8Z Compression
Evidence-first lossless compression researchRaziskave brezizgubnega stiskanja, vodene z dokazi

Smaller archives.
No missing bytes.Manjši arhivi.
Brez izgubljenih bajtov.

8zip is not one magic codec. It is a verified arena: build several serious candidates, account for every byte, reject anything that cannot restore the exact source tree, and keep the smallest valid result. The compatible standard-ZIP branch and the experimental 8zipLab branch are deliberately separate.

8zip ni en čudežni kodek. Je verificirana arena: zgradi več resnih kandidatov, obračunaj vsak bajt, zavrni vse, kar ne obnovi natančnega izvornega drevesa, in obdrži najmanjši pravilen rezultat. Združljiva veja standardnega ZIP-a in eksperimentalna veja 8zipLab sta namenoma ločeni.

See the measured resultsPoglej izmerjene rezultate Open the 8Z hubOdpri 8Z središče
Exact restore requiredZahtevana natančna obnova Standard ZIP branchVeja standardnega ZIP-a Experimental FPLB branchEksperimentalna veja FPLB Live snapshot · 19 July 2026Živi posnetek · 19. julij 2026
19/24
ReZIP items completed; item 20 is runningReZIP arhivov končanih; arhiv 20 trenutno teče
Latest supplied log snapshotZadnji priloženi dnevnik
14.45%
saved on completed ReZIP archive 19prihranjeno pri končanem ReZIP arhivu 19
73,107,031 B → 62,541,850 B
0.325%
v0.2.1 FPLB gain over its standard-ZIP baselinedobiček FPLB v0.2.1 proti standardnemu ZIP izhodišču
Small, real, fully chargedMajhen, resničen, v celoti obračunan
9.07%
smaller v0.2.2 migration artifact than v0.2.1manjši migracijski artefakt v0.2.2 od v0.2.1
Exact-dedup migration proof, not the new live finalDokaz migracije z exact-dedup, ne novi živi finale
Two separate branchesDve ločeni veji

Compatibility here.
Research freedom there.Združljivost tukaj.
Raziskovalna svoboda tam.

The separation prevents a good experimental result from being mistaken for an ordinary ZIP result. It also lets the mainline stay useful today while the laboratory explores deeper structures for tomorrow.

Ločitev prepreči, da bi dober eksperimentalni rezultat zamenjali za navaden ZIP rezultat. Hkrati omogoča, da je glavna veja uporabna že danes, laboratorij pa raziskuje globlje strukture za jutri.

8zip + ReZIP

The standard-ZIP mainlineGlavna veja standardnega ZIP-a

The output remains a normal .zip archive using Store(0) and Deflate(8). Ordinary unzip tools restore it. Internally, 8zip may compare fresh 7‑Zip, ECT, KZIP, imported streams, and its own per-file candidates.

Izhod ostane običajen arhiv .zip z metodama Store(0) in Deflate(8). Obnovijo ga navadna unzip orodja. Znotraj arene lahko 8zip primerja sveži 7‑Zip, ECT, KZIP, uvožene tokove in lastne kandidate po datotekah.

  • Standard ZIP compatibility is a hard requirement.Združljivost s standardnim ZIP-om je trda zahteva.
  • The smallest candidate wins only after strict tree and byte verification.Najmanjši kandidat zmaga šele po strogi preverbi drevesa in bajtov.
  • ReZIP optimizes existing archives for cold storage without changing their meaning.ReZIP optimizira obstoječe arhive za hladno hrambo, ne da spremeni njihov pomen.
8zipLab

The experimental FPLB laboratoryEksperimentalni laboratorij FPLB

8zipLab is allowed to use its own FPLB container, embedded trained or procedural dictionaries, content-addressed blobs, and research-only restoration logic. It is not presented as standard ZIP.

8zipLab sme uporabljati lasten vsebnik FPLB, vgrajene naučene ali proceduralne slovarje, vsebinsko naslovljene bloke ter raziskovalno obnovo. Ni predstavljen kot standardni ZIP.

  • All model, dictionary, manifest, and container bytes are charged.Vsi bajti modela, slovarja, manifesta in vsebnika so obračunani.
  • Train, validation, and sealed holdout are separated by exact-content group.Train, validation in sealed holdout so ločeni po skupinah enake vsebine.
  • FPLB is not standard ZIP. A custom-format win does not imply universal dictionary generalization; CROSS_ARCHIVE_REUSE remains unproven without independent external corpora.FPLB ni standardni ZIP. Zmaga lastnega formata še ne pomeni univerzalne posplošitve slovarja; CROSS_ARCHIVE_REUSE ostaja nedokazan brez neodvisnih zunanjih korpusov.
The shared disciplineSkupna disciplina

Do not guess the winner. Build the arena.Ne ugibaj zmagovalca. Zgradi areno.

This is the MDL×DCC idea in practical form: multiple candidate descriptions compete, full cost decides, and verification is part of the score rather than an afterthought.

To je praktična oblika ideje MDL×DCC: več kandidatnih opisov tekmuje, odloča celoten strošek, verifikacija pa je del rezultata in ne naknadna misel.

1 · Inventory1 · InventarPaths, sizes, hashes, empty directoriesPoti, velikosti, hashi, prazne mape
2 · Candidates2 · KandidatiCodecs, streams, dictionaries, dedup layoutsKodeki, tokovi, slovarji, dedup postavitve
3 · Full MDL cost3 · Celoten MDL strošekPayload + metadata + models + containerPayload + metadata + modeli + vsebnik
4 · Restore4 · ObnovaExact paths and byte-for-byte contentsNatančne poti in bajtno enaka vsebina
5 · Promote5 · PromocijaSmallest verified no-regression championNajmanjši verificirani prvak brez regresije
Measured evidenceIzmerjeni dokazi

What we know so farKaj za zdaj vemo

The encouraging part is not a slogan. It is that several different mechanisms have now produced verified gains: stream selection in standard ZIP, embedded dictionaries in FPLB, and exact-content deduplication in FPLB3.

Spodbuden del ni slogan. Več različnih mehanizmov je že prineslo verificirane prihranke: izbor tokov v standardnem ZIP-u, vgrajeni slovarji v FPLB in odstranjevanje natančnih dvojnikov v FPLB3.

8zipLab v0.2.1 · verified

A small but real dictionary resultMajhen, a resničen rezultat slovarjev

On a 320-file, 293.59 MB source corpus, the final artifact restored every file exactly and was deterministically rebuilt byte-identically.

Na korpusu 320 datotek z 293,59 MB izvornih podatkov je končni artefakt natančno obnovil vsako datoteko in bil deterministično ponovno zgrajen bajtno identično.

ArtifactArtefaktBytesBajtiChangeSprememba
Standard ZIP baselineStandardni ZIP baseline38,805,947—
Dictionaryless FPLB2FPLB2 brez slovarjev38,772,676−33,271
Final FPLB2 + dictionariesKončni FPLB2 + slovarji38,679,805−126,142 vs ZIP

The dictionary layer added 92,871 bytes of net gain after all model and manifest costs. The strongest real-data contribution came from a train-derived 8 KiB dictionary.

Slovarski sloj je po vseh stroških modela in manifesta dodal 92.871 bajtov neto prihranka. Največji prispevek na realnih podatkih je dal iz train podatkov naučen 8 KiB slovar.

v0.2.2 · migration proof

Exact duplicates were the larger opportunityNatančni dvojniki so bili večja priložnost

The same v0.2.1 corpus contained 320 paths but only 286 unique contents. A production FPLB3 migration stored each unique content once and kept all original paths.

Isti korpus v0.2.1 je imel 320 poti, vendar le 286 različnih vsebin. Produkcijska migracija FPLB3 je vsako vsebino shranila enkrat in ohranila vse izvorne poti.

FPLB2 v0.2.1
38,679,805 B
FPLB3 v0.2.2
35,172,243 B
  • 3,507,562 bytes smaller than the v0.2.1 champion.3.507.562 bajtov manj od prvaka v0.2.1.
  • 9.07% gain versus FPLB2 v0.2.1.9,07 % prihranka proti FPLB2 v0.2.1.
  • 9.36% gain versus the original standard-ZIP baseline.9,36 % prihranka proti prvotnemu standardnemu ZIP baselineu.

This is a verified migration result on the known corpus, not the final result of the new live v0.2.2 search.

To je verificiran migracijski rezultat na znanem korpusu, ne končni rezultat novega živega iskanja v0.2.2.

ReZIP · archive 19

A practical standard-ZIP winPraktična zmaga standardnega ZIP-a

Original ZIPIzvirni ZIP73,107,031 B
Verified ReZIP outputVerificirani ReZIP izhod62,541,850 B
SavedPrihranek10,565,181 B · 14.45%
WinnerZmagovalecinternal mixed MDL + imported streams
VerificationVerifikacijaPASS + independent reverify
Interpretation

Why both small and large gains matterZakaj so pomembni majhni in veliki prihranki

A 0.3% gain on a difficult corpus can validate a new mechanism. A 9–15% gain on a suitable archive can already matter operationally. The right goal is not “beat everything everywhere”; it is find the best valid description for this data and never regress when a simpler method is better.

0,3 % prihranka na težkem korpusu lahko potrdi nov mehanizem. 9–15 % na primernem arhivu je lahko že praktično pomembnih. Pravi cilj ni »premagati vse povsod«, temveč najti najboljši pravilen opis teh podatkov in ne regresirati, kadar je preprostejša metoda boljša.

Research lessonsRaziskovalne lekcije

The project is becoming clearer, not smallerProjekt postaja jasnejši, ne manjši

Each run removes one kind of uncertainty. That is progress: we now know which gains are container effects, dictionary effects, exact-duplicate effects, and which claims still need external corpora.

Vsak run odstrani eno vrsto negotovosti. To je napredek: zdaj vemo, kateri prihranki izvirajo iz vsebnika, slovarjev, natančnih dvojnikov in katere trditve še zahtevajo zunanje korpuse.

Verification is part of compressionVerifikacija je del stiskanja

A smaller archive that cannot prove exact restoration is not a winner. The verifier, tree hash, corruption tests, and deterministic rebuild are part of the product.

Manjši arhiv, ki ne dokaže natančne obnove, ni zmagovalec. Verifier, hash drevesa, testi poškodb in deterministična ponovna gradnja so del izdelka.

Split by content, not only by filenameDelitev po vsebini, ne le po imenu

v0.2.1 revealed that identical content under different paths can weaken holdout independence. v0.2.2 keeps exact duplicates inside one split.

v0.2.1 je pokazala, da ista vsebina pod različnimi potmi oslabi neodvisnost holdouta. v0.2.2 drži natančne dvojnike v istem splitu.

One learned dictionary survived stricter testsEn naučeni slovar je preživel strožje teste

The reusable signal became smaller after duplicate-safe analysis, but it did not disappear. That makes the next cross-archive test worth doing.

Signal ponovne uporabe se je po duplicate-safe analizi zmanjšal, ni pa izginil. Zato je naslednji cross-archive test smiseln.

Exact dedup changed the scaleExact dedup je spremenil merilo

On this corpus, storing identical content once produced far more gain than refining dictionary recipes alone. The best roadmap follows measured leverage.

Na tem korpusu je enkratno shranjevanje enake vsebine prineslo mnogo več kot samo izboljševanje slovarskih receptov. Najboljši načrt sledi izmerjenemu vzvodu.

Live campaign snapshotPosnetek živih runov

Both runs are healthyOba runa sta zdrava

Snapshot from the latest project-folder logs supplied on 19 July 2026. These values will be updated when the campaigns advance.

Posnetek iz najnovejših dnevnikov projektne mape, priloženih 19. julija 2026. Vrednosti bomo posodabljali, ko bosta kampanji napredovali.

8zipLab v0.2.2

Real run: corpus constructionPravi run: gradnja korpusa

All standalone selftests, the FPLB3 red-team suite, the triple-gate tests, positive control, resume control, and E2E smoke passed. The master then began scanning D:\ to build the real 320-file corpus.

Vsi standalone selftesti, FPLB3 red-team paket, triple-gate testi, positive control, resume kontrola in E2E smoke so prestali. Master je nato začel pregledovati D:\ za gradnjo pravega 320-datotečnega korpusa.

[2026-07-19 15:33:58] [SCAN] walking D:\

No traceback, error, or failed gate in the supplied log.V priloženem dnevniku ni tracebacka, napake ali padlega gatea.

8zip ReZIP v0.2.5B

Archive 20 of 24Arhiv 20 od 24

Archive 19 finished with a verified 10.57 MB saving. On archive 20, the fresh one-thread 7‑Zip candidate completed at 64,879,746 bytes after 5:25:55; ECT9 MT1 is now running.

Arhiv 19 se je končal z verificiranim prihrankom 10,57 MB. Pri arhivu 20 je sveži enonitni kandidat 7‑Zip končal pri 64.879.746 bajtih po 5:25:55; zdaj teče ECT9 MT1.

[20/24] original=75,192,592 B
7-Zip=64,879,746 B · PASS
current=ECT9 MT1

No crash or non-zero return code in the supplied log.V priloženem dnevniku ni padca ali neničelnega return code.

Realistic upsideRealistični potencial

What could this become?Kaj lahko iz tega nastane?

There is no honest universal “maximum compression” number. The ceiling depends on the data. But the measured mechanisms already support three realistic development scenarios.

Poštena univerzalna številka »maksimalne kompresije« ne obstaja. Strop je odvisen od podatkov. Izmerjeni mehanizmi pa že podpirajo tri realistične razvojne scenarije.

Useful nowUporabno že zdaj

Verified ZIP optimizerVerificirani ZIP optimizator

ReZIP becomes a practical cold-storage tool that can recover anything from zero to low double-digit percentages, depending on how weak or redundant the original archive is, while preserving ordinary ZIP compatibility.

ReZIP postane praktično orodje za hladno hrambo, ki lahko glede na šibkost ali redundanco izvirnega arhiva prihrani od nič do nizkih dvomestnih odstotkov in ohrani običajno ZIP združljivost.

Strong research pathMočna raziskovalna pot

Specialized FPLB archivesSpecializirani arhivi FPLB

For project bundles, versioned JSON/CSV/log datasets, and backup trees with repeated content, exact dedup plus charged dictionaries can realistically produce persistent mid- to low-double-digit gains.

Pri projektnih paketih, verzioniranih JSON/CSV/log podatkih in backup drevesih s ponavljajočo vsebino lahko exact dedup ter obračunani slovarji realno prinesejo trajne srednje do nizke dvomestne prihranke.

High-upside frontierVisoko-potencialna meja

Near-duplicate chunks and deltasNear-duplicate bloki in delte

If content-defined chunking and base+delta candidates work without verifier complexity exploding, highly repetitive snapshots could go well beyond today’s 9% migration proof. This is plausible on the right data, not promised for arbitrary files.

Če content-defined chunking in base+delta kandidati delujejo brez eksplozije kompleksnosti verifierja, bi lahko zelo ponavljajoči snapshoti močno presegli današnji 9-odstotni migracijski dokaz. To je verjetno na pravih podatkih, ne obljuba za poljubne datoteke.

The most important positive constraint: already-compressed JPEG, MP4, encrypted, or high-entropy data may offer almost nothing. That does not weaken the architecture. It is why the arena keeps Store/fallback candidates and promotes a new method only when the complete verified result is genuinely smaller.

Najpomembnejša pozitivna omejitev: že stisnjeni JPEG, MP4, šifrirani ali visoko-entropijski podatki lahko ponudijo skoraj nič. To ne slabi arhitekture. Prav zato arena ohrani Store/fallback kandidate in novo metodo promovira le, kadar je celoten verificirani rezultat res manjši.

Next evidence gatesNaslednji dokazni gatei

The route forwardPot naprej

The project advances by freezing what is proven, then adding one major mechanism at a time so the source of every gain remains understandable.

Projekt napreduje tako, da zamrzne dokazano, nato pa doda po en večji mehanizem, da ostane izvor vsakega prihranka razumljiv.

Finish the v0.2.2 real runKončaj pravi run v0.2.2

Audit the new 320-file review bundle, compare FPLB2 and FPLB3, and verify archive-specific, group-safe, and external cross-archive claims separately.

Auditiraj novi 320-datotečni review bundle, primerjaj FPLB2 in FPLB3 ter ločeno preveri archive-specific, group-safe in zunanji cross-archive rezultat.

Finish ReZIP 24/24 and activate the R5 resume pathKončaj ReZIP 24/24 in aktiviraj R5 resume pot

Keep successful items, reprocess only missing or legacy cases, and calculate total weighted savings, median gain, winner provenance, and strict reverify outcomes.

Ohrani uspešne arhive, ponovno obdelaj le manjkajoče ali legacy primere ter izračunaj skupni uteženi prihranek, mediano, izvor zmagovalcev in stroge reverify rezultate.

Only then build v0.2.3Šele nato zgradi v0.2.3

The strongest candidate directions are content-defined chunking, near-duplicate block reuse, base+delta encoding, and cluster-specific dictionaries — always behind a no-regression champion gate.

Najmočnejše smeri so content-defined chunking, ponovna uporaba skoraj enakih blokov, base+delta kodiranje in cluster-specific slovarji — vedno za champion gateom brez regresije.