PP-258 local CSV CLI comparison: earlier hands-on check record, 2026-10-01 UTC Historical record. This check ran on October 1, 2026, before Practical Picks adopted its research-only method on October 3, 2026. It is kept as a dated observation of the versions named below, not as current testing. The comparison was re-decided from public evidence on October 19, 2026. Environment: Linux x86-64. Miller 6.22.0, qsv 23.0.1, csvkit 2.2.0, DuckDB CLI 1.5.6. Official GitHub release archives for Miller, qsv and DuckDB were SHA-256-checked against release API digests. csvkit was installed from PyPI in a task-local Python virtual environment. No account or payment. Valid fixture: valid.csv, SHA-256 cc10e91bb04bf16af374e39194c7c576467bf102de8e6d2c1024a7dbee61d3c9 Malformed fixture: malformed.csv, SHA-256 7cba156ca944a4b8168024b99d3b63b59094d2ed49eba3282bfa3e06b5d190d8 The only difference is a two-field data row before the last normal row. Expected task: keep amount >= 8, group by team, sum amount. North = 20; South = 13. Preserve quoted comma, doubled quote and UTF-8 values when selecting owner/note fields. Reject malformed input before handoff. Miller: mlr --icsv check valid.csv exit 0 mlr --icsv check malformed.csv exit 1; row-5 length mismatch mlr --icsv --ocsv filter '$amount >= 8' then stats1 -a sum -f amount -g team valid.csv exit 0; team,amount_sum / North,20 / South,13 mlr --icsv --ocsv filter '$amount >= 8' then cut -f owner,note valid.csv exit 0; tested quoted and accented values retained Direct transformation of malformed.csv exited 1 but had already sent an earlier result to stdout. Do not hand off redirected stdout on failure. csvkit: csvclean --length-mismatch malformed.csv exit 1; short row reported csvsql --query "SELECT team, SUM(amount) AS total FROM valid WHERE amount >= 8 GROUP BY team ORDER BY team" valid.csv exit 0; North,20.0 and South,13.0 csvcut -c owner,note valid.csv exit 0; values retained The analogous csvsql query on malformed.csv returned exit 0 and the same filtered totals; it did not itself reject the short row. csvclean 2.2.0 requires an explicit check flag and has no --dry-run. These defaults began in csvkit 2.0.0, per its May 1, 2024 changelog. DuckDB CLI: duckdb -csv -c "SELECT team, SUM(CAST(amount AS INTEGER)) AS total FROM read_csv('valid.csv', header=true, columns={'team':'VARCHAR','owner':'VARCHAR','amount':'VARCHAR','note':'VARCHAR'}) WHERE CAST(amount AS INTEGER) >= 8 GROUP BY team ORDER BY team" exit 0; North,20 and South,13 read_csv('malformed.csv', all_varchar=true) without explicit columns returned exit 0 with one whole-line column, not the intended four. Explicit four-column schema on malformed.csv exited 1 with sniffer error. qsv: qsv validate valid.csv exit 0; 4 columns/4 records qsv validate malformed.csv exit 1; record 4, line 5, byte 108 qsv sqlp valid.csv 'SELECT team, SUM(amount) AS total FROM valid WHERE amount >= 8 GROUP BY team ORDER BY team' exit 0; North,20 and South,13 qsv sqlp valid.csv 'SELECT owner,note FROM valid WHERE amount >= 8 ORDER BY owner' exit 0; tested quoted and accented values retained qsv sqlp malformed.csv 'SELECT * FROM malformed' exited 0 and padded the short row with two empty cells. Validation must happen before SQLp for this task. Do not use SQLp success as a CSV check. Safe handoff exercise: a local shell wrapper refused to overwrite an existing destination, preflighted with Miller check or qsv validate, wrote a .part file, and renamed only after command success. Both tools saved the expected valid totals. Both exited 1 for malformed input and left no destination or .part file. Input checksums remained unchanged. The same-directory rename was tested; power-loss durability, concurrent changes, larger files and cross-platform behavior were not. Exact wrapper used for the staged handoff, also retained at research/PP-258/fixture/handoff-check.sh. Run it as `bash handoff-check.sh valid.csv result.csv miller` or replace `miller` with `qsv`. Put the verified executable on PATH, or set PP258_MILLER_BIN or PP258_QSV_BIN to its path. The output and its .part file must not exist before the run. Use a separate copy of the fixture when replaying. #!/usr/bin/env bash set -euo pipefail input=$1 output=$2 tool=$3 scratch="${output}.part" qsv_bin=${PP258_QSV_BIN:-qsv} miller_bin=${PP258_MILLER_BIN:-mlr} if [[ -e "$output" || -e "$scratch" ]]; then printf 'Refusing to overwrite an existing output or scratch file.\n' >&2 exit 3 fi trap 'rm -f -- "$scratch"' EXIT case "$tool" in qsv) "$qsv_bin" validate "$input" >&2 "$qsv_bin" sqlp "$input" \ 'SELECT team, SUM(amount) AS total FROM _t_1 WHERE amount >= 8 GROUP BY team ORDER BY team' \ --output "$scratch" >&2 ;; miller) "$miller_bin" --icsv check "$input" >&2 "$miller_bin" --icsv --ocsv \ filter '$amount >= 8' then stats1 -a sum -f amount -g team then sort -f team \ "$input" > "$scratch" ;; *) printf 'Unknown tool: %s\n' "$tool" >&2; exit 2 ;; esac mv -- "$scratch" "$output" This wrapper prevents a failed command from becoming the named handoff file in our single-process test. It does not make validation and query atomic: the input could change between those commands. It also does not guarantee a durable rename after power loss or prevent another process from creating the destination during the run. Archive SHA-256: Miller 6.22.0 Linux amd64: 0551f8696e8f42dae1c67c782bbe04e0589e718268722fa98561e2566e4623b8 qsv 23.0.1 Linux x86-64: df3a52004d3f3d7aaacec9326749967eaf2aa60cb8552253123b8487a0c6796c DuckDB CLI 1.5.6 Linux amd64: 6e89deac1ebbc36eed0291caf8b567b030c7b86ac35998f71854e22b3c5d5e2f This is one small synthetic fixture, not a speed, scale, security, privacy, reliability, general correctness or real-world data-quality study.