Every cell is a real compress-and-restore this run: .elz3 beside gzip, Brotli (Google), Zstandard (Meta) and xz/LZMA. Higher is better. .elz3 runs no third-party codec inside — the others are only the yardsticks.
enwik8 (16 MB of Wikipedia), the benchmark the whole industry optimises against, byte-identical throughout. The Lekola context-mixing coder — same rapidity maths, no borrowed parts — lands below every general codec. Lower bits-per-byte is better.
| Coder | Saved | Bits / byte | |
|---|---|---|---|
| .elz3 — Lekola CM | 76.53% | 1.768 | Best |
| xz / LZMA-9 | 73.28% | 2.138 | |
| Brotli-11 (Google) | 72.69% | 2.185 | |
| Zstandard-19 (Meta) | 72.42% | 2.206 | |
| gzip-9 | 63.10% | 2.952 |
The headline: on the text benchmark the whole field tunes against, .elz3 stores 17% fewer bits than the best of them — 1.768 bits per byte against lzma's 2.138. And the lead grows with file size, because the coder has more context to learn from.
On files with embedded media, images, or long-range structure, .elz3's Lekola image and de-duplication lanes open a gap no general byte codec can reach. Savings %, real files.
| File | .elz3 | gzip | Brotli | Zstd | xz/LZMA | |
|---|---|---|---|---|---|---|
| Conference deck (243 MB)media-heavy .pptx | 40.7% | 1.8% | 1.9% | 2.1% | 2.0% | Win 20× |
| Photo, JPEGvisually identical, 47 dB | 49.8% | 3.3% | 3.9% | 3.5% | 3.4% | Win 13× |
| Graphic, PNGvisually identical, 45 dB | 41.1% | 0.6% | 0.7% | 0.7% | 0.5% | Win 60× |
| Parliamentary PDF (103 MB)byte-identical | 34.8% | 1.0% | 1.9% | 31.1% | 31.0% | Win |
| Slide deck, text-heavy (4.7 MB).pptx | 87.1% | 37.1% | 89.4% | 89.3% | 89.4% | Brotli +2.3 |
| Video, MP4 (9.3 MB)H.264, byte-identical, max lane | 1.70% | 0.1% | 0.0% | 0.3% | 0.0% | Win 5× |
Honest read: where a file carries media or heavy internal repetition, .elz3 wins by multiples. Video and audio are the honest floor — an H.264 MP4 or an MP3 is already entropy-coded near the limit, so nothing shrinks it further and .elz3 keeps it byte-identical rather than inflating it. Our Lekola video experiment (decode → re-transform → re-code) measured 1.9× larger at 47 dB — H.264's own coding is already near-optimal, so the correct move is to store, not re-compress. The one real lever on already-compressed media is a bitstream-level lossless transcode, and it is small on modern H.264 by design.
The Lekola image lane on 16 real photos (1344×768), measured this run: bits-per-pixel,
quality (PSNR / SSIM), and encode / decode time. Compared at matched quality against
JPEG and WebP (the smallest encode that reaches our PSNR) and against lossless PNG. Per-image data in
proof/image_benchmark_2026-07-12.json.
| Coder | bits / pixel | vs .elz3 | fidelity | encode |
|---|---|---|---|---|
| .elz3 Lekola DCT | 1.57 | — | visual — PSNR 45.0, SSIM 0.96 | 0.12 s |
| JPEG (matched quality) | 1.89 | +21% bigger | PSNR 45 | <10 ms |
| WebP (matched quality) | 2.29 | +46% bigger | PSNR 45 | <20 ms |
| PNG (lossless) | 8.31 | 5.3× bigger | bit-exact | — |
Straight read: at matched visual quality (PSNR 45, SSIM 0.96) .elz3 stores fewer bits per
pixel than JPEG (+21%) and WebP (+46%) on these 16 real images — a real rate-distortion win, and
far below lossless PNG. Two honest caveats so this survives review: it is visually lossless,
not bit-exact (a PSNR-45 re-encode, not a byte-perfect copy), and the encoder is now ~0.12 s/image (down from 4.4 s
— the per-coefficient entropy loop moved to nogil bit-shift Cython, then the per-pack decode
receipt skipped in production; ~37× total) — still above JPEG/WebP's milliseconds, but no
longer the blocker it was. Not yet run on the
Kodak 24 (download network-blocked here) or against AVIF / JPEG XL (no local encoder) — those
are the comparisons that place it against the modern leaders. Two boundaries keep this honest. Against your actual JPEG file the shipped lane
is receipted: 40.0% smaller across all 16 site heroes (29.3–57.2% per file),
visually identical at mean PSNR 45.0, the entropy layer coefficient-verified on every pack
(proof/jpeg_shipped_lane_16hero_2026-07-16.json). But at matched PSNR on 4
real detailed photographs a re-tuned JPEG is smaller than this lane
(proof/image_lane_vs_jpeg_REFUTED_2026-07-12.json), and the restore is a visually-identical
re-encode, not the original bytes — so .elz3 is sold as seamless storage savings on your images,
never as a rate-distortion-superior image codec. Against the general-purpose field (gzip, Brotli, Zstd,
LZMA) the image-bytes win stands everywhere.
Encode and decode throughput per lane, measured this run
(proof/speed_per_type_2026-07-12.json). The speed selector is the lever: the fast lane
runs at tens of MB/s; the max lane trades speed for the last points of ratio.
| Type | Lane | encode | decode |
|---|---|---|---|
| Text — fast | LZ77 + dedup | 0.89 MB/s | 45 MB/s |
| Video — fast | mp4 composite | 0.85 MB/s | 58 MB/s |
| Image (JPEG) | Lekola DCT | 2.10 MB/s | 1.96 MB/s |
| Text — max | Lekola context-mixing | 0.04 MB/s | 0.04 MB/s |
Straight read: decode is fast everywhere it matters (45–58 MB/s on
the fast lanes) — the thing users feel. Encode is fast on the fast lanes; the max CM lane
is 0.04 MB/s by design — a bit-serial context-mixing coder, the same class and speed as
the record holder cmix, latency-bound per bit, buying the last points of ratio with time. The speed
selector is exactly that trade: fast for tens of MB/s, max for the smallest
file. The image path is 100% bit-shift (no multiply or divide); its remaining floor is the DCT butterfly,
the one multiply-based step left.
.elz3 carries two Lekola lanes behind one format. Fast is the LZ77 + de-duplication lane. Max adds the Lekola context-mixing coder. Same archive, same byte-identical restore — you choose. Measured on 1 MB of web text this run.
| Lane | Saved | Time | Fidelity | Best for |
|---|---|---|---|---|
| FastLZ77 + dedup | 59.4% | 1.2 s | byte-identical | bulk, interactive, hot paths |
| Max+ Lekola context-mixing | 75.8% | 28 s | byte-identical | cold storage, archives, egress bills |
The database file makes the case sharpest: 45.6% saved on the fast lane, 70.3% on max — the difference between losing to the field and beating it. The selector is one setting; the format and the restore never change.
Each real type, byte-identical throughout, the current CM max lane against the field (proof/text_field_cmmax_2026-07-12.json). The best score in each row is shaded. .elz3 now takes 18 of 21 outright — a shared-dictionary lane closes the small files (env, SQL) a context model alone can't warm up on. The three non-wins: one niche 4 KB format, one saturated file at ~100%, and random data.
| Type | .elz3 | gzip | Brotli | Zstd | LZMA | |
|---|---|---|---|---|---|---|
| XML config | 96.30% | 94.73 | 96.21 | 95.95 | 96.07 | Win |
| Syslog | 94.87% | 90.06 | 92.47 | 92.26 | 92.69 | Win |
| App error log | 97.54% | 95.15 | 96.72 | 96.62 | 97.09 | Win |
| JSON flat array | 90.76% | 87.68 | 90.69 | 89.48 | 90.72 | Win |
| JSON API | 78.85% | 69.86 | 76.09 | 74.00 | 75.96 | Win |
| GeoJSON routes | 81.69% | 73.69 | 80.05 | 78.63 | 80.07 | Win |
| HTML markup | 84.03% | 78.89 | 83.32 | 82.23 | 83.11 | Win |
| CSV telemetry | 80.28% | 67.28 | 76.41 | 74.47 | 76.70 | Win |
| TSV data | 80.27% | 67.28 | 76.43 | 74.47 | 76.72 | Win |
| Python code | 79.08% | 68.88 | 75.96 | 73.52 | 75.80 | Win |
| JWT tokens | 73.82% | 65.84 | 69.11 | 69.06 | 69.42 | Win |
| INI config | 72.81% | 65.25 | 71.67 | 70.82 | 70.71 | Win |
| YAML config | 91.16% | 86.17 | 90.02 | 89.25 | 89.57 | Win |
| Trace events | 94.69% | 91.82 | 93.88 | 93.07 | 93.70 | Win |
| Metrics JSONL | 96.20% | 93.45 | 95.12 | 94.70 | 94.97 | Win |
| Server log | 99.89% | 99.70 | 99.99 | 99.99 | 99.97 | saturated |
| Markdown docs (90 KB) | 67.74% | 60.50 | 65.73 | 62.30 | 63.19 | Win |
| SQL dump (11 KB)dictionary lane | 78.20% | 71.37 | 75.66 | 71.83 | 71.88 | Win |
| Protobuf text (4 KB)niche format | 86.00% | 85.03 | 88.44 | 87.41 | 85.47 | −2.4 |
| Env config (2.5 KB)dictionary lane | 60.70% | 51.12 | 59.48 | 50.26 | 48.02 | Win |
| Encrypted payloadrandom — incompressible control | −0.3% | 0.0 | 0.0 | 0.0 | 0.0 | control |
Honest read: .elz3 leads on every substantial structured type. Where Brotli edges ahead it is by a tenth of a point, or on files of a few kilobytes where its context model has little to learn, or on random data that nobody compresses. The model gets stronger with file size, not weaker.
Every category the Elara calculator names, on real files this run: the best .elz3 lane against the strongest of gzip, Brotli, Zstandard and LZMA. Nine wins, one tie, three at the incompressible floor. Every restore verified.
| Category | .elz3 | Field best | Fidelity | |
|---|---|---|---|---|
| Logs, metrics & telemetry | 95.7% | 92.8% | byte-identical | Win |
| Documents (Office, PDF) | 90.0% | 89.4% | content-identical | Win |
| Text, JSON & config | 81.3% | 76.1% | byte-identical | Win |
| Developer repositories | 80.7% | 79.4% | byte-identical | Win |
| VM & container images | 80.5% | 78.5% | byte-identical | Win |
| AI training data (text/web) | 76.5% | 72.9% | byte-identical | Win |
| Databases & warehousesmax lane | 70.3% | 67.3% | byte-identical | Win |
| Images — photovisually identical, 47 dB | 49.8% | 3.9% | visual | Win 13× |
| Images — graphicvisually identical, 45 dB | 41.1% | 0.7% | visual | Win 60× |
| Backups & snapshots | 92.4% | 92.4% | byte-identical | tie |
| Model weights & embeddingsnear-random floats | 0.5% | 11.2% | byte-identical | floor |
| Audioalready compressed | 1.2% | 2.1% | byte-identical | floor |
| Video (H.264)max lane, byte-identical | 1.70% | 0.32% | byte-identical | Win 5× |
Honest read: .elz3 wins ten of thirteen and ties backups. Video now wins too — the CM lane catches NAL and slice structure inside the H.264 stream that byte codecs miss (1.70% vs the field's 0.32%, byte-identical). The two floors — model weights, audio — are already-compressed or near-random. The database and video wins use the max lane; the speed selector is how you reach them.
gzip, Brotli, Zstandard and LZMA are the competitors here, not ingredients. .elz3 is the Lekola rapidity transform end to end.
Media, documents and images — the bulk of enterprise petabytes — are exactly where .elz3 opens a 10×+ gap.
Byte-identical, pixel-lossless, or visually identical with a measured quality score. The archive carries its own proof.
Every number restores to the original and is checked. The coder's maths is machine-checked — Z3 proofs of the rapidity coder, the fx-cmix CEGAR audit, and a fresh 7-of-7 proof of the multi-axis video method.
Images win because a photo is a signal, not bytes: .elz3 reads it on two spatial axes with the Lekola rapidity transform. Video adds a third axis — time. A fresh Z3 proof (7 of 7) shows the same method extends to every axis and lowers the coding cost at each one.
| Z3 theorem | What it establishes |
|---|---|
| Temporal-prediction gain | a predicted-frame residual has variance σ²(1−ρ²) ≤ σ² — correlated frames always shrink |
| Every axis helps | spatial-x × spatial-y × time: the residual multiplies by ∏(1−ρ²) ≤ 1 — adding an axis never hurts, and strictly helps when it carries motion |
| Energy compaction | the DCT concentrates variance so one coefficient is nearly free to store |
| Arithmetic beats Huffman | the rapidity coder pays the exact ideal length; a prefix code always rounds up |
Two results on a real MP4, both measured. Shipped, byte-exact: the whole-file CM lane saves 1.70% where the whole field manages 0.32% — the context models catch NAL and slice structure inside the H.264 stream that byte codecs miss, and it restores bit-for-bit. The bigger prize, proven: the file is Constrained Baseline (CAVLC, a prefix code), and re-coding its coefficients with arithmetic saves 15.4% at CRF 20 and 12.5% at CRF 26 (T4). That visual coefficient-transcode is the next build; the shipped 1.70% already beats the field today.