ZJX / Benchmark ledger
Every result, in its own row.
Every published ZJX campaign, including failed and invalid runs, separated by release and experiment so no figure borrows another run's context.
RAW STRUCTURED HISTORY342.4MB
VERIFIED ZJX ARCHIVE4.41MB
98.7% smaller
Ledger
Results by release and experiment.
CURRENT marks the newest qualified evidence on the current lineage. That includes the two REDLINE verified-creation campaigns on the released 1.29.0 package (two rows: they measured different cohorts) and the 2026-09-06 comparator campaign on the released 1.28.0 package, whose compact and fast objectives 1.29.0 keeps unchanged. PASS met its contract. SUPERSEDED marks a campaign replaced by a newer run of the same procedure on the same input, kept as lineage. FAIL is rejected or unreleased. INVALID could not answer its question and is never pooled with valid evidence.
| Release | Workload | Objective | ZJX result | Comparator / baseline | Advantage | Status | Evidence |
|---|---|---|---|---|---|---|---|
| 1.29.6.6 | JSON-lines and general inputs; One laboratory host: AMD Ryzen 3 7330U, 4 cores / 8 threads, Linux | Archive identity, creation and verification time, and peak memory against 1.29.6.5 | A reliability release for JSON-lines data; every other archive byte-identical to 1.29.6.5, with creation and verification times unchanged | ZJX 1.29.6.5 SIGHTLINE, the released package | Reliability release; no measured regression | CURRENT | Results since 1.29.0 |
| 1.29.6.5 SIGHTLINE | Already-compressed media of 1 MiB to 64 MiB; One laboratory host: AMD Ryzen 3 7330U, 4 cores / 8 threads, Linux | Verified creation, fast objective | 1.14x to 12.80x faster than 1.29.6; archive size within +1% of 1.29.6 | ZJX 1.29.6 IRONFLOOR and 1.29.0 REDLINE | The fast objective on already-compressed media of 1 MiB to 64 MiB; every other archive byte-identical to 1.29.6 | PASS | Results since 1.29.0 |
| 1.29.6 IRONFLOOR | 8 corpora, 6 CPU placements (48 cells) plus sealed and fresh holdout tiers; One laboratory host: AMD Ryzen 3 7330U, 4 cores / 8 threads, Linux | Verified creation, CPU, and peak memory, ultrafast objective | 1.74x faster verified creation than 1.29.0 (1.20x in the slowest cell); 1.50x less CPU; 3.27x lower peak memory | ZJX 1.29.0 REDLINE, the released package | Faster than 1.29.0 in every measured cell (48 primary cells plus sealed and fresh holdout tiers) with byte-identical archives | PASS | Results since 1.29.0 |
| 1.29.5 GRIDMIND | 8 corpora, 6 CPU placements (48 cells); One laboratory host: AMD Ryzen 3 7330U, 4 cores / 8 threads, Linux | Verified creation and CPU, ultrafast objective | 1.36x faster verified creation and 1.24x less CPU than 1.29.0 overall | ZJX 1.29.0 REDLINE, the released package | Overall gain; superseded by 1.29.6, which is faster in every cell | SUPERSEDED | Results since 1.29.0 |
| 1.29.0 REDLINE | Sealed development cohort: silesia, enwik8, and a synthetic deduplication case; 12 paired rounds, 468 of 468 cells usable; quiesced laboratory host; executable from the released package | Verified creation (create plus independent verify), ultrafast objective, four workers | 1.247x the speed of default tar | zstd -1; 95% [1.223, 1.267]; Bonferroni-adjusted lower bound 1.221; archive bytes 1.0005x; CPU about 1.19x | tar | zstd -1 (default). Against tar | zstd -1 -T4 the same configuration reads 0.820 (ZJX slower; 95% [0.807, 0.832]) | Faster than the default comparator; slower than its four-thread setting; development cohort | CURRENT | Source note |
| 1.29.0 REDLINE | Four sealed members plus six privately held-out cases (about 1.86 GB, six workload classes) selected by a rule frozen before any measurement; ten classes, 12 paired rounds, 360 of 360 cells usable, run once | Verified creation, ultrafast objective, four workers; preregistered gate P4 | 1.1959x default tar | zstd -1; 95% [1.1867, 1.2043]; 6 of 6 held-out cases restored exactly before the timed rounds | Gate against tar | zstd -1 -T4: class-weighted 0.8236, 95% bootstrap lower bound 0.8170 against a required 0.80; ZJX slower on 7 of 10 cases; the six held-out cases alone 0.7917 [0.7832, 0.8004] | Gate PASS; reported apart from the development figure | CURRENT | Source note |
| 1.29.0 REDLINE | Real service telemetry, 342,409,609 B JSONL (2026-06-03 capture); compact objective, unchanged from 1.28.0 (REDLINE gate P6: byte-identical archives); bytes written by the released 1.28.0 Linux package | Complete archive bytes, compact objective, versus 14 world archive baselines | 4,412,984 B; zjx test passed; restored file matched the source SHA-256 | Strongest of 14 completed baselines: zpaq -m5 at 7,125,674 B; every comparator archive round-tripped | 38.1% smaller than the strongest comparator | CURRENT | Source note |
| 1.29.0 REDLINE | Same telemetry capture, fast objective, unchanged from 1.28.0 (REDLINE gate P6: byte-identical archives) | Complete archive bytes, fast objective | 4,412,984 B; zjx test passed; restored file matched the source SHA-256 | Strongest of 14 completed baselines: zpaq -m5 | 38.1% smaller than the strongest comparator | CURRENT | Source note |
| 1.27 CIPHERGLASS | Same telemetry capture, previous build (2026-09-05 campaign, development host) | Complete archive bytes, compact and fast objectives, versus the same 14 world archive baselines | 4,412,968 B on both objectives; zjx test passed; restored file matched the source SHA-256 | Strongest of 14 completed baselines: zpaq -m5 at 7,125,674 B | 38.1% smaller than the strongest comparator; the 1.28.0 archive is 16 B larger (manifest-integrity footer) | SUPERSEDED | Source note |
| 1.26 RELICWIRE | Large single-file text corpus | Compact pack wall time and process CPU | About 28% lower wall time and CPU; 3/3 favourable paired comparisons | ZJX 1.25.1 compact, same host and corpus | Lower packing cost | PASS | Source note |
| 1.26 RELICWIRE | Large mixed public corpus | Compact pack wall time and process CPU | About 43% lower wall time and CPU; 3/3 favourable paired comparisons | ZJX 1.25.1 compact, same host and corpus | Lower packing cost | PASS | Source note |
| 1.26 RELICWIRE | Repaired second-boot reproduction, two large corpora | Archive identity and exact restoration | 6/6 byte-identical archive pairs; 12/12 exact restores; peak memory about 1.00x | Paired 1.25.1 reference and source-tree restoration checks | Exact measured output | PASS | Source note |
| 1.25.1 NIGHTMARKET | Four qualification corpora (mixed repo, Calgary, firmware snapshots, telemetry export) | Fast pack wall time | 0.3637 equal-family median wall fraction (~2.75x faster, ~63.6% less wall time) | Paired compact control, same host and corpus | ~2.75x relative speed | PASS | Source note |
| 1.25.1 NIGHTMARKET | 48-start qualification campaign | Exact restoration | 48/48 exact restores; size penalty 0/0/0 (median/p90/max) | Byte-identity contract | Zero measured size penalty | PASS | Source note |
| Pre-NIGHTMARKET telemetry | Local real telemetry, 342,409,609 B JSONL | Complete archive size | 4,412,968 B archive; 98.7% raw reduction | zpaq-m5 at 7,125,674 B (strongest of 14) | 38.1% smaller | PASS | Evidence note |
| Pre-NIGHTMARKET telemetry | Combined local telemetry, 1,256,880,590 B | Complete archive size | 21,881,739 B archive | Best completed at 25,717,553 B | Smaller retained archive | PASS | Evidence note |
| Pre-NIGHTMARKET generated | Sharded NDJSON, 256,002,048 B across eight shards | Complete archive size | 2,451,123 B archive | Strongest completed at 3,976,428 B | Smaller retained archive | PASS | Evidence note |
| 1.21 STREAMWALL | Silesia v1, 211,938,580 B, 12 files | Peak encode RSS under ceilings | 390,225,920 B observed peak RSS | 8,580,669,440 B historical reference | ~95.5% lower (~22x less), exact restore | PASS | Source note |
| 1.21 STREAMWALL | Deterministic scaling, 1 MiB through 3 GiB | Completion under enforced ceilings | 3 GiB input under 1 GiB host / 768 MiB ZJX ceilings | Enforced ceiling contract | ~30.1 MB observed encode RSS at 3 GiB | PASS | Source note |
| 1.23 BLACKWIRE | Seven-corpus native qualification, Campaign 3 | Native production boundary + archive identity | Exact candidate-control archive identity; clean-host lifecycle | Frozen native control | One qualified production wire | PASS | Source note |
| 1.22 NULLSPOOL | Bounded qualification campaign | End-to-end wall time and CPU | +2.4742% median wall time, +2.4714% CPU regression | Same-harness baseline | None: rejected, not shipped | FAIL | Chronicle |
| 1.24 GHOSTLANE | First materiality campaign | Fast-path treatment observation | All eight qualifying commands selected chunked storage; treatment never ran | Campaign validity contract | None: INVALID, retained separately | INVALID | Source note |
| 1.24 GHOSTLANE | Corrected campaign + repository-wide validation | Release qualification | 4/4 paired wins, 22.35% median improvement; then R12 terminal FAIL (2 failures, 7 errors) | Release contract | None: not released, version consumed | FAIL | Chronicle |
Historical releasePre-NIGHTMARKET benchmark generations