Joakimpalm-Zen/Xyntetik-Genesis-Native-record

GENESIS: Native, the record (MN-000, MN-000b and their addenda)

Research evidence dataset. No model weights. Part of the collection Xyntetik Research: GENESIS. Everything here was preregistered before its data. Amendments are dated, and unflattering results come first.

Dataset summary

The thesis (full text in THESIS.md). A machine-native model is designed from scratch around one specific computer, using its measured cache, bandwidth and processors, rather than around a class of devices. The thesis has six pillars: regenerated weights, a looped core sized to the cache, a knowledge store, parallel decoding, device-shaped design, and a personal model. This track tests one narrow slice: (b) a looped, weight-shared block sized to a cache, inside (e) device-shaped design. The test is a transfer test. A shape fitted to one machine should do better there and lose that edge on another machine.

What is new and what is not (PRIOR_ART.md, verified). Device-specialised models losing on other devices is published: HAT, ProxylessNAS, FBNet, HAQ and HW-NAS-Bench, 2018-2021. What stays unclaimed is narrow:

• cache capacity as the varied host property;
• a looped block's size as the knob;
• toy-scale language-model cores;
• quality and speed measured on more than one host.

Hosts. Single-threaded CPU on three machines:

• A: AMD Zen 5 (1 MB L2 per core, 32 MB L3 per 8 cores);
• B: Intel Kaby Lake (256 KB L2 per core, 8 MB L3);
• C: Apple M1 (12 MB L2 shared by the performance cores).

MN-000: the instrument and the mechanism. Verdict: GO (report/MN-000.md)

No training. It compares a looped block (one block, three passes) with three unique blocks at equal FLOPs, at widths 32-1024, on random weights from fixed seeds.

• The timing instrument works on all three hosts. The A/A half-width is 0.2-1.2% against a 3% limit. Weights are bit-identical across x86 and arm64, and logits agree to 3×10⁻⁷.

• The mechanism appears only when one row is decoded at a time. At batch 1, R (looped over unique, GFLOP per CPU-second) peaks at a different width on each host:

  host	peak width	R at the peak
  B	256	1.15
  C	384	1.73
  A	640	2.64 (a plateau of 2.5-2.7 over 512-704 in a quiet repeat)

  The peaks follow each host's largest cache: 8, 12 and 32 MB. At 64 rows the effect is at most 10%. At the language search's own batch size (4,096 rows) it is absent.

• The weak point, stated first in the report: the preregistered knee rule separated the hosts only through a censored knee. It was replaced, by the owner's decision, with "the width where R peaks".

MN-000b: the pilot. Verdict: AMEND (report/MN-000b.md)

The core's own shape (G0s), Looped and Unique at widths 384 and 640, and the published frozen core were fitted to GENESIS's published replay buffer. They were then searched one program at a time, at 30 CPU-seconds per task, on the same 300 pilot tasks on hosts A and C.

• Power fails at every N ≤ 3,000. Charging a full training run per task leaves no search budget.
• Speed in the sampler does not follow R. Looped/Unique programs per CPU-second is 2.09 and 1.23 on A, and 1.08 and 0.59 on C.
• At width 384, looping helps on A and hurts on C (+15 against −12 solved, out of 300). A seed replicate on C repeats C's sign (−0.040 and −0.050), each interval touching 0. Pilot numbers, one seed per shape on A.
• Every refitted shape solved far fewer tasks than the published core (72-107 against 149 on A).

What each file contains.

• THESIS.md, INTAKE-MN-2026-10-09.md, PRIOR_ART.md: the claim, the scope and the owner's decisions, and the verified literature.
• prereg/: MN-000 and MN-000b with every dated amendment, addendum and pinned version.
• report/: both reports in the theory-open order.
• runs/: the analysis summaries and reports (MN-000 on three hosts; MN-000x fine grids; the cost probe on A and C; MN-000b on A and C; the seed replicate).
• code/: the harness (profiler, cost probe, pilot fits and searches, analyses, runners and tests).

Reproduction. The code runs on CPU with torch 2.11 and numpy. The weights and inputs come from numpy seeds and are identical on every host. The full run records, with raw per-cell timings, are on the GENESIS research repository. Host-identifying fields are left out of this public copy.

Limits, read before quoting.

• Toy scale: sub-million to 15M-parameter cores.
• One task family: GENESIS's machine-generated DSL tasks.
• Single-threaded CPU only.
• The pilot's rates are discarded and feed the power planning only.
• Nothing here says anything about billion-parameter models, GPUs or NPUs.

Status. MN-001, the confirmatory transfer test, is not designed yet. Its currency and budget are open decisions.