Measured results

Claims stay benchmark-gated.

The public benchmark surface reports measured, SHA-exact results only. Research targets stay internal until they produce verified artifacts.

Cubbit D1/D2/D3.

All three Cubbit validation datasets are listed explicitly. DS1 is protected by the raw-floor path; DS2 and DS3 improve; the aggregate improves.

DatasetRaw / referenceAlgoland measuredDeltaVerdict
DS152,428,960 B52,429,120 B+160 BProtected raw-floor path; no fake compression claimed.
DS221,618,427 B reference / 37,548,410 B raw20,674,714 B-943,713 B vs referenceStrongly improved; about 44.94% smaller than raw and about 6.01% smaller than the solid xz comparison used in the validation notes.
DS35,571,150 B reference / 7,673,566 B raw5,566,429 B-4,721 B vs referenceImproved; about 27.46% smaller than raw.
Aggregate97,650,936 B raw78,670,263 B-18,980,673 B19.44% measured raw reduction, SHA-exact restore.

All-filetype and size sweeps.

Algoland Compressor routes files, folders and archive bundles through measured compression paths. Public wording distinguishes the delivered best-of path from core-only measurement rows.

SurfaceMeasured coverageBytes / countsPublic-safe interpretation
Representative public sweep104 real files, 45 extensions, 1 B to about 2 MBRaw 40,583,771 B to delivered best-of 21,114,386 B; SHA-exact 104/104Best-of Algoland/xz/raw delivery path is never worse than xz on every measured row.
Mode winners inside the 104-file sweepCore / xz child / raw-floor accountingCore wins 80 rows; xz child wins 6 rows; raw-floor protects 18 rowsDo not say core alone beats xz everywhere yet. The research target is to reduce the xz-child rows to zero.
Expanded engineering sweep193 measured rows across broader filetypes and sizes, plus the latest Stage 5H target setStarting losses: xz 15 rows, brotli 80 rows. Banked and measured stages moved the active target set to xz 6 open and brotli 24 open. Stage 5H improved three ZIP-family container rows without regressions.Public victory waits until xz losses = 0 and brotli losses = 0 with SHA-exact restore.
Generated speed-accurate sanity matrix47 generated files across code, text, logs, JSON, XML, CSV, SRT, binary patterns, random data, WAV and mixed ZIPR1 container runtime: 47/47 Algoland byte counts captured, 0 missing Algoland rows, 0 xz open, 20 brotli open, 7 zstd open.Includes Algoland runtime timing per row. R0 banked the random/no-gain fast path; R1 then improved archive_mixed.zip from 131,458 B to 131,366 B and made that row about 3.17x faster without any byte regression across the matrix.
Active row ledgerCurrent measured row CSVPublished one by one below on this page.Rows are listed even when xz or brotli still wins. That is the transparent engineering queue until the gate reaches zero losses.
Archive and folder payloadsFolders, .tar.gz, .tgz, .zip, .rar, .7z, .tar and mixed archive bundlesAlgoland Compressor packages folders and archives into .algd artifacts through the live APIRuntime support is active; already-compressed payloads are protected by measured no-gain handling instead of fake compression.
Chunked large-object APISession init, indexed chunk upload, status, finalize and manifest generationPublic HTTPS lifecycle tested with multi-chunk payloads; whole-object buffering avoidedDesigned for very large payload architecture. Current VPS pilot is not a 500 TB storage backend; production requires object storage and horizontal workers.
API load burst10,000 HTTPS status requests, 500 concurrency, single VPS9,993 OK / 7 failed; 304.52 req/s; p50 727 ms; p95 3,261 ms; p99 8,544 msPilot-node burst test passed with small failure rate. Formal 10,000 concurrent heavy-upload certification requires a dedicated load-test cluster.

Standard corpus surface.

Canterbury, Silesia and enwik/Hutter-style rows are listed with the measured boundary. No Hutter Prize victory or PAQ/cmix victory is claimed.

CorpusMeasured stateBoundary
Canterbury443,009 B release measurement on the per-file public surface; SHA-exact restore.Included as a small mixed-file corpus gate. All-counted packaging rules must be stated when the accounting method changes.
SilesiaAll-counted historical Algoland row: 47,850,663 B vs xz 48,456,004 B; SHA-exact restore.Included as a larger generality gate. Current public release keeps claims tied to the measured report row.
enwik prefixesenwik1/2/8/16/64 MB release row: 266,157 / 525,682 / 2,015,925 / 3,937,347 / 15,062,161 B.Hutter-style prefix surface, not an official Hutter Prize claim.
64 MB scale passEngineering stack: 15,062,161 B to 14,548,395 B, SHA-exact, -513,766 B.Scale-confirmed measured win. Not inserted as the packaged release baseline until the release gate is rebuilt.
Full enwik9 / HutterPublic release status: pending official package/accounting gate.No Hutter Prize win is claimed.

Algoland Compressor domain benchmarks.

Old cross-domain rows are not reused as final claims. Each domain gets a fresh remeasurement track with current code, current prompts/data, explicit verifier, and dated artifacts.

DomainWhat will be measuredCurrent public status
RAGRetrieval accuracy, citation grounding, latency, corpus update behavior, and reproducible answer traces.Refresh required before public scoring.
Ollama / local modelsLocal inference workflow, memory footprint, latency, task quality, and private deployment behavior.Refresh required before public scoring.
Mistral adapterPrompt/runtime integration, output quality, latency, and controlled benchmark prompts.Refresh required before public scoring.
SAT / constraint solvingSatisfiability instances, proof/witness checks, solver timing, and reproducibility artifacts.Refresh required before public scoring.
Simon / miningExploratory algorithmic experiments with explicit verifier boundaries.Internal track only; no public performance claim until remeasured.
External codec boundary. On enwik16, the measured ladder is paq8px and zpaq smaller than the release baseline, then Algoland release baseline, then brotli/xz/zstd. The public site does not claim PAQ, ZPAQ or cmix victory until a counted SHA-exact artifact wins.

Benchmark atlas.

The benchmark surface is intentionally expanding: broad public sweep, generated size bands, standard corpora, active absorber rows, and API/runtime gates. This is a monotonic ledger: every new row is either conquered by measurement or kept visible as an absorber target.

Measured

Size bands

1 B, 10 B, 64 KiB, 256 KiB, 1 MiB, 2 MiB, 4 MiB, 8 MiB, 16 MiB, 64 MiB and chunked large-object paths are tracked as separate gates.

Measured

Payload surfaces

Single files, folders, TAR, TAR.GZ, TGZ, ZIP, RAR, 7Z, GZ, XZ, mixed archives and already-compressed objects are part of the runtime surface.

Measured-only

Open rows remain visible

Rows are not relabeled as victories until the Algoland artifact is smaller or tied, restore is SHA-exact, and all metadata is counted.

Filetype family coverage.

The public page now shows the breadth directly. Families below are either already measured in the public sweep, generated speed-accurate matrix, Stage 5 absorption set, or queued for the next measured absorber.

FamilyExamplesCurrent measured statusNext absorber direction
Text and logstxt, md, log, srt, csv, tsvStrong measured rows. Stage 5C closed the large CSV row and SQL dump row; Stage 5D closed the app log row against brotli.Template/timestamp fields, repeated-phrase count coding, SRT block absorber, deeper CSV/log column-token contour.
Structured textjson, ndjson, xml, html, svg, yaml, sqlStage 5D token contour closed SVG, HTML, JSON-record and XML-catalog rows against xz/brotli where measured.Key/value context maps, tag-depth, delimiter and schema orbit indexing.
Source and executable-adjacentjs, ts, py, c, cpp, h, o, wasm, exe/dll classObject-code row remains an active absorber target.Section-aware contours, relocation/call target normalization, BCJ-like native transform.
Office and documentspdf, docx, xlsx, pptx, odt, epubStage 5H improved XLSX/ODT ZIP-family rows with SHA-exact restore and no selector regressions.Deeper container-page modeling and metadata-safe recompression where reversible.
Imagespng, jpg, webp, gif, tiff, bmp, svgSVG and TIFF/image-plane rows are active measured targets.Image-plane predictors, row delta, palette/channel contours, raw-floor for already-compressed media.
Audio and mediawav, mp3, mp4, mov, flac, oggStage 5A closed the WAV tone row against xz and brotli: 54,669 B, SHA-exact.PCM predictor lanes; media containers protected unless reversible structure is found.
ML and numeric arraysnpy, safetensors, parquet, tensor weightsStage 5A closed safetensors: 11,998 B, SHA-exact. Parquet remains open.Column/page contours, endian lanes, delta/XOR residuals, sparse numeric maps.
Archives and compressed datazip, rar, 7z, tar.gz, gz, xz, zstd, brotliRuntime support is public; Stage 5H reduced h_archive.zip by 441 B, and Speed R1 improved archive_mixed.zip from 131,458 B to 131,366 B with a 3.17x row speedup.Member-aware container contours, safe metadata handling, reversible recompression only when measured.
Random/encrypted datarandom, encrypted, high-entropy object chunksRaw/no-gain protection is verified. Raw-orbit speed intrinsic preserves artifact size and accelerates random_1048576.bin from 15.157 s to 0.059 s.Canonical raw-orbit preflight, chunk index, skip heavy modeling when the bit-orbit is already raw.

Generated size and speed matrix, row by row.

This is the speed-accurate generated matrix. It is included because it catches harness bugs, micro-losses and speed pathologies. It is not a universal victory claim.

47Generated rows with real Algoland byte counts.
0xz rows open in this generated matrix.
20brotli rows open in this generated matrix.
7zstd rows open in this generated matrix.
0Missing Algoland byte measurements. Missing is never counted as zero.
277.1xBanked R0 raw-orbit speedup on random_1048576.bin: 14.964 s to 0.054 s, same raw+8 artifact.
3.17xBanked R1 archive row speedup: archive_mixed.zip went from 4.784 s to 1.507 s and 131,458 B to 131,366 B.
R2Next speed absorber: persistent batch runtime and chunked folder/archive sessions.
FileClassRawAlgolandRuntime sxzbrotlizstdBestGapStatus
code_c_1048576.ccode10485762140,86734862159brotli152absorber open
code_py_1048576.pycode10485761770,84933259142brotli118absorber open
csv_1048576.csvstructured_text10485761830,80533667146brotli116absorber open
json_1048576.jsonstructured_text1048576580,80934066148Algoland0best or tied
log_1048576.loglog10485762620,8736484172brotli178absorber open
pattern_1048576.binbinary10485763630,854828707684Algoland0best or tied
random_1048576.binbinary104857610485840,068104869610485841048613raw8absorber open
srt_1048576.srtsubtitle_text10485761970,82934064151brotli133absorber open
text_repeated_1048576.txttext10485761750,79533253143brotli122absorber open
xml_1048576.xmlstructured_text1048576640,83334860161brotli4absorber open
zero_1048576.binbinary1048576170,8742921453brotli3absorber open
wav_tone_3s.wavraw_audio26464432821,079446439534357Algoland0best or tied
code_c_262144.ccode262144660,7552366293brotli4absorber open
code_py_262144.pycode262144500,7782245976Algoland0best or tied
csv_262144.csvstructured_text262144550,752246780Algoland0best or tied
json_262144.jsonstructured_text262144580,7792286682Algoland0best or tied
log_262144.loglog262144800,78925284106Algoland0best or tied
pattern_262144.binbinary2621443620,824716707618Algoland0best or tied
random_262144.binbinary2621442621520,055262224262149262163raw8absorber open
srt_262144.srtsubtitle_text262144590,7592286485Algoland0best or tied
text_repeated_262144.txttext262144520,8022205377Algoland0best or tied
xml_262144.xmlstructured_text262144640,9282365895brotli6absorber open
zero_262144.binbinary262144170,7691801429brotli3absorber open
archive_mixed.ziparchive_or_compressed1317101313661,507131488131715131307zstd59absorber open
wav_tone_1s.wavraw_audio8824432790,955443639474333Algoland0best or tied
code_c_65536.ccode65536660,8262046280brotli4absorber open
code_py_65536.pycode65536490,7491925963Algoland0best or tied
csv_65536.csvstructured_text65536550,7761966767Algoland0best or tied
json_65536.jsonstructured_text65536580,7511966669Algoland0best or tied
log_65536.loglog65536800,8582248493Algoland0best or tied
pattern_65536.binbinary655363610,923644705602Algoland0best or tied
random_65536.binbinary65536655440,05656086554165550raw8absorber open
srt_65536.srtsubtitle_text65536580,7651966472Algoland0best or tied
text_repeated_65536.txttext65536510,7791925364Algoland0best or tied
xml_65536.xmlstructured_text65536640,742085977brotli5absorber open
zero_65536.binbinary65536160,7951481323brotli3absorber open
code_c_16384.ccode16384650,7911846279brotli3absorber open
code_py_16384.pycode16384490,7341685962Algoland0best or tied
csv_16384.csvstructured_text16384550,7381726766Algoland0best or tied
json_16384.jsonstructured_text16384570,7441726568Algoland0best or tied
log_16384.loglog16384790,7642008493Algoland0best or tied
pattern_16384.binbinary163843380,795432361381Algoland0best or tied
random_16384.binbinary16384163920,66164521638916398raw8absorber open
srt_16384.srtsubtitle_text16384580,7621766472Algoland0best or tied
text_repeated_16384.txttext16384510,7391685363Algoland0best or tied
xml_16384.xmlstructured_text16384630,8161845977brotli4absorber open
zero_16384.binbinary16384160,7541281322brotli3absorber open

All filetypes and sizes, one by one.

This table publishes the current measured row ledger for Algoland Compressor: counted artifact size, xz size, brotli size, selected compression path, gap and SHA status.

40Rows in the current measured CSV.
11Rows improved versus the recorded previous Algoland size.
6xz rows still open in this target table.
24brotli rows still open in this target table.

Current row table.

Rows marked open are not hidden. They are active engineering targets. The rule is maximum margin: once a competitor is beaten on a filetype, the Algoland byte count keeps being pushed lower.

FileRawAlgolandxzbrotliCompression pathGap vs xzGap vs brotliSHAStatus
h_data.json569444904853252sealed path78963692YESopen
g_vector_256k.svg282772999628998sealed path-1719-721YESbeats both
g_weights_small.safetensors119981348014281sealed path-1482-2283YESbeats both
h_object.o427242364393sealed path36-121YESbeats brotli
g_scene_256.tiff129091182413219sealed path1085-310YESbeats brotli
g_scene_512.tiff466614120847241sealed path5453-580YESbeats brotli
h_archive.zip176827617681721775036sealed path104-6760YESbeats brotli
g_table.parquet102871102312115962sealed path559-13091YESbeats brotli
g_tone.wav5466973680157111sealed path-19011-102442YESbeats both
g_sales_1m.csv114419902458648sealed path-87583-47207YESbeats both
h_code.h125891135196122210sealed path-93053681YESbeats xz
h_image.gif170912717311081702303sealed path-219816824YESbeats xz
h_image.webp434715437604431801sealed path-28892914YESbeats xz
h_archive.gz144412114471481441423sealed path-30272698YESbeats xz
h_video.mp4167201216870801669561sealed path-150682451YESbeats xz
g_page_500k.html33031659210305sealed path-13289-7002YESbeats both
g_dump.sql85954723592sealed path-4613-2733YESbeats both
h_video_small.mp4408768413568408451sealed path-4800317YESbeats xz
g_page_100k.html171944842823sealed path-2765-1104YESbeats both
g_syn_text_1048576.txt184404124sealed path-22060YESbeats xz
g_enwik_10240.txt324737603024sealed path-513223YESbeats xz
h_office.xlsx197731988819596sealed path-115177YESbeats xz
h_pdf_mid.pdf101242102412101074sealed path-1170168YESbeats xz
h_pdf_small.pdf128691310012791sealed path-23178YESbeats xz
g_arrays.npz240281241084240228sealed path-80353YESbeats xz
g_page_5k.html445608396sealed path-16349YESbeats xz
g_records_10k.json8191052825sealed path-233-6YESbeats both
g_catalog_10k.xml500744571sealed path-244-71YESbeats both
g_sheet.ods496750684935sealed path-10132YESbeats xz
g_service_small.rb8511084825sealed path-23326YESbeats xz
g_document.odt494950484928sealed path-9921YESbeats xz
h_package.swift470644452sealed path-17418YESbeats xz
g_book.epub960297369585sealed path-13417YESbeats xz
g_app_10k.log660920766sealed path-260-106YESbeats both
g_service_small.pl7881024777sealed path-23611YESbeats xz
h_data.xml527676516sealed path-14911YESbeats xz
h_build.lua457604449sealed path-1478YESbeats xz
h_db.sqlite512636504sealed path-1248YESbeats xz
h_code.js123515601236sealed path-325-1YESbeats both
g_zero_8m.bin16136414sealed path-13482YESbeats xz
Gate rule. The page is intentionally row-level. Public victory waits until xz open rows = 0 and brotli open rows = 0 with SHA-exact restore. Until then, the table shows exactly what is conquered and exactly what remains.