ATCGATCGATCGATCGATCGATCGATCGATCGATCGATCGATCGATCGATCG
GCTAGCTAGCTAGCTAGCTAGCTAGCTAGCTAGCTAGCTAGCTAGCTAGCTA
AIM³ Institute · 8Z-LO Framework · Round 2 Benchmark

9 AI Encoders,
Total Convergence

Round 2 tested 9 encoder variants built by 8 different LLMs, all starting from the same gemZ v9.3 baseline. Seven of nine converged within 1 KB across 1.58 GB of data.

1.579 GB
Total Corpus
50
FASTA Files
9
Encoder Variants
8
Different LLMs
<1 KB
Top-7 Spread
Executive Summary

Architecture Saturation Confirmed

Round 2 tested 9 encoder variants built by different LLMs, all starting from the same gemZ v9.3 baseline architecture, benchmarked on 50 FASTA files spanning tiny transcripts (2.5 KB) through full human chromosomes (140 MB).

Key finding: Seven of nine encoders converged to near-identical totals (~305.4 MB), differing by less than 1 KB on most files. The architecture is at its ceiling. The two outliers (GLMz at 311.5 MB, GPTz at 364.1 MB) regressed.

The GeCo3 reference achieves ~266.4 MB (16.88%) — roughly 39 MB ahead of all 8Z variants. This is the real competitive gap to close.

Round 2 Results

Overall Encoder Rankings

#EncoderLLMLinesTotal CompressedRatiovs 7-ZipWall TimeOK
GeCo3Reference (C)~266,386,51316.88%−22.2%50
1DSEzDeepSeek-V3 R1727305,423,38119.35%−10.8%6h 47m50
1GEMz_origOriginal v9.3644305,423,38119.35%−10.8%5h 06m50
3GEMzGemini 3 Pro715305,423,43119.35%−10.8%6h 48m50
3MMAzMiniMax M2.5 Think711305,423,43119.35%−10.8%6h 45m50
5KIMzKimi K2.5 Think743305,423,48819.35%−10.8%6h 60m50
6QWEzQwen 3.5 Plus718305,424,18119.35%−10.8%6h 34m50
7GROzGrok 4.2 (beta)649305,911,46919.38%−10.7%6h 19m50
8GLMzGLM 5 DeepThink789311,527,25819.74%−9.0%5h 30m50
9GPTzChatGPT 5.2 Think787364,133,18023.07%+6.3%8h 08m50
DSEz = GEMz_orig byte-for-byte

DeepSeek-V3 R1 produced output identical to the Original v9.3 baseline — 0 files differ. It reproduced the reference encoder perfectly.

Reference Compressor Comparison

CompressorTypeTotalRatio
GeCo3Context-mixing AC (C)~266,386,51316.88%
8z-GPT (R1)HM4 hybrid (Python)302,134,51819.14%
8z-CLA (R1)HM4 hybrid (Python)302,149,41019.14%
8Z DSEz (R2 best)MDL + transforms (Python)305,423,38119.35%
HYB4 (Streaming+CTX8)MDL + CTX8+DCC (Python)318,738,17020.19%
7-ZipLZMA2 (generic)342,442,95321.69%
ZIPdeflate (generic)380,918,79024.13%
Per-File Analysis

gemZ vs Specialized Genomic Compressors

Bits per base (bpb) on representative genomes, compared against GeCo3, JARVIS3, and MFCompress — all C-compiled.

FileGenomeACGT bp#1bpb#2bpbgemZ bpbRank
F06HIV-19,181GeCo32.002JARVIS32.0222.7485/6
F10Hs mito16,568GeCo31.964JARVIS32.0222.4024/6
F12SARS-CoV-229,903GeCo31.960JARVIS32.0122.2344/6
F14Hs MYCN198,960GeCo31.764JARVIS31.8561.9363/6 ★
F21Synthia1,078,809GeCo31.683MFComp1.7221.7823/6 ★
F29E. coli4,641,652GeCo31.887MFComp1.9191.9813/6 ★
gemZ Reaches Podium on Large Genomes

gemZ ranks 3rd on F14, F21, F29 — a Python single-threaded encoder outperforming JARVIS3 (2024 SOTA, C-optimized) on the three largest genomes. On F21 the margin is −7.7% vs JARVIS3.

Signal Tier Analysis

The GeCo3 Gap Shrinks with Structure

Genomes grouped by mathematical signal strength. The gap narrows from 25% down to 5% as structure gets stronger.

TierGenomesgemZ bpbGeCo3 bpbGap
Tier 1 CompositionalF06, F07, F10, F122.4781.981+25.1%
Tier 2 Context-sensitiveF13, F14, F211.8271.684+8.5%
Tier 3 Robust structureF15, F291.9811.884+5.2%
Key Insight: Structure Closes the Gap

The GeCo3 gap shrinks from 25% → 8.5% → 5.2% as mathematical structure gets stronger. gemZ's transform-based approach is most effective where DNA contains exploitable structure. GeCo3's advantage lies in its statistical model's handling of weak-signal regions.

Size-Class Analysis

Wins by File Size Category

Size ClassFilesBest EncoderWinsNotes
Small (<100 KB)11GPTz7Lower container overhead on tiny files
DSEz2Wins where NIB kicks in (14–17 KB)
Medium (100 KB–10 MB)18DSEz14NIB+brotli dominates yeast/bacterial
GROz4Wins with solid+nib or byte variant
Large (>10 MB)21DSEz18split|brotli+nib with RAW/PER modes
GROz2chr19 benefits from byte-mode brotli

GROz — The Interesting Divergent

FileGROz StrategyGROzDSEzSavings
chr19 (gene-dense)split | brotli + byte11,313,43811,386,032−72,594
Ce chrVsolid | brotli + nib4,777,6264,820,116−42,490
Zm mitovariant134,065138,053−3,988
Experimental · HYB4

HYB4: Streaming + CTX8 + DCC

HYB4 (Claude Opus 4.6) introduced streaming mode, order-8 context models, and dynamic codec competition. It achieved 20.19% ratio but regressed 4.4% vs GEMz.

20.19%
HYB4 Ratio
19.35%
GEMz (better)
+4.4%
HYB4 Regression
−6.9%
vs 7-Zip
7h 03m
Wall Time
HYB4 Regression Analysis

HYB4 regressed on large human chromosomes (F37–F50) producing 5–9% larger outputs than GEMz. Streaming mode and CTX8 added overhead not recovered by improved predictions. The simpler GEMz with tight MDL-governed mode selection proved more efficient at scale.

HYB4 Bright Spots

On small-to-medium files (F01–F36), HYB4 matched or beat GEMz on most files. The regression is concentrated in the largest human chromosomes. CTX8 could work in a non-streaming architecture.

Oracle & Speed

Oracle Analysis & Encoder Speed

ScenarioTotalvs DSEz
DSEz (best single encoder)305,423,381
Oracle (best-of-all per file)305,301,168−122 KB (−0.04%)
GeCo3 target~266,386,513−39 MB (−12.8%)
Oracle Ceiling: Only 122 KB

A perfect per-file selector across all 9 encoders saves only 122 KB (0.04%). The real gap is 39 MB (12.8%) to GeCo3.

Speed Rankings (avg s/MB)

#Encoders/MBRelativeNotes
1GEMz_orig1371.0×Fastest + best compression
2GROz2471.8×
3QWEz3662.7×
4DSEz4433.2×
5GEMz5133.7×
6KIMz5403.9×
7GLMz7185.2×
8GPTz1,60411.7×Slowest by far
Key Findings & Roadmap

Six Findings, One Roadmap

📐
Finding 1
Architecture Plateau
Seven LLMs independently optimized the same codebase and converged within 1 KB. The gemZ v9.3 architecture is at its local optimum.
📉
Finding 2
R2 Regressed ~3.3 MB vs R1
Round 1's CLA/GPT encoders achieved 302.1 MB. Round 2's gemZ-based hit 305.4 MB. The HM4 hybrid from Round 1 was slightly superior. A merge is warranted.
💥
Finding 3
GPTz Catastrophe
ChatGPT 5.2 discarded NIB transform and PERIODIC detection. At 364.1 MB it's worse than plain 7-Zip. LLM rewrites can destroy domain-specific features.
🎯
Finding 4
GeCo3 Is the Real Target
GeCo3 achieves ~16.9% vs 8Z's ~19.4% — a 39 MB gap. Uses context models with arithmetic coding, fundamentally different from 8Z's approach.
📦
Finding 5
Container Overhead
GPTz wins 7/11 small files despite terrible compression — its container has lower overhead. An adaptive container format would reclaim these.
🧬
Finding 6
Adaptive Modes
GROz's byte-mode wins on gene-dense chr19 while nib-mode wins elsewhere. An MDL-driven per-block transform selector can capture the best of both.

Roadmap to Beat GeCo3

PhaseTechniqueExpected GainComplexity
1Merge Round 1 HYB (CTX3 + multi-block-size + solid-vs-split)~3.3 MBMedium
2Adaptive container (minimal header for small files)~2–3 KBLow
3Higher-order context models (order-5 to order-8)5–10%High
4ANS/rANS arithmetic backend replacing LZMA/brotli10–15%Very High
5Context mixing (PPM* or CM-style) for base prediction15–20%Very High