← Benchmarks · Unpublished 5.0.0-rc5

Smaller numeric logs.
Complete costs, exact bytes.

The requested 30% size threshold is exceeded on named real logs, including later source windows. The complete 30-source panel improves 5.81% Native / 5.07% Hybrid. Compression and decoding take more work. This is a representation improvement, not a new entropy coder or a universal win.

The full panel

Matched ordinary thin profiles against frozen pre-change sources. Every original archive is retained as an incumbent. All 80 Native/Hybrid cases across 40 inputs restore exact bytes and SHA-256; production models match independently prepared Python bytes and complete-cost predictions.

Matched profile	Old bytes	New bytes	Reduction	Improved / 30
native_thin	805,043	758,258	5.81%	12
hybrid_thin_L9	726,532	689,715	5.07%	10

Against the best previous tested settings

Each previous minimum includes Native thin, Hybrid thin, Hybrid thin with template factorization, and forced line sorting in both engines. The 22 additional sorting controls restore exact source bytes; none improves these previous minima. New minima evaluate the first three modes. This is a finite evaluated set, not an exhaustive option search. Six later, complete-line windows start after 1/2 MiB, beyond prior test/training prefixes. Model parameters were frozen first. These are disjoint bytes from three previously seen families, not six independent datasets.

Source	Previous best bytes	New best bytes	Reduction	Selection
03_loghub_spark.log	17,257	14,567	15.59%	Inspected development
04_loghub_zookeeper.log	7,160	5,457	23.78%	Inspected development
09_loghub_windows.log	15,015	9,642	35.78%	Inspected development
13_loghub_healthapp.log	22,724	18,083	20.42%	Inspected development
api_payload.json	9,536	9,536	0.00%	Inspected development
spark_1024k.log	22,865	18,254	20.17%	Later window
spark_2048k.log	22,292	17,926	19.59%	Later window
windows_1024k.log	12,578	7,350	41.56%	Later window
windows_2048k.log	13,936	7,562	45.74%	Later window
healthapp_1024k.log	26,346	16,547	37.19%	Later window
healthapp_2048k.log	26,643	17,766	33.32%	Later window

The time tradeoff

Actual subprocess wall time includes launch and file IO: two warmups plus three measured medians, alternating old/new order. Each number sums eleven source medians on this shared host. No fastest-codec claim follows from this study. All slower outcomes remain in the raw CLI trials.

Profile · 11 sources	Encode old → new ms	Encode change	Verified decode old → new ms	Decode change
native_thin	365.3 → 542.5	+48.5%	107.0 → 127.6	+19.3%
hybrid_thin_L9	394.8 → 592.6	+50.1%	102.1 → 137.9	+35.1%
hybrid_thin_L9_template	436.3 → 624.4	+43.1%	105.1 → 130.6	+24.2%

Stronger density references

All raw reference frames are independently decoded and SHA-verified. PIZD pays both its model and original-source integrity framing; raw reference formats keep their own framing. Q11/L19 are stronger settings, not matched effort. No size or timing result is copied from a different input window.

Source	PIZD best complete bytes	C Brotli Q11 raw	C Zstd L19 raw	Gzip L9
03_loghub_spark.log	14,567	15,901	17,717	25,641
04_loghub_zookeeper.log	5,457	12,297	13,857	18,423
09_loghub_windows.log	9,642	11,048	13,217	22,301
13_loghub_healthapp.log	18,083	29,172	32,089	42,764
api_payload.json	9,536	19,492	20,200	24,200
spark_1024k.log	18,254	23,989	25,880	36,899
spark_2048k.log	17,926	22,903	23,964	34,522
windows_1024k.log	7,350	9,567	10,857	19,477
windows_2048k.log	7,562	10,384	11,575	21,384
healthapp_1024k.log	16,547	30,795	33,354	50,189
healthapp_2048k.log	17,766	30,693	33,845	52,625

Peak memory samples

One whole-process macOS /usr/bin/time -l sample per engine/version on unmodified larger source prefixes. These are observations, not worst-case bounds or IoT-memory guarantees. Additional workspaces, numeric scalar arrays and skeletons cost memory. Decoder values use bounded cursors. Large-file streaming limits per-frame source size.

Source prefix	Engine	Version	Input bytes	Peak RSS MiB
Windows	native	old	3,145,728	22.9
Windows	native	new	3,145,728	30.0
Windows	hybrid	old	3,145,728	28.8
Windows	hybrid	new	3,145,728	41.7
HealthApp	native	old	3,145,728	24.5
HealthApp	native	new	3,145,728	30.0
HealthApp	hybrid	old	3,145,728	29.3
HealthApp	hybrid	new	3,145,728	46.1

Every production result

Unchanged cases include all tested API fixtures and the GitHub JSON windows: this numeric model does not solve their remaining density gap. Default/fast and dictionary/server paths retain their previous route. Rejected numeric models can still add thin encoding work.

Source	Backend	Old bytes	New bytes	Reduction	Container
01_loghub_hdfs.log	native_thin	43,756	43,756	0.00%	5
01_loghub_hdfs.log	hybrid_thin_L9	37,697	37,697	0.00%	5
02_loghub_hadoop.log	native_thin	17,013	14,244	✓ 16.28%	11
02_loghub_hadoop.log	hybrid_thin_L9	14,896	12,907	✓ 13.35%	11
03_loghub_spark.log	native_thin	22,979	16,481	✓ 28.28%	11
03_loghub_spark.log	hybrid_thin_L9	19,439	14,567	✓ 25.06%	11
04_loghub_zookeeper.log	native_thin	8,964	6,358	✓ 29.07%	11
04_loghub_zookeeper.log	hybrid_thin_L9	8,049	5,457	✓ 32.20%	11
05_loghub_openstack.log	native_thin	41,821	41,821	0.00%	5
05_loghub_openstack.log	hybrid_thin_L9	37,772	37,772	0.00%	5
06_loghub_bgl.log	native_thin	17,984	13,883	✓ 22.80%	11
06_loghub_bgl.log	hybrid_thin_L9	17,416	12,615	✓ 27.57%	11
07_loghub_hpc.log	native_thin	20,956	16,671	✓ 20.45%	11
07_loghub_hpc.log	hybrid_thin_L9	19,351	15,956	✓ 17.54%	11
08_loghub_thunderbird.log	native_thin	40,422	39,508	✓ 2.26%	11
08_loghub_thunderbird.log	hybrid_thin_L9	35,294	35,166	✓ 0.36%	11
09_loghub_windows.log	native_thin	17,470	10,889	✓ 37.67%	11
09_loghub_windows.log	hybrid_thin_L9	15,015	9,642	✓ 35.78%	11
10_loghub_linux.log	native_thin	31,830	31,830	0.00%	7
10_loghub_linux.log	hybrid_thin_L9	26,071	26,071	0.00%	5
11_loghub_mac.log	native_thin	22,240	22,240	0.00%	7
11_loghub_mac.log	hybrid_thin_L9	20,156	20,156	0.00%	5
12_loghub_android.log	native_thin	40,960	40,960	0.00%	7
12_loghub_android.log	hybrid_thin_L9	38,974	38,974	0.00%	7
13_loghub_healthapp.log	native_thin	34,560	20,526	✓ 40.61%	11
13_loghub_healthapp.log	hybrid_thin_L9	29,399	18,083	✓ 38.49%	11
14_loghub_apache.log	native_thin	25,470	25,470	0.00%	7
14_loghub_apache.log	hybrid_thin_L9	23,973	23,973	0.00%	7
15_loghub_openssh.log	native_thin	16,327	13,401	✓ 17.92%	11
15_loghub_openssh.log	hybrid_thin_L9	14,084	11,885	✓ 15.61%	11
16_loghub_proxifier.log	native_thin	25,976	24,743	✓ 4.75%	11
16_loghub_proxifier.log	hybrid_thin_L9	22,943	22,943	0.00%	5
17_beats_nginx_access.log	native_thin	774	774	0.00%	5
17_beats_nginx_access.log	hybrid_thin_L9	644	644	0.00%	5
18_beats_nginx_error.log	native_thin	663	663	0.00%	5
18_beats_nginx_error.log	hybrid_thin_L9	561	561	0.00%	5
19_beats_postgresql_log.log	native_thin	2,978	2,538	✓ 14.78%	11
19_beats_postgresql_log.log	hybrid_thin_L9	2,264	2,264	0.00%	5
20_beats_mysql_slowlog.log	native_thin	2,883	2,485	✓ 13.81%	11
20_beats_mysql_slowlog.log	hybrid_thin_L9	2,365	2,213	✓ 6.43%	11
21_beats_mongodb_log.log	native_thin	1,784	1,784	0.00%	5
21_beats_mongodb_log.log	hybrid_thin_L9	1,519	1,519	0.00%	5
22_beats_redis_log.log	native_thin	1,505	1,505	0.00%	7
22_beats_redis_log.log	hybrid_thin_L9	1,163	1,163	0.00%	7
23_beats_elasticsearch_audit.log	native_thin	1,633	1,633	0.00%	5
23_beats_elasticsearch_audit.log	hybrid_thin_L9	1,336	1,336	0.00%	5
24_beats_traefik_access.log	native_thin	1,058	1,058	0.00%	5
24_beats_traefik_access.log	hybrid_thin_L9	926	926	0.00%	5
25_nasa_http_jul95.log	native_thin	36,151	36,151	0.00%	7
25_nasa_http_jul95.log	hybrid_thin_L9	35,057	35,057	0.00%	7
26_nasa_http_aug95.log	native_thin	37,283	37,283	0.00%	7
26_nasa_http_aug95.log	hybrid_thin_L9	35,956	35,956	0.00%	7
27_github_events_20150101_00.jsonl	native_thin	76,935	76,935	0.00%	5
27_github_events_20150101_00.jsonl	hybrid_thin_L9	69,507	69,507	0.00%	5
28_github_events_20150101_06.jsonl	native_thin	71,978	71,978	0.00%	5
28_github_events_20150101_06.jsonl	hybrid_thin_L9	65,775	65,775	0.00%	5
29_github_events_20150101_12.jsonl	native_thin	68,120	68,120	0.00%	7
29_github_events_20150101_12.jsonl	hybrid_thin_L9	62,753	62,753	0.00%	5
30_github_events_20150101_18.jsonl	native_thin	72,570	72,570	0.00%	7
30_github_events_20150101_18.jsonl	hybrid_thin_L9	66,177	66,177	0.00%	5
api_payload.json	native_thin	9,944	9,944	0.00%	7
api_payload.json	hybrid_thin_L9	9,536	9,536	0.00%	7
api_batch1.json	native_thin	5,623	5,623	0.00%	7
api_batch1.json	hybrid_thin_L9	5,374	5,374	0.00%	7
10_rest_api_batch.json	native_thin	5,762	5,762	0.00%	7
10_rest_api_batch.json	hybrid_thin_L9	5,473	5,473	0.00%	7
blank_line.jsonl	native_thin	211	211	0.00%	5
blank_line.jsonl	hybrid_thin_L9	118	118	0.00%	5
escaped.jsonl	native_thin	211	211	0.00%	5
escaped.jsonl	hybrid_thin_L9	117	117	0.00%	5
long_string.jsonl	native_thin	149	149	0.00%	5
long_string.jsonl	hybrid_thin_L9	118	118	0.00%	5
no_final_lf.jsonl	native_thin	210	210	0.00%	5
no_final_lf.jsonl	hybrid_thin_L9	116	116	0.00%	5
schema_change.jsonl	native_thin	212	212	0.00%	5
schema_change.jsonl	hybrid_thin_L9	115	115	0.00%	5
uniform.jsonl	native_thin	211	211	0.00%	5
uniform.jsonl	hybrid_thin_L9	117	117	0.00%	5
variable_separator.json	native_thin	154	154	0.00%	5
variable_separator.json	hybrid_thin_L9	129	129	0.00%	5

Compatibility and evidence

New numeric containers are version 10 (SHA-256) or 11 (BLAKE3); older readers reject them explicitly. Existing versions 1–9 remain readable. Nested numeric chains, external patterns, malformed dimensions and output expansion beyond the declared budget are rejected. The full suite passes 283 tests across 58 targets; this is not an exhaustive security audit.

Smaller numeric logs.Complete costs, exact bytes.