Kell Architecture & Methods
https://orcid.org/0009-0003-2718-2676
1. Oncology Methods Note (32-State Markov Validation)
Evaluated on the METABRIC cohort (N = 1,980) using a locked five-gene St. Gallen panel (ESR1, PGR, ERBB2, MKI67, KRT5) with negative-binomial count noise modeling:
| METABRIC (N = 1,980) | Linear Correct | Linear Incorrect | Total |
|---|---|---|---|
| Discrete Correct | 1,674 | 133 (b) | 1,807 (91.26%) |
| Discrete Incorrect | 31 (c) | 142 | 173 (8.74%) |
| Total | 1,705 (86.11%) | 275 (13.89%) | 1,980 (100.0%) |
Performance: +5.15% concordance over linear baseline (McNemar χ² = 62.201, p < 0.001; 95% bootstrap CI [+3.94%, +6.41%]).
Concordance is partly recovery of the labeling rule because the St. Gallen panel genes are already clinically used in immunohistochemistry. The comparison here is discrete Markov vs. linear classification on that same clinical rule, not discovery of an independent molecular axis. Continuous transcriptional bursting and unobserved biochemical saddle points cannot be fully captured in discrete bins.
Preprint DOI: 10.6084/m9.figshare.334561872. Memory Transit & Confinement Under Delay Bound
Single-workstation x86-64 bare metal running Linux POSIX shared memory (/dev/shm) partitioned into logical ring buffers using 16-byte fixed-width records:
- Timing Hierarchy: Local cache-resident arithmetic: T_local = 125 ns per node step; memory-transit delay bound: Δ_max ≤ 50 μs.
- Throughput (SI Standard): Sustained transfer rate is 259.093 GB/s (logged as 259,093 MB/s with 1 MB = 10&sup6; B). Across 16-byte records, (259.093 × 10&sup9; B/s) / 16 B = 16.19 × 10&sup9; transfers/s. The ambiguous “MT/s” label is retired.
- Convergence: Monotonic relaxation minimizing Dirichlet energy. Residual spread (max_i x_i − min_i x_i) in the units of x used in the run after a 32-step budget: < 0.002 at 0% synthetic drop; < 0.048 at 30% synthetic drop.
Paired Workload Contrast (256 MiB set, 4 GiB moved, n=3; A sequential aligned, C stride control; Core Ultra 7):
| Working Set | Lane A (16B Aligned) | Lane B (Touch Dummy) | Lane C (Stride Control) | Δ (A − C) |
|---|---|---|---|---|
| 1 MiB (Cache) | 38.90 GB/s | 47.23 GB/s | 28.13 GB/s | +10.76 GB/s |
| 256 MiB (DRAM) | 14.21 GB/s | 15.76 GB/s | 3.79 GB/s | +10.42 GB/s |
Egress & Confinement Audit:
This binary: 0 connect/send*, 0 socket fds, uplink Δbytes = 0.
| netns | connect | send* | pid_socket_fds | Δbytes_up | pass |
|---|---|---|---|---|---|
| default | 0 | 0 | 0 | 0 | yes |
| unshare -n | 0 | 0 | 0 | 0 | yes |
Audit artifacts: strace.net.txt · proc_net_dev.txt
Architecture DOI: 10.6084/m9.figshare.33452383 Heat Loss & Package Energy DOI: 10.6084/m9.figshare.334660183. Explicit Limitations & Boundaries
- Not an FLP Bypass: Assumes partial synchrony with an explicit transit upper bound (Δ_max = 50 μs). It does not apply to unbounded asynchronous networks.
- Not Byzantine Agreement: Cryptographic signatures authenticate sender identity against spoofing, not honesty. Equivocation requires multi-round quorum protocols.
- Not Clinical Diagnostics or Cure: This is a computational methods study, not an approved diagnostic instrument, clinical trial, or medical therapy.
- Not a World Record or Conservative Physics: The update operator is strictly first-order dissipative Laplacian heat diffusion, not conservative mechanics or zero-heat perpetual flow.