# Source synthesis for Eve's Embrace of Extreme Quantization

## Evidential boundary

The two attached releases are implemented **lossless byte codecs**. Neither quantizes neural weights, measures LLM perplexity, evaluates downstream tasks, implements a quantized matrix kernel, or establishes an industry PTQ improvement. Their striking compression figures concern deliberately structured numeric/fork sources. They support mechanisms and correctness arguments for a new proposal, not the proposal's performance claims. A tiny-model experiment can validate implementation and expose failure modes; it cannot certify world leadership.

## Precisely transferable source mechanisms

| Source | Exact location | Transfer | Required qualification |
|---|---|---|---|
| Gi-Alpha | LaTeX Section 3, `Causal plastic predictor: precise construction`; `gi_plasticity.py` `_plastic_loop`, `plastic_transform`, `plastic_inverse` | Decoder-synchronized expert selection and residual coding; context-local surprise-controlled decay | Adaptation must use **decoded quantized labels**, never original floating weights or calibration-only quantities at decode time |
| Gi-Alpha | Section 4, `Second-order Planck-bit curves: reversible numeric operators`; `double_delta`, `word_predict` | Exact lane organization, modular first/second differences and XOR residual candidates | Adjacency/smoothness of neural weight arrays is not established; try and price, do not assume |
| Gi-Alpha | Section 5, `Exact supporting proxies, bounded patches and reference DAGs`; `patch_encode`, `patch_decode` | Prior quantized block prediction plus exact sparse label innovations | Every core/reference must be present or explicitly conditioned on; avoid comparing a small delta with a self-contained model |
| Gi-Alpha | Section 6, `Full-archive selection and non-regression theorem` | Select among complete valid representations, including baseline | The guarantee is over actual included candidates and actual wrapper cost; it does not guarantee speed, global optimum, or validation accuracy |
| EWECZKA | `docs/FORMAL_DERIVATIONS.md` Theorem 1; manuscript chapters `4. Xorilithry...`, `5. Global cores...`; `ewczka.py` `factor_rows`, `_encode_tile` | Exact GF(2) rank factorization of packed bitplanes with explicit basis blocks and coordinates | Real-valued numerical rank and GF(2) label-bit rank are different. Typical random bitplanes are full rank |
| EWECZKA | Theorems 2 and 6, manuscript finite rank-family chapters | A conditional coding budget and source-family certificate | Principal cost is r(m+q-r) **bits**, plus anchor indexing, framing, padding, scales and tails. Family lower bounds are not individual-file lower bounds |
| EWECZKA | Theorem 3, `Cost-optimal paid cores...`; Chapter `6. Vexhyrration...` | Search exact algebraic gauges and orientations by measured total bytes | Greedy optimality applies only to additive basis costs with fixed mask costs; compressed-factor optimum is not proved |
| EWECZKA | Theorem 5, `Sparse innovations...`; Chapter `7. Sparse residuals can destroy low rank` | Rank-core-plus-exact-sparse-innovation candidates | Small sparse noise can create full GF(2) rank; robust clean-factor discovery is not automatically polynomial or globally optimal |

## Strongest defensible joint method

The credible synthesis is an **activation-aware rate-distortion quantizer with exact structural coding of its labels**. Separate the two guarantees:

1. Floating weights W are approximated by dequantized weights Q. This is the lossy PTQ step and requires independent accuracy/perplexity testing.
2. Q's quantization labels, scales, zero points, codebook mapping, factor records and exact innovations are serialized without further loss. This is the inherited exact-code step and admits byte-for-byte reconstruction tests.

For one linear layer and calibration inputs X, set H=XX^T/n and D(Q)=tr((Q-W)H(Q-W)^T). Optimize a concrete finite candidate family under either a distortion cap or a Lagrangian `D(Q)+lambda * actual_archive_bits(Q)`. It is crucial to include metadata and all innovations in the bit cost. Entropy, GF(2) rank, or zero counts alone are not the final objective.

### Zero-distortion label gauges

A b-bit quantizer has K=2^b numeric levels. Choose a bijection p on the K labels and invert it at decode (or permute the stored level table). The real dequantized weights remain **identical**: only their storage labels change. Search all 24 permutations at b=2; at b=3, all 40,320 permutations are finite but a bounded search can be practical. The selected permutation or its index must be paid in the archive.

Evaluate each label gauge under raw bit packing, Gi-like causal/modular/XOR residual transforms, and EWECZKA-like exact tiled GF(2) factorization. All 24 permutations at **b=2 are affine GF(2) maps**: AGL(2,2) has 4*6=24 elements. Their linear components preserve rank; translations can change unaugmented rank by at most one, and preserve the homogeneous augmented rank. They can still alter byte serialization and conventional compressor behavior. At b>=3, genuinely nonlinear label permutations exist and can change bitplane rank more substantially. A Gray-like label gauge can improve correlated ordinal labels, but IID weight-label entropy gives no expected miracle.

This is a useful, implementable synthesis that preserves the original quantizer's measured distortion and may reduce deployment **storage**. Whether this precise integration is historically new requires prior-art comparison with entropy-constrained quantization and coded quantizers; the mathematical components themselves are classical.

### Exact distortion increments for compression-aware rounding

Write E=Q-W. Changing one dequantized weight Q_ij by delta has exact calibration increment

`Delta D = 2 delta (E H)_ij + delta^2 H_jj`.

Thus a bounded coordinate refinement can test nearby quantization labels and accept changes using `Delta D + lambda * Delta actual_bits`. Recompute or update the affected row's `E H` after each accepted move. This derives the score from calibration geometry, while Gi supplies causal coding and EWECZKA supplies paid core structure.

To avoid expensive whole-model recompression on every trial, use a local estimate only for proposal screening, then compare the **actual** serialized touched tile before accepting a candidate. Surrogate bit-cost reductions are not evidence of real compression. A simple distortion cap can retain the baseline whenever no improved feasible candidate exists.

### Optional exact sparse exceptions

The compact code can permit a few high-sensitivity entries to retain different labels or higher precision. Their positions and values must be paid, with sorted sparse patches or enumerative support. This corresponds to source innovations, but choosing outliers by activation-weighted loss is a new task. It has strong prior art in sparse outlier PTQ and must be compared with those methods.

## What cannot legitimately be inferred

- The Gi float32 waveform reduction does not forecast FP32 neural weight compression; source smoothness was synthetic and explicitly identified.
- EWECZKA's constructed 178-fold compression does not forecast LLM compression; hidden planted XOR cores are absent from its real-file controls.
- A shared hash seed cannot represent arbitrary trained weights; the innovation retains any independent information.
- Exact coding of an already quantized model does not improve that model's accuracy. It only improves bytes if label structure exists.
- Decompressing every tensor to FP16/FP32 restores the corresponding resident memory cost. A disk archive benefit is distinct from packed resident memory and fused inference benefit.
- Calibration loss dominance does not imply held-out perplexity or downstream-task dominance.
- Finite candidate fallback can prove nonregression of that candidate objective, not superiority over every untested GPTQ/AWQ/QuIP#/AQLM/GGUF configuration.

## Minimal falsifiable evaluation

Report true parameter count and full archive bytes/bits per parameter including scales, mappings and sparse corrections. Compare at equal total bytes and a second equal-distortion view. Retain a standard nearest/activation-aware baseline, causal-only ablation, factor-only ablation, label-gauge-only ablation, and rate-distortion-refinement ablation. Use a held-out activation set. Include IID Gaussian/random-weight controls to expose missing structure. Verify every packed archive reconstructs exactly the candidate's quantized tensor. LLM evidence ultimately needs pretrained language-model perplexity, task scores and actual memory/latency measurements; synthetic layers are mechanism studies.

## Potential scientific contribution

The source-faithful contribution would be to **change the PTQ optimization objective from local numeric approximation alone to activation-weighted approximation plus a fully paid structural decoder description**, with a distortion-preserving label-gauge stage and explicit residual patches. The decisive empirical question is whether trained LLM labels have, or can acquire within an acceptable distortion budget, enough causal/GF(2) structure to pay for the decoder factors. Negative results are scientifically useful: they would locate the practical boundary of transferring universal byte-codec ideas to learned neural weights.