Research evidence dataset. No model weights. Part of the collection Xyntetik Research: GENESIS (https://huggingface.co/collections/Joakimpalm-Zen/xyntetik-research-genesis-6abcfa50b6707dec9bfee356). Not machine-only: the scanned models were trained on human text, and humans chose the patterns, the diagnostic, the rewrite rule, the controls and every threshold, so nothing here may be combined with GENESIS's machine-only claims. The complete record of the genesis-rewrite track (RW-001 to RW-005b): the intake, the prior-art audit, six preregistrations, six reports, the code and the raw scan records, failures included. The unflattering result comes first: neither candidate pattern was found in any trained model scanned, so no real model was rewritten.

Xyntetik GENESIS: Rewrite record

Dataset summary

Label: not machine-only. The models scanned were trained on human text. The machine only did the measuring.

Question tested. The owner's brief, paraphrased in the record: can inefficiencies that optimization leaves in a trained network be found by a diagnostic and algebraically rewritten into a smaller, faster computation that preserves capability, with no gradient-based retraining, distillation or recovery fine-tuning? Pruning, merging, low-rank compression or a benchmark gain on its own does not count. Two candidate patterns were tested:

The odd unit pair (RW-001 to RW-002d): two units of one layer with antiparallel input weights (bias included) and output weights that cancel with the reciprocal scale. Because φ(z) − φ(−z) = z for ReLU, SiLU and GELU, the pair computes a linear function and can be replaced, exactly, by one linear unit. Look-alikes that do necessary computation (|z| pairs, offset ramps) must not be flagged.
Cross-layer write-and-cancel (RW-005, RW-005b): a pair of units in two layers whose writes to the residual stream cancel on real activations.

Models involved. No model was trained or changed. Weights were read, never redistributed:

RW-001: small MLPs trained here on a synthetic teacher, with planted pairs and look-alikes.
RW-002: Qwen3-0.6B (the registered BF16 file, bitwise equal to the publisher's snapshot), with Qwen3-0.6B-Base and Qwen3-1.7B as descriptive checks.
RW-002b: Pythia-160m, OPT-125m, Gemma-3-270m, Granite-4.0-350m-base and the Granite-3.1-1B-A400M MoE.
RW-002d: GPT-2, SmolLM2-135M, CodeGen-350M-mono, BLOOM-560m, TinyLlama-1.1B-3T, Falcon-RW-1B, Phi-1.5, OLMo-2-0425-1B and OLMoE-1B-7B-0924.
RW-005: GPT-2, Pythia-160m, OPT-125m, SmolLM2-135M, Granite-4.0-350m-base, CodeGen-350M-mono, Qwen3-0.6B-Base, BLOOM-560m and TinyLlama-1.1B-3T, with 2,048 calibration tokens of wikitext-2 validation text per model; RW-005b: Falcon-RW-1B and Phi-1.5.

Every model is pinned by repository, revision and weight sha256 in its preregistration.

Method.

Preregistration. Every stage was preregistered and pushed before its data, with dry runs on random or unregistered models disclosed. RW-001 and RW-002 were approved by the owner; RW-002b reopened the track on other families at the owner's instruction, and RW-002d to RW-005b ran under the owner's standing pre-approvals, given before each draft existed (paraphrased in this copy).
The diagnostic. For odd pairs, s = ρ_in + ρ_out, with RW-001's thresholds fixed on development networks before any held-out network existed: θ_safe 0.04369 (the instance criterion) and θ_BA 0.06471. Input norms are folded into the input rows; gated MLPs and MoE experts use the gated form. For write-and-cancel, R is the RMS of the two writes' sum over the smaller single write's RMS, with an instance at R ≤ 0.1 and both units alive.
Nulls and controls. Null families of 999 draws (shuffled, sign-randomised, initialisation, re-paired or token-permuted, per stage), positive controls planted in copies of real layers, and family-wise alpha 0.01 over each stage's models.
Compute. CPU only, except RW-002, which used one GPU process.

What each file contains.

record/ is the track directory as on the branch, apart from the edits in PUBLIC-COPY-CHANGES.md:
- README.md, INTAKE-2026-10-06.md (the brief, paraphrased), PRIOR_ART.md (with its 2026-10-14 addendum), PARENTS.md and RESEARCH_STATE.md;
- prereg/RW-001.md to prereg/RW-005b.md, with every model's pins and licence and the dated amendments;
- report/RW-001.md to report/RW-005b.md, rendered from each stage's decision.json, unflattering results first;
- the code: diagnostics, scans, nulls and analyses (rw00*_core.py, rw00*_run.py, rw00*_analyze.py), the self-committing chains and gates (*.sh), and the unit tests (test_rw*.py).
runs/RW-001.tar.gz to runs/RW-005b.tar.gz each restore record/runs/<stage>/: decision.json, the per-model scans and null draws (summary statistics of weights and activations: pair scores, cosines and ladders, no tensors), Phase 0, the pins check (preflight.json) and the chain logs.
PUBLIC-COPY-CHANGES.md lists every difference from the branch. SHA256SUMS hashes every file except itself and this card.

Reproduction.

Restore the runs: tar -xzf runs/<stage>.tar.gz -C record.
Re-read a verdict from the archived records with the stage's analysis script (rw00*_analyze.py), as each chain (rw00*_chain.sh) calls it.
Rerun a scan: fetch the registered weights at the pinned revision (sha256 in the preregistration) and run the stage's runner on CPU. Python 3 with numpy and torch; the larger scans took hours.

Result. The unflattering result comes first.

No natural instance of either pattern in any trained model scanned.
- Odd pairs: none at θ_safe in Qwen3-0.6B (smallest s 1.30; none even at s ≤ 0.5), its base model or Qwen3-1.7B (RW-002, so the candidate stopped), nor in the 14 further models of RW-002b and RW-002d. The smallest s per model ranged from 0.2816 (OLMoE-1B-7B) to 1.318 (TinyLlama). Phi-1.5's RW-002d verdict was first INCONCLUSIVE on an instrument check (its descriptive sign null was stopped by hand) and read NO NATURAL INSTANCE after RW-005b's label repair, which added no null information.
- Write-and-cancel: no cross-layer pair at R ≤ 0.1 in any of the 11 models of RW-005 and RW-005b (the smallest recorded cross-layer R was 0.1637, in SmolLM2-135M), so that candidate stopped too.
The structure that does exist leans the other way. In Qwen3-0.6B the most antiparallel gate pairs carry parallel, |z|-like outputs, not cancelling ones: 22 of 28 layers, two-sided sign test p = 0.0037 (a registered secondary reading; ten post hoc initialisation draws showed no consistent direction). The sign-randomised null produced more antiparallel gate rows than the trained model.
What passed is a stage gate on planted pairs (RW-001). In 12 held-out small MLPs the diagnostic separated rewritable odd pairs from look-alikes with balanced accuracy 0.994 [0.988, 1.000], where a per-unit view scored 0.500 and an input-only view 0.860. The exact rewrite's worst relative error was 7.4e-16 over 174 planted exact pairs, and joint rewrites kept the error at most 0.016 against a 0.03 bar. The MAC count fell from 19,584 to 18,944: counted, not measured. These networks themselves contained no natural instance.
The positive controls passed: in RW-002, an exact pair planted in a copy of a real Qwen3 layer was found at s = 0 and flagged exact; RW-002b and RW-002d record their positive-control checks (I3) as passed for every model.
ATLAS (a re-pricing in its separate not-machine-only block): AT-001b kept the nulls of RW-002 and RW-002b.

Limits, read before quoting

Two hand-chosen patterns, small and mid-sized open models only. A null here says these exact patterns, at these thresholds, were not found in these checkpoints. It does not say trained networks hold no rewritable redundancy.
No rewrite of a real model was made, so nothing here is a speed, memory, energy or capability result. Counts of units or MACs are arithmetic.
No novelty is claimed (record/PRIOR_ART.md): the narrow open point was a weight-only diagnostic for odd pairs, validated against look-alikes, and a count of natural instances against a null.
Not machine-only. Pretrained models trained on human text were scanned; keep these results apart from GENESIS's machine-only claims.

Licences

This record: Apache-2.0 (the project's code, text and measurements).
No weights, model outputs or dataset text are included. The scan records hold summary statistics of each model's weights (and, in RW-005 and RW-005b, of its activations on wikitext-2 validation text, which is not included).
The scanned models' licences, as the preregistrations record them:
- Apache-2.0: Qwen3-0.6B, Qwen3-0.6B-Base, Qwen3-1.7B, Pythia-160m, Granite-4.0-350m-base, Granite-3.1-1B-A400M, SmolLM2-135M, TinyLlama-1.1B-3T, Falcon-RW-1B and OLMoE-1B-7B-0924; OLMo-2-0425-1B is Apache-2.0 with Ai2's Responsible Use Guidelines;
- MIT: GPT-2 and Phi-1.5; BSD-3-Clause: CodeGen-350M-mono;
- OPT-125m: the OPT-175B License Agreement, non-commercial research use;
- BLOOM-560m: the BigScience RAIL License v1.0, with use-based restrictions; its weights were only read;
- Gemma-3-270m: the Gemma Terms of Use, read through a mirror (unsloth/gemma-3-270m) whose weight file has the same sha256 as Google's; no gated terms were accepted.

About Xyntetik

This repository is published by Xyntetik, the Zenova AB program for running capable language models on hardware you own and proving what they did. Its released engine is Xyntetik Runner: a free, single-binary GGUF engine in C for CPU, CUDA and Metal that serves, scores and trains through the quantized file it deploys and records every run as a receipt. The measurements behind the program's claims are collected at xyntetik.com/evidence. The program's organization on Hugging Face is Xyntetik.

Downloads last month: 216

Total file size:

11.2 MB

Collection including Joakimpalm-Zen/Xyntetik-Genesis-Rewrite-record

Xyntetik Research: GENESIS

Collection

Genesis: preregistered research on machine-only learning in three tracks (language, existence, crossing), failures included. Research artifacts only. • 11 items • Updated 19 days ago