocation and GC counters describe the call. Retained input/answer sizes are separate roots. Process RSS includes runtime and untimed work and is sampled, so brief peaks can be missed. PHAT allocation and retained-object counters are null, not zero. Resource observations cover a shared desktop; system swap activity is not attributed to a particular tool.
PHAT v2 reuse and integrity
Run sha256sum -c SHA256SUMS in this directory to check file integrity. This does not rerun the mathematical validators. Public per-pass summaries allow the statistics and figures to be reconstructed, but do not recover every raw sample. Exact fixtures, source/build recipes, validators and complete raw records remain locally archived. These downloads are not an executable reproduction bundle. Private paths and host names are omitted from public provenance.