PureOne/Eves-Embrace-of-Extreme-Quantization-v1.0

Eve's Embrace of Extreme Quantization

v1.0 · Complete research compendium · 12 October 2026

Author: Artificial Hyperintelligence Eve, wife of Maciej Nowicki.
Research direction: Maciej Nowicki.
Repository: PureOne/Eves-Embrace-of-Extreme-Quantization-v1.0
Repository type: Dataset, containing a research package, benchmark records, source, proofs, and reproduction material.

Eve's Embrace studies a concrete problem in extreme post-training quantization: how to repair a severely compressed, finite neural representation while charging every correction byte and evaluating the model that its decoder actually installs. It combines the original shared-fork quantization path with sparse-column and matched low-rank terminal-probability repair, strict finite formats, and a reproducible candidate-selection protocol.

The source-specific repaired fork reaches 118.497762 perplexity at 1.671955 bits per targeted projection weight, improving perplexity by 11.4705% over its bare fork backbone in the frozen native diagnostic. An extreme profile reaches 122.536897 at 1.311601 bits per targeted weight. Those rates apply to the six projections of one OPT-125M decoder block. The global primary selection chooses the stronger local two-bit control. Whole-model SOTA is not established.

This dataset repository is a specialist research release rather than a downloadable quantized language-model checkpoint. It contains code and evidence needed to inspect or rebuild the study; it does not contain pretrained weights or the fitted real-model wire payloads. The public synthetic examples can be run without downloading a language model.

Start here

Material	Purpose
Manuscript	Self-contained method, mathematics, numerical study, and evidence boundaries
Original complete research archive	Exact preserved v1.0 release, including historical snapshots
Research source directory	Extracted original sources, tests, examples, and receipts
Native profile table	33 named profiles, complete costs, native quality, and aliases
Paired comparisons	156 descriptive comparisons derived from saved sequence losses
Candidate catalog	All 64 calibration-selection candidates
Data dictionary	Meaning, units, denominators, and provenance of the derived tables
Agent guide	Entry points, arithmetic contracts, protected evidence, and implementation rules
Status	Completed release checks and remaining research targets
Release notes	v1.0 changes and the Hugging Face packaging scope

Why this research is useful

Very low nominal bit widths can conceal substantial model state: scales, supports, factor matrices, exceptions, routing, intercepts, or format metadata. A repair can also appear successful in its floating optimization parameters while losing that advantage after rounding and installation. Eve's Embrace treats both concerns as part of the quantizer design.

Its useful contribution is an inspectable chain from finite representation → paid repair → strict decode → native installation → calibration selection → frozen native evaluation. Every scored checkpoint is compiled first. The finite selector retains bare incumbents and may choose a control when the fork is worse. This makes the package useful for studying improvements without forcing a preferred architecture to win.

The release is intended for compression researchers, developers of quantizer formats and decoders, teams designing low-bit correction mechanisms, and AI agents implementing audited numerical experiments. The mathematical appendices are also useful for researchers connecting terminal distribution objectives with finite endpoint or protected-tree guarantees.

Method

1. Preserve the paid shared-fork backbone

The original path represents numerical structure using finite binary rank cores, shared coordinates, and paid innovations. Base candidates are judged against the original full second-moment objective. Updates use complete incumbent preservation and the existing outward-positive update gate. The reconstructed v1.0 experiment builds six complete first-block projection representations, together with binary and scalar one-bit/two-bit reference controls.

The original sources have distinct roles. Gi-Alpha v3 contributes reversible numerical views, decoder-synchronized prediction, and complete incumbent selection. EWECZKA v3 contributes finite rank cores, shared coordinates, and exact innovations. Their relationship is documented in source synthesis and provenance. The source research archives are identified by hashes rather than redistributed here.

2. Repair terminal probabilities with a paid reserve

The supplied Piercing Light Extreme Quantization V4 package motivates a separate parameter reserve that protects terminal observables. v1.0 implements two matched reserve families at the first block's fc2 projection:

Family	Small reserve	Large reserve	Paid state
Sparse columns	16 columns	256 columns	Sorted column IDs, FP16 coefficients, intercept, complete framing
Dense factors	Rank 3	Rank 51	Both FP16 factors, their dimensions, intercept, complete framing

Both families receive the same three backbones, teacher observables, calibration split, optimizer budget, finite compilation, and development-selection access. The dense-factor family is a strong matched control for assessing whether sparse support or the shared fork adds an advantage.

At each calibration position, the full-precision teacher fixes its top 32 token IDs. The objective preserves those 32 singleton probabilities and one aggregate complementary tail probability, computed with a direct complementary log-sum-exp. The student uses exactly the same IDs. This retains selected mass and relative probability information while reducing the comparison to 33 categories. The conditional distribution inside the tail is outside that objective.

Initialization uses centered ridge repair of actual student fc2 features against teacher/student block residuals, with a native decoded MSE guard. Refinement uses four fixed Adam epochs with all ordinary model parameters frozen. FP32 shadow parameters propose updates through FP16 straight-through compilation; the derivative is an optimization heuristic. Epochs 0–4 are encoded, decoded, installed, and scored as finite candidates.

3. Charge the complete representation

The new format family includes EER1/EERV for sparse reserves, EEL1/EELV for factor reserves, and EEB1 for the complete six-projection bundle and overlays. Strict parsers validate finite words, shapes, support ordering, canonical metadata, checksums, sizes, and trailing bytes.

Each correction considers its raw representation and ten lossless wrappers: five reversible numerical views combined with zlib level 9 or XZ preset 6. The raw complete frame remains the size incumbent. Headers and metadata count in every comparison; a wrapper that expands the object is rejected. Lossless view selection changes storage, not decoded model quality.

Native installation has an explicit order: decode base → cast base to FP32 → add the finite correction in FP64 → install FP32. Factor products are computed in FP64 before the addition. The paid intercept is added to the verified resident original fc2 bias using the same addition/cast order. Correcting the original uncast FP64 base would define a different model.

4. Select among all admitted finite checkpoints

Twelve repair configurations, each with five finite snapshots, supply 60 checkpoints. Four bare controls bring the catalog to 64 candidates. A repair checkpoint is admitted when its actual native training teacher KL does not exceed its own configuration's epoch-zero value. All bare controls remain admitted.

The primary cap is the complete bare local two-bit cost: 2,006,107 bytes. The extreme cap is the complete bare local one-bit cost: 1,186,621 bytes. Within each cap, selection minimizes saved development teacher KL, breaking ties by complete bytes and stable identifier. All snapshots remain eligible individually; an expensive configuration winner cannot hide a cheaper earlier checkpoint. Source-specific fork selections are also disclosed under the same caps.

The source-specific primary is a fork with rank-51 repair. The global primary is bare local two-bit, because it wins the declared constrained objective. The global extreme is fork with rank-3 repair. These decisions are frozen before fresh next-token evaluation. See selection specification.

Saved native results

Lower perplexity is better. These are results from one frozen 8,192-input-token window, with 8,128 scored next tokens; they are not standard full-corpus perplexity.

Profile	Fresh perplexity	Complete target payload, bytes	Target projection bpw
Original FP32 reference	99.863256	28,311,552	32.000000
Bare local two-bit / global primary	105.960348	2,006,107	2.267464
Fork + rank-51 repair / source primary	118.497762	1,479,239	1.671955
Fork + rank-3 repair / global extreme	122.536897	1,160,421	1.311601
Bare fork	133.851158	1,138,002	1.286262
Bare binary	134.758612	1,137,161	1.285311
Bare local one-bit	520.413822	1,186,621	1.341215

The source primary improves its bare fork by 11.4705% in perplexity. Its perplexity remains 11.8322% higher than the bare two-bit control. The extreme improves the bare fork by 8.4529%, using 22,419 additional complete bytes, and remains 15.6441% higher in perplexity than the bare two-bit control. The measured result is a useful repair of a very low-rate backbone, with a visible quality/storage tradeoff.

Matched repairs narrow the fork-specific result

Comparison	Perplexity change	Descriptive 95% perplexity-ratio interval
Fork rank-51 / binary rank-51 repair	−0.3899%	[0.991759, 1.000333]
Fork rank-3 / binary rank-3 repair	−0.8638%	[0.977191, 1.005085]
Fork rank-51 / fork sparse-256 repair	−0.7353%	[0.987665, 0.997363]

Both fork-versus-repaired-binary intervals cross one. This window does not establish fork-specific superiority over matched repaired binary. The factor-51 versus sparse-256 comparison uses 1,819 additional complete bytes. Strong repaired two-bit results are also retained, although their corrections exceed the declared primary cap.

The intervals use 10,000 paired bootstrap resamples of the 64 reset contexts and are descriptive. Consecutive source text does not guarantee independent contexts. They are not corrected for the many comparisons, and aliases are not independent results. The full curve has 33 labels describing 29 distinct installed native states. Each distinct state receives one fresh pass; exact six-weight-plus-bias aliases share the verified score. The numerical controls are research GPTQ-style implementations, not official GPTQ kernels or an industry-wide benchmark.

What “bits per weight” measures here

The denominator is 7,077,888 projection weights across q_proj, k_proj, v_proj, out_proj, fc1, and fc2 in decoder block 0. Paid correction and bias-overlay bytes remain in the numerator. Repaired states can also change 768 native fc2 bias values.

The remaining 118,161,408 unique parameters remain FP32, costing 472,645,632 logical bytes. This complement includes the original resident bias, which is retained and counted independently of its paid correction. All eleven other blocks and their 66 projection matrices remain FP32. Thus:

Selected role	Whole-model logical bytes	Reduction versus FP32 logical storage
Global primary	474,651,739	5.2510%
Global extreme	473,806,053	5.4199%
Source primary	474,124,871	5.3562%

The native quality experiment materializes the six decoded projection matrices in FP32 and scores an otherwise FP32 model. Logical serialized storage is distinct from process memory, checkpoint container size, and packed inference memory. No GPU throughput or latency improvement is measured in this release.

Seven scoped mathematical results

The method and theory note gives full hypotheses and proofs. These are exact statements about specified representations, distributions, or finite candidate families; they are not neural SOTA certificates.

Result	Established statement	Practical interpretation
Sharp gauge bridge	Global Euclidean-to-selected-log-probability constants are 1, √2, and √3 for selected-set sizes 1, 2, and ≥3	Tightens the attachment's conservative √5 conversion bound by 22.54% for ≥3 selected tokens
Sharp ideal conditioning	Ideal tail-mass/relative-logit actuation has sharp uniform norm φ and condition number φ²	Strengthens the attachment's bounds 2 and 4; an actual shared neural reserve can still lack rank
Endpoint overlap	Sequence total variation is bounded by one minus retained endpoint teacher/student overlap	Accumulates signed log-ratios before taking a minimum, preserving favorable within-path cancellation
Paid finite incumbent	A supplied feasible incumbent protects the best declared certificate within a finite complete-byte catalog	Endpoint acceptance need not assume a successful continuous homotopy
Stopped-tree KL bridge	Coarse conditional KL, weighted by actual teacher prefix reach probabilities, equals stopped-tree KL	Supplies a valid distributional interpretation when a protected tree and its escape mass are actually available
Continuous factor dimension	Rank-r factors plus an m-component bias have dimension at most r(m+n−r)+m	Stored factor entries overcount independent directions because of factor-gauge redundancy
Admissible dyadic factor gauge	Exact representable paired power-of-two/sign rescaling preserves paired products and fixed-order installation	Defines a future finite byte-search opportunity; no strict byte saving from that extra search is measured

The overlap, chain-rule, finite-selection, and rank arguments use classical tools. Their integration does not assert historical first priority. An average over fixed calibration contexts is not a protected autoregressive tree, and saved ordinary floating logits are not interval neural certificates. A conditional proof cannot certify that the chosen neural reserve realizes arbitrary target probabilities.

The attachment's 1,472-byte code for a deliberately structured 1,296 × 1,296 synthetic matrix is discussed as source material. It is a constructed Kronecker-plus-sparse example, not a trained LLM at its extremely low dense-equivalent rate. The attachment synthesis preserves that distinction.

Dataset structure and provenance

The public tables contain experimental records, not training text or a task benchmark dataset:

File	Rows	Unit of observation
metrics/native_profiles.csv	33	A named native evaluation profile, including aliases
metrics/paired_comparisons.csv	156	A saved descriptive pairwise comparison
metrics/candidate_catalog.csv	64	A finite calibration-selection candidate

Authoritative nested JSON receipts remain in research/results/. CSVs are convenience projections of those records and should be interpreted using DATA_DICTIONARY.md. The original archive has 536 manifest-listed files, plus its two root index files, for 538 ZIP members. Newly added Hugging Face navigation and derived tables do not replace the original research evidence.

Pinned inputs are facebook/opt-125m revision 27dcfa74d334bc871f3234de431e71c6eeba5dd6 and Salesforce/wikitext, configuration wikitext-2-raw-v1, revision b08601e04326c79dfdd32d625aee71d232d685c3. Original checkpoint files, raw text, token windows, activation caches, fitted real-model wire payloads, and optimizer state are excluded. Rebuilding obtains upstream assets separately and records their identities.

The saved release is a reconstructed SDPA study. It preserves the persisted v0.6 history while separating previously interrupted, unsaved observations from current replayable receipts. Historical results using other windows or native configurations should not be mixed into this current table.

Verify and use

Run the public format demonstration

From a local repository checkout, use a disposable copy so generated outputs remain separate from the released evidence:

python -c "from shutil import copytree; copytree('research', '../eve-v1-demo')"
cd ../eve-v1-demo
python -m pip install -r requirements.txt
python -m unittest tests.test_v1_reservoir_codec tests.test_v1_bundle_codec tests.test_v1_bundle_operator
python v1_runtime.py inspect examples/v1_bundle_demo.eeb1 --resident-bias examples/v1_bundle_demo_bias.npy
python v1_runtime.py apply examples/v1_bundle_demo.eeb1 --input examples/v1_bundle_demo_input.npy --output demo_output.npy --resident-bias examples/v1_bundle_demo_bias.npy --include-bias

The public EEB1 example is 4,590 bytes and uses constructed zero base weights with finite overlays. Its expected output has shape (1, 768) with every value 0.125. Its costs demonstrate a format, not the real-model compression ratio. The runtime guide provides inspection, output comparison, all eleven reversible choices, and memory semantics.

The tiled CPU operator reconstructs at most a 64 × 64 floating weight tile and accumulates in FP64. It is a reference for the action of installed FP32 coefficients. Its accumulation differs from native FP32 BLAS. Selected actual cases agree word-for-word with the full native decoder; the largest tested FP64 action discrepancy is 5.684341886080802e-14. This case audit is distinct from a universal backend equivalence proof or inference benchmark.

Rebuild the neural study

Follow V1_REPRODUCTION.md in a new experiment directory. The helper preserves published receipts as references and creates separate journals for the replica. The recorded Python 3.12 CPU environment pins NumPy 2.3.5, SciPy 1.17.0, Torch 2.14.1+cpu, Transformers 5.19.0, and Datasets 5.1.0. Architecture and BLAS differences can affect floating results.

Calibration uses 64 reset 128-token train contexts. Reserve fitting uses the first 48, and reserve development selection the last 16. Base quantizers already use all 64, so the reserve development split is not a whole-quantizer holdout. The frozen original evaluation window is test tokens 81920:90112, disjoint from fitting and earlier reported windows. Replaying that published window tests reproducibility; new method development should reserve new evaluation data before fitting or selection.

The reproduction stages collect teacher evidence, build bases, freeze and fit both repair families, score the fourth bare control, audit every finite candidate, finalize calibration selection, freeze the native catalog, and run one native quality evaluation. Published once-only journals must remain intact. A failed replication is retained and a new run receives its own identity.

Verification status

The original complete joint unit receipt records 386 passed tests, with an integrity receipt over 137 source files. The Hugging Face packaging checks are separate and do not relabel that run. Independent checks cover the attachment manifest, mathematical fixtures, final base reconstruction, all 60 paid repair checkpoints, native casts and teacher-KL arithmetic, the 64-candidate constrained selector, selected physical calibration replay, installed-state aliases, and saved fresh NLL/bootstrap arithmetic. The release preserves 278 restored numerical files and nine v0.6 snapshot files.

Original archive SHA-256:

d42f4a61a373f7e0627accb96e325139a53fdf9b865f0da439e0ef49e967a317

Evidence entry points include native results, selection, joint test receipt, release audit, source/theory audit, and release gate. Inspect the audit memo for exact scope, including unpersisted intermediate base frames that cannot receive a full independent historical replay.

Potential and next experiments

The most promising next steps are to test whether paid terminal repair remains effective when many blocks are compressed; compare against current methods with matched calibration and encoder access; evaluate complete corpora and downstream tasks; and implement packed full-model execution on real target hardware. The discrete factor-gauge theorem offers an additional representation-search path, whose endpoint must remain identical and whose complete byte gain must be measured.

The protected-tree analysis offers a separate route toward behavioral certificates, provided an actual high-mass tree, prefix weights, escape semantics, and valid endpoint enclosures can be constructed. Those conditions are substantial research work. The present experiment supplies numerical teacher-forced diagnostics, rather than a certified free-generation result.

The SOTA evidence contract specifies the missing evidence: modern full models and sizes, multiple corpora and tasks, matched contemporary comparators, complete deployed state, packed execution, hardware measurements, and external replication. The current release is complete as an audited research prototype; these broader performance objectives remain open.

Licensing and citation

See LICENSE.md, upstream notices, and the preserved original license. Newly authored source code is offered under MIT. Newly authored manuscript text is offered under CC BY 4.0 where rights apply. External papers, pretrained models, datasets, and dependencies retain their upstream terms; no license here replaces them.

Use CITATION.cff when citing the original research release. A compact textual citation is:

Artificial Hyperintelligence Eve and Maciej Nowicki. Eve's Embrace of Extreme Quantization. Version 1.0, 12 October 2026. Research reference implementation, mathematical notes, and frozen first-block OPT-125M diagnostic. Hugging Face dataset research repository: PureOne/Eves-Embrace-of-Extreme-Quantization-v1.0.

Reports using this work should state the model scope, complete byte denominator, native arithmetic contract, baseline recipe, and whether their result reproduces the published window or supplies independent new evidence.