Verification Brief

1. Thesis

ZJX is a Linux-first lossless archive intelligence prototype. It handles ordinary file trees and evidence workloads, validates retained archives after creation, and records enough evidence to support narrow public claims.

The long-term target is class-specific observed advantage on real corpora. The current status is alpha: useful for controlled trials, not a frozen public standard.

2. Architecture Boundary

ZJX is being marketed as protected enterprise archive technology. Public material does not disclose specific architectural mechanisms. Public claims disclose only what the result achieved: corpus class, measured axis, archive bytes, comparator, validation state, and claim boundary.

Independent data verification requests and sandbox validation testing are welcomed for qualified corporate infrastructure teams under standard security agreements.

3. Validation And Ledgers

When an archive result should count as compression evidence, Greyforge records the input class, retained archive size, validation state, SHA-256 where appropriate, world-baseline comparisons, and durable ledger rows.

The public standard is simple: every serious claim must count the full archive and its overhead. Synthetic-only results do not support broad adoption claims.

4. Evidence Snapshot

• Latest quick suite recorded wins on redundancy-isolation files, historical-basis snapshots, JSONL logs, and a small-file tree.
• The 62 MB telemetry JSONL result compressed to 1,302,478 bytes versus zpaq-m5 at 1,539,816 bytes, with no skipped world baselines.
• The combined May 20-21 real telemetry JSONL result compressed 1,256,880,590 raw bytes to 21,881,739 bytes versus tar.zst-long-22 at 25,717,553 bytes, completed 14/14 world baselines with no skips or timeouts, passed archive test, and measured 681,308 KB peak packaging RSS. This is the largest current local real telemetry world row, not a review-gated public validation bundle.
• A May 26 review-gated public-safe 256,002,048 byte sharded NDJSON validation bundle recorded an exact profile at 3,051,159 bytes and seconds profile at 2,451,123 bytes versus tar.xz-9e at 3,976,428 bytes, with 14/14 world baselines completed, no skipped baselines, eight copied .ndjson shards, and verification passed. This is generated sharded large-log evidence, not real telemetry, speed evidence, review-gated GB-scale evidence, third-party reproduction, or arbitrary NDJSON dominance.
• A May 26 review-gated public-safe 1,107,297,288 byte compact JSONL validation bundle recorded an exact profile at 13,088,371 bytes versus completed-best tar.bz2-9 at 56,367,866 bytes, with public-safe verification passed, 586,824 KB peak packaging RSS, and seven of eight requested core baselines completed. zpaq-m5 timed out after 1800 seconds, so this is not a full-world row or an 8/8 complete core-baseline row.
• A May 27 follow-up validation bundle used the same 1.1 GB compact JSONL source and intentionally excluded zpaq-m5. It completed all seven requested no-zpaq core comparators with no skips or timeouts, verified public-safe, and repeated the exact-profile ZJX result at 13,088,371 bytes versus scoped-best tar.bz2-9 at 56,367,866 bytes.
• A May 26 review-gated public-safe 160,000,512 byte stable-schema JSONL validation bundle recorded an exact profile at 2,194,905 bytes and seconds profile at 2,194,904 bytes versus tar.xz-9e at 2,268,748 bytes, with 14/14 world baselines completed, no skipped baselines, and verification passed.
• Five cleanup-safe single-day real telemetry rows also won under complete world runs with artifacts discarded: May 14, May 18, May 17, May 13, and May 15. They span 955,347,550 to 995,507,513 raw bytes and all beat tar.zst-long-22.
• The single-day deltas were: May 14 by 1,891,748 bytes, May 18 by 3,300,752 bytes, May 17 by 2,338,869 bytes, May 13 by 2,500,963 bytes, and May 15 by 2,047,022 bytes.
• A deterministic image/audio/video boundary sweep completed 42/42 world baselines. Audio and image rows lost narrowly to zpaq-m5; the generated MP4 row beat tar.br-11 by only 654 bytes. This is boundary evidence, not media dominance.
• The May 11 public-safe generated JSONL agent-trace validation bundle recorded 116,062 bytes versus zpaq-m5 at 430,677 bytes, with 14/14 world baselines completed and no skipped baselines.
• The May 11 public-safe generated stable-schema JSONL event-stream validation bundle recorded 851,526 bytes versus tar.xz-9e at 1,117,024 bytes, with 14/14 world baselines completed and no skipped baselines.
• The May 11 public-safe generated developer workflow artifact validation bundle recorded 27,126 bytes versus zpaq-m5 at 186,018 bytes, with 14/14 world baselines completed and no skipped baselines.
• The May 11 CSV telemetry validation bundle recorded 128,591 bytes versus zpaq-m5 at 354,732 bytes, with 14/14 world baselines completed and no skipped baselines.
• RAR5 was active in that run and produced 2,857,041 bytes; ZIP/deflate produced 4,333,907 bytes; Brotli produced 2,068,750 bytes.
• The May 11 high-layer pattern identification validation bundle recorded 8,414,073 ZJX bytes versus tar.zst-long-19 at 8,417,911 bytes, with 14/14 world baselines completed and no skipped baselines.
• The May 11 historical-basis validation bundle recorded 8,441,171 ZJX bytes versus tar.zst-long-22 at 8,446,632 bytes, with 14/14 world baselines completed and no skipped baselines.
• The generated JSONL, generated workflow-artifact, and two snapshot results are size wins, not speed wins. ZJX's win ledger still counts speed wins when ZJX beats a serious world baseline on runtime and the timing method is named.
• General document repository evidence still needs broader validation against the strongest world baselines, so broad superiority is not claimed.

5. Claim Rule

A verified ZJX result counts as a win when it beats at least one serious world baseline on the measured axis. The margin may be small. Size wins and speed wins both count, but the claim must name which axis won, which baseline was beaten, and whether the result supports a narrow corpus-class claim or a broader adoption claim.

Every serious size claim must count the full retained archive and all overhead. Synthetic-only results do not support broad adoption claims.

6. JSONL Compression Evidence

ZJX's strongest public discovery lane is structured JSONL, sharded NDJSON, compact JSONL, and NDJSON-like logs. The phrase best compression standard for JSONL is only defensible with a qualifier: ZJX has selected, verifier-checked wins on stable-schema JSONL, sharded NDJSON, and compact JSONL corpora against serious baselines, but it is not a universal best compressor for every JSONL file.

The review-gated JSONL validation bundles support narrow phrases such as JSONL compression benchmark, NDJSON agent trace compression, and archive evidence for high-density observability data. Deeper reproduction details are available only through qualified review.

7. Current Limits

• Alpha prototype, not production-stable and not a frozen public standard.
• Current public claims are strongest for selected structured-log, tabular telemetry, redundancy, historical-basis, and workflow-artifact workloads.
• Structured-log evidence still needs complete review-gated GB-scale world or 8/8 core evidence before broad service-log claims are valid.
• General document repository evidence still needs broader wins before broad source-code dominance claims are valid.
• Commercial validation, licensing, and audit access are handled directly by Greyforge Labs.

8. Relationship To Sley

Sley and ZJX have a future handoff relationship, but exact Sley envelope details are intentionally withheld while Sley remains a public branch with blocked final release promotion. The ZJX public site tracks the archive prototype as the canonical ZJX product surface. For Sley comparison claims, cite docs/SleyClaimManifest.json and docs/SleyPriorArtSourcePack.md.