QPA v1 result data

These files support the QPA comparison, completed 2026-10-03, for candidate qpa-v1-529681519b5d1b40. They report the frozen 220-request study, including its 20 unresolved QPA tensor comparisons. They do not contain the full executable benchmark suite.

Files and units

Times are seconds; allocations and exported-answer sizes are bytes. Process RSS in the summary is MiB (2²⁰ bytes). Runtime allocation counters are not retained memory, and Julia's counter does not include native FLINT allocations. Exported JSON byte counts measure an output file's size, not the memory occupied by the mathematical answer.

Row meanings

case is the fixed request identifier. workflow names Q1–Q10 as described in the report; variant distinguishes requests such as projective/injective resolution and pushout/pullback. area is one of the five weighted mathematical areas. field is QQ, F2, F3, or F101. split records development or the reserved evaluation subset; that subset was evaluated after the source freeze and is no longer unseen.

The phase values are:

• native_construction: build native inputs from the prepared category and decoded input.
• complete_query: produce the complete requested answer from native inputs.
• construction_and_query: directly time both stages together; the primary comparison.

In observations.json, each tool records five pass_medians_seconds and five sample_counts. A complete pass median uses three independently reset samples. median_seconds is the median of those five process medians, also exposed as TamerOp_seconds or QPA_seconds in the CSV. Allocation columns summarize the accepted per-timer counters; they are not allocation totals for a whole pass.

ratio is the geometric mean of the five paired QPA/TamerOp ratios. Consequently it need not equal the quotient of the two displayed median times. interval95 (CSV: ci_low, ci_high) is the approximate paired-block t interval of the log ratio, with four degrees of freedom. pass_range and pass_ratios retain variation that the point estimate does not show.

weight is the case's exact rational share of the original 220-case suite; it repeats across phases and must not be counted three times. For an aggregate, select a phase, use mathematically matched complete pairs, then normalize the retained weights within that subset. The primary subset has 200 cases and original weight 9/10. classification uses a 10% practical margin: win above 1.10, loss below 1/1.10, and near_tie between them. A point classification does not imply that its interval or every pass has the same classification.

JSON null and empty CSV fields mean unavailable, not zero. The 20 tensor cases remain present. Their QPA query and combined timings are missing. Native construction can have a time even when the requested final answer failed; mathematically_matched and tool-level verified flags prevent treating that as a completed comparison. The public phase aggregates consistently use the same 200 matched requests.

input_stalk_sum sums dimensions of the supplied vector spaces at poset vertices, counting each supplied module record once for that request. poset_vertices counts vertices. coefficient_max_bits describes the largest numerator/denominator bit length in the supplied module structure matrices; it is not a bound on intermediate arithmetic. Together these describe inputs, not a single universal complexity measure.

Integrity and reuse

From this directory on a system with sha256sum, run:

sha256sum -c SHA256SUMS

This checks the downloaded files, not the mathematical computations. The source archive and dependency hashes identify the private retained evidence without claiming that the archive is available for download here. No absolute local paths, host names, or unrelated project data are included.

The five medians per case are sufficient to recompute the reported speed aggregates and intervals using the procedure in the comparison page. They cannot recover the individual raw samples or independently rerun the answer validators. A future portable reproduction bundle must supply those materials and be identified separately from this results publication.