ZJX 1.25 NIGHTMARKET: The Fastest Compressor Is Not Always the Best Compressor
Understand how ZJX 1.25 NIGHTMARKET turned compression into an explicit customer choice between a size-first default and a bounded fast schedule, why the quickest schedule was rejected, why performance authority moved to a dedicated laboratory host, and why the first release was withdrawn.
Understand how ZJX 1.25 NIGHTMARKET turned compression into an explicit customer choice between a size-first default and a bounded fast schedule, why the quickest schedule was rejected, why performance authority moved to a dedicated laboratory host, and why the first release was withdrawn.
- Edition
- Release qualification record 1.0
- Published
- Sep 2, 2026
- Preview updated
- Sep 2, 2026
- Reading time
- 12 minutes
- Full edition
- 2,100 words
- Status
- current
Who this is for
- Compression and storage engineers deciding how much packing effort a customer should be allowed to trade away.
- Release engineers designing frozen qualification contracts that guard downstream operations, not only headline speed.
- Technical leads who need a defensible measurement law for sub-second operations and contention-free performance evidence.
Not for
- Readers seeking ZJX source, candidate schedules, codec allowlists, private corpora, host topology, or a universal speed claim.
Detailed contents
- 01
Why NIGHTMARKET
Two prior experiments that optimized proxy metrics and did not ship.
- 02
Size first, not size only
The five costs of an archive and why retained bytes stay the flagship.
- 03
COMPACT and FAST
The default size-first search and the explicit guardrail contract for the reduced schedule.
- 04
The fast mode we refused to ship
The 1.83x test and 1.61x unpack result against a 1.25 limit, and the measurement-law repair it forced.
- 05
Clean lab
Why development-host timings were superseded and how a quiesced laboratory host became release authority.
- 06
The final campaign
Four corpora, six pairs, 48 starts, per-corpus fractions, zero size penalty, 48 of 48 exact restores.
- 07
The release we withdrew
A contradictory machine-readable status, a one-string repair, and a fresh exact-byte requalification.
- 08
Permanent doctrine
What NIGHTMARKET changes for every future ZJX experiment.
Named artifacts
Fast objective guardrail table
Wall, CPU, size-penalty, auxiliary, and memory limits that define the fast contract.
Final campaign results table
Equal-family medians, per-corpus archive bytes and wall fractions, exact restores, and start ledger.
Rejected-schedule record
Why the fastest schedule failed the lifecycle guardrail and how the replacement was promoted.
Status-repair release record
The withdrawn v1.25, the v1.25.1 repair, and the fresh qualification cycle.
Size first, not size only
An archive has at least five costs: the final archive bytes, the packing time, the packing CPU, the peak memory, and the cost of every later operation that consumes the archive: test, inspect, unpack.
A compressor that reports only its ratio is hiding four of them. A compressor that reports only its speed is hiding four different ones. NIGHTMARKET treats the set as a frontier. The flagship objective is retained bytes, because customers pay for storage every month and pay for packing once.
Packing time is tradeable, but only inside an explicit contract, and only when the downstream operations stay cheap, because customers consume archives far more often than they create them. Under NIGHTMARKET's frozen four-corpus clean-lab campaign, fast packed in an equal-family median of 0.3637 of compact's wall time with zero complete-archive size penalty and auxiliary costs within one percent of compact.
The public source note remains free and independently indexable.
Public release boundary
- Default
- compact runs the canonical size-first search and keeps the smallest complete archive.
- Fast
- A fixed reduced schedule qualified under explicit wall, CPU, size, auxiliary, and memory guardrails.
- Rejected
- The schedule that packed quickest exceeded the 1.25 test and unpack limit and did not ship.
- Full edition
- The rejected-schedule analysis, measurement-law repair, clean-lab doctrine, and withdrawal record require an unlock; bounded final measurements remain public.
Evidence and method
measured: Frozen objective surface and thresholds, paired compact and fast arms with alternating order on a quiesced laboratory host, batched short-operation sampling, paired-ratio sign-test decisions, exact restoration checks, and release validation of the exact published bytes.
Sources
Limitations
- Campaign ratios are bounded to the frozen laboratory host, four corpora, six pairs per corpus, commands, and thresholds. They are not universal speed claims.
- Zero size penalty on four corpora does not make the compact search redundant, and fast carries a published size budget that other inputs may spend.
Access and updates
Purchase includes lifetime read access to this edition, email-based recovery, and revisions published to the same edition.
Public companion: Read the NIGHTMARKET public source note