Speed: jubarte, docxodus, SuperDoc, docxide, LibreOffice

Every tool gets the same planned inputs, one tool at a time on one machine (Apple M4 Pro, 12 cores, 24 GB, macOS). Times are per call. Nothing produced is kept: each redline or PDF is deleted once it is timed. The raw per-call rows are in the bench repository under results/speed_10k/.

Redline speed, 10,000 planned pairs

One call per pair, engines run one after another on the same machine. Node lanes time an in-memory compare(base, next); jubarte-rust spawns the CLI per pair (process start and file I/O included); SuperDoc's SDK is file based (open, compare, apply, save). No redline is kept. Base size quartile cuts: [10601, 19102, 56197] bytes. Machine state: ENV.txt. A call past the per-pair timeout (120 s) is a failure; its row keeps the timeout as its time.

Coverage

docxodus: 10000 of 10000 planned pairs
jubarte-rust: 10000 of 10000 planned pairs
jubarte-rust-inproc: 10000 of 10000 planned pairs
jubarte-wasm: 10000 of 10000 planned pairs
superdoc: 10000 of 10000 planned pairs

Same pairs for every tool

tool	pairs	n	ok	failed	median ms	mean ms	p95 ms	p99 ms	max ms	ok total s
docxodus	10000 common	10000	9933	67	79.1	1731.7	7197.7	40675.7	116815.3	17200.6
jubarte-rust	10000 common	10000	10000	0	45.0	83.0	246.8	671.7	3818.1	830.1
jubarte-rust-inproc	10000 common	10000	10000	0	38.9	74.3	234.6	621.0	3818.8	743.3
jubarte-wasm	10000 common	10000	10000	0	70.9	140.6	461.3	1202.8	5859.9	1406.1
superdoc	10000 common	10000	798	9202	75.7	418.9	1623.1	6350.0	31746.8	334.3

By pair category

tool	category	n	ok	failed	median ms	mean ms	p95 ms	p99 ms	max ms	ok total s
docxodus	all	10000	9933	67	79.1	1731.7	7197.7	40675.7	116815.3	17200.6
docxodus	grid	5961	5928	33	116.1	2064.5	8934.3	45458.4	116815.3	12238.5
docxodus	identity	300	300	0	0.1	0.3	1.0	5.9	7.5	0.1
docxodus	reverse	1000	990	10	52.4	1317.8	4990.0	45081.4	81329.7	1304.6
docxodus	word_compare	2739	2715	24	45.7	1347.1	4643.2	35400.5	109684.8	3657.4
jubarte-rust	all	10000	10000	0	45.0	83.0	246.8	671.7	3818.1	830.1
jubarte-rust	grid	5961	5961	0	55.7	101.3	313.5	783.3	3028.3	604.0
jubarte-rust	identity	300	300	0	5.2	9.0	27.2	59.8	97.7	2.7
jubarte-rust	reverse	1000	1000	0	34.5	57.6	172.5	466.6	2546.2	57.6
jubarte-rust	word_compare	2739	2739	0	35.4	60.5	169.4	457.2	3818.1	165.8
jubarte-rust-inproc	all	10000	10000	0	38.9	74.3	234.6	621.0	3818.8	743.3
jubarte-rust-inproc	grid	5961	5961	0	48.5	90.3	296.1	695.7	2992.7	538.3
jubarte-rust-inproc	identity	300	300	0	2.4	6.1	22.7	54.4	119.1	1.8
jubarte-rust-inproc	reverse	1000	1000	0	28.5	50.9	157.8	419.1	2405.7	50.9
jubarte-rust-inproc	word_compare	2739	2739	0	30.2	55.6	164.1	468.3	3818.8	152.3
jubarte-wasm	all	10000	10000	0	70.9	140.6	461.3	1202.8	5859.9	1406.1
jubarte-wasm	grid	5961	5961	0	89.2	172.7	569.7	1494.8	5305.6	1029.5
jubarte-wasm	identity	300	300	0	4.4	18.6	76.5	294.7	474.3	5.6
jubarte-wasm	reverse	1000	1000	0	56.2	99.6	308.1	759.8	5146.3	99.6
jubarte-wasm	word_compare	2739	2739	0	54.0	99.1	295.9	793.4	5859.9	271.4
superdoc	all	10000	798	9202	75.7	418.9	1623.1	6350.0	31746.8	334.3
superdoc	grid	5961	16	5945	93.4	336.2	3585.5	3585.5	3585.5	5.4
superdoc	identity	300	265	35	343.2	1095.9	3403.1	12359.0	31746.8	290.4
superdoc	reverse	1000	144	856	65.6	76.4	129.4	208.3	228.7	11.0
superdoc	word_compare	2739	373	2366	59.0	73.7	156.1	219.3	291.8	27.5

By base size quartile

tool	quartile	n	ok	failed	median ms	mean ms	p95 ms	p99 ms	max ms	ok total s
docxodus	Q1	2499	2499	0	21.4	216.5	435.9	4267.3	38258.4	541.1
docxodus	Q2	2497	2495	2	45.5	471.0	1168.3	11683.2	77293.0	1175.0
docxodus	Q3	2502	2485	17	117.9	2842.6	13489.2	65597.5	116815.3	7063.9
docxodus	Q4	2502	2454	48	230.7	3431.4	19517.9	64970.6	115067.5	8420.6
jubarte-rust	Q1	2499	2499	0	11.2	36.8	121.1	342.3	3028.3	92.0
jubarte-rust	Q2	2497	2497	0	26.5	48.2	136.9	467.1	2176.4	120.3
jubarte-rust	Q3	2502	2502	0	59.4	87.6	208.2	599.4	2407.2	219.2
jubarte-rust	Q4	2502	2502	0	113.4	159.3	450.1	1063.5	3818.1	398.6
jubarte-rust-inproc	Q1	2499	2499	0	7.4	30.5	109.0	359.4	2016.5	76.2
jubarte-rust-inproc	Q2	2497	2497	0	21.7	41.6	125.3	419.1	1955.6	104.0
jubarte-rust-inproc	Q3	2502	2502	0	52.4	78.9	188.1	552.5	2992.7	197.4
jubarte-rust-inproc	Q4	2502	2502	0	104.4	146.2	411.6	965.3	3818.8	365.7
jubarte-wasm	Q1	2499	2499	0	12.1	58.7	211.5	615.2	3810.8	146.7
jubarte-wasm	Q2	2497	2497	0	38.9	80.1	229.5	883.2	3453.3	200.0
jubarte-wasm	Q3	2502	2502	0	96.6	154.3	405.1	1182.5	4454.4	386.0
jubarte-wasm	Q4	2502	2502	0	185.1	269.2	762.0	1974.9	5859.9	673.4
superdoc	Q1	2499	585	1914	63.3	86.3	173.5	282.7	3585.5	50.5
superdoc	Q2	2497	85	2412	236.1	285.4	685.0	1512.0	1512.0	24.3
superdoc	Q3	2502	73	2429	696.1	1212.9	5847.2	7270.1	7270.1	88.5
superdoc	Q4	2502	55	2447	1623.1	3109.2	12359.0	31746.8	31746.8	171.0

Redline speed run, notes

Run of 2026-09-29 (00:45 to 09:23 EDT)

jubarte-rust-inproc, jubarte-rust (CLI) and jubarte-wasm ran all 10,000 planned pairs with no failures (jubarte-redlines 0.9.3 @673aff74, same commit for all three lanes).
docxodus (npm 12.6.4, .NET WASM) started at 01:35 and had finished 1,335 pairs (p00000 to p01334, all word_compare) when the run was stopped at 09:23. The harness then had no per-pair timeout and wrote per-pair rows only at the end of a lane, so nothing had been written.
It was not hung. Single pairs took up to 7,162 s (p01320), 3,212 s (p01235) and 2,242 s (p01122). 21 of the 1,335 pairs took more than 120 s, and together they account for about 18,600 of the 21,700 s the lane had used. Pair p01335 (base 491 KB, next 232 KB) had been running for more than an hour when the run was stopped.
The 1,335 timings were read from the live process through the Node inspector (SIGUSR1, then Runtime.queryObjects over live arrays to find the lane's samples, sampleKeys and outSizes). The keys are contiguous from p00000 to p01334, so no pair in that range failed.
per_pair/docxodus.jsonl holds them with "source": "salvaged". The 120 s per-pair timeout now used by every lane is applied to them after the fact: the 21 slower pairs are rows with ok: false, timeout: true, ms: 120000 and their real time in measured_ms.
SuperDoc and the rest of docxodus (from p01335, with the per-pair timeout) are run separately.

Load

The machine was shared: two 1-ce-5 Python processes at 100% CPU for the whole run and bursts of another session's Rust builds (see ENV.txt). The lanes ran one after another, so each saw the same background.

Speed on 10,000 varied inputs