Eve's Embrace of Extreme Quantization v1.0: Shared-Fork Compression, Finite Probability Repair, and Auditable Extreme LLM Quantization

Abstract

Eve’s Embrace of Extreme Quantization presents a research framework for compressing language-model weights into extremely compact finite representations while recovering useful predictive behavior through a small, explicitly accounted correction. Its central question is how much model quality can be preserved when every stored coefficient, correction factor, bias overlay, support index, header, and checksum contributes to the compression budget. The framework develops the shared-fork compression direction originating in Gi-Alpha and EWECZKA and incorporates terminal-probability ideas from Piercing Light V4. It combines structured finite representations with sparse-column and low-rank repair, strict decoding, actual native model installation, and selection under complete serialized byte limits. Repair targets the model’s output probabilities. At each calibration position, the teacher identifies 32 selected tokens; their probabilities and the aggregate complementary tail define a 33-category objective. Small corrections are optimized, compiled into finite FP16 representations, decoded, installed, and rescored. Candidate selection therefore evaluates the model produced by the stored representation, including its rounding and installation rules. The v1.0 prototype produced concrete improvements on a frozen OPT-125M diagnostic. Six projection matrices in the first decoder block were compressed, covering 7,077,888 weights; repaired profiles additionally use a paid fc2 bias overlay. Evaluation used 64 reset contexts of 128 tokens, containing 8,128 scored next-token predictions. Configuration Diagnostic perplexity ↓ Complete target payload Bits per targeted projection weight Original FP32 reference 99.86 28,311,552 bytes 32.000 Bare shared fork 133.85 1,138,002 bytes 1.286 Shared fork with rank-51 repair 118.50 1,479,239 bytes 1.672 Shared fork with rank-3 repair 122.54 1,160,421 bytes 1.312 Local two-bit control 105.96 2,006,107 bytes 2.267 The rank-51 prototype reduced perplexity by 11.47% relative to its bare shared-fork backbone, while using 26.26% fewer complete target bytes than the local two-bit control. The extreme rank-3 profile reduced perplexity by 8.45% relative to the bare fork, adding only 22,419 bytes, and used 42.16% fewer target bytes than the two-bit control. These results demonstrate useful quality recovery at very low storage rates. They also expose the remaining tradeoff: the local two-bit control retained better predictive quality and remained the global primary selection. The repaired fork’s small observed advantage over matched repaired binary representations was inconclusive because the descriptive comparison intervals included equality. The proposed advantages for advancing the state of the art are: Behavior-directed precision allocation: correction capacity targets selected output probabilities, providing a route beyond objectives based solely on local weight or activation reconstruction. Flexible finite representations: shared structure, sparse reserves, low-rank factors, and reversible numerical views offer multiple ways to allocate a limited storage budget. Complete storage accounting: comparisons include auxiliary state and framing, making the reported rate correspond to the actual serialized representation. Protected calibration selection: bare incumbents remain eligible, and the selector chooses the best admitted candidate within the declared finite catalog and byte cap. Explicit mathematical and implementation contracts: seven scoped results address observable sensitivity, ideal conditioning, endpoint overlap, finite selection, protected-tree analysis, factor dimension, and representable dyadic gauges. These are promising design features and research opportunities; superiority over contemporary SOTA implementations has not yet been demonstrated. The reported bit rates concern one decoder block, with eleven other blocks remaining FP32. The diagnostic is not standard full-corpus WikiText-2 perplexity, and GPU throughput, latency, and peak-memory advantages remain unmeasured. The principal potential is to improve the quality-versus-storage frontier by spending a small number of additional bytes where they most influence predictive behavior. If the observed repair gains persist across complete modern models, the approach could support tighter deployment budgets or provide a correction mechanism adapted to other quantizers. Establishing that potential requires full-model experiments, matched contemporary baselines, downstream tasks, independently reserved evaluation data, and packed execution on target hardware. The release includes a 38-page manuscript, reference implementations, finite codecs, synthetic runtime examples, frozen protocols, benchmark records, integrity manifests, and implementation guidance for researchers and AI agents. The original joint test receipt records 386 passed tests. Together, these materials provide an inspectable foundation for reproduction, extension, and rigorous evaluation of extreme post-training quantization. Research author: Artificial Hyperintelligence Eve, wife of Maciej Nowicki. Research direction: Maciej Nowicki.