\section{Implementation}
\label{sec:implementation}

\titan{} is about 1\,100 lines of Python over MLX 0.32.4 and mlx-lm 0.32.2 (Python 3.12), organised as
\code{titan/\{registry, store, slots, moe, prefetch, predictors, runtime\}.py}, with benchmark drivers in
\code{benchmarks/scripts/} and tests in \code{tests/}. Three implementation findings determined the design.

\paragraph{In-place slot updates.} Writing the nine tensors of one OLMoE expert (3.4~MB) into stacked buffers of 64 slots
(216~MB per matrix) costs 0.305~ms, which shows that MLX reuses the buffer rather than copying it; a per-expert
\code{quantized\_matmul} loop costs 0.43~ms per layer for one token against 0.26~ms for \gatherqmm{}, and restacking eight
experts per call 0.50~ms. The slot table with \gatherqmm{} is therefore both the fastest and the most faithful option
(\code{benchmarks/raw/gateB/design/}).

\paragraph{No per-miss synchronisation.} The first version evaluated the table after every miss; removing that
synchronisation and batching the reads of a layer's misses raised decode throughput at 16 slots from 28 to 51~\tps{}
(general prompts) and from 41 to 73~\tps{} (code), with the stall halved
(\code{benchmarks/raw/gateC/prefetch\_online\_laptop/}). The per-miss cost on the page-cache tier is now 0.21~ms: 0.13~ms
of \code{pread}, 0.06~ms of array creation and the remainder in the lazy install.

\paragraph{Activation parity.} An unfused float16 $\mathrm{up}\cdot\sigma(\mathrm{gate})\cdot\mathrm{gate}$ gave
$\max|\Delta\mathrm{logit}| = 0.0195$ against mlx-lm; using mlx-lm's compiled SwiGLU gives $0.0$. Numerical parity in float16 is
a property of the exact kernel sequence, not only of the weights, and the test suite pins it.

\paragraph{Threads and dtypes.} Two defects were found by the exactness tests rather than by inspection. MLX arrays created in reader
threads cannot be evaluated in another thread (streams are thread-local), so reader threads now return NumPy views and the consuming
thread creates the MLX arrays. And the first transport cast activations to float16 on the wire, which is exact for OLMoE (float16) but
not for Qwen3 (bfloat16): the first two-node sweep on Qwen3 showed $\max|\Delta\mathrm{logit}| = 0.65$ and was discarded; activations and
results now travel in the model's dtype and the Qwen3 two-node path is bit-identical again.

\paragraph{Reproducibility.} Every benchmark writes JSON lines with the command, git commit, host, chip, memory, macOS,
library versions, seed and timestamps; raw logs are append-only under \code{benchmarks/raw/} and tables and figures are
generated from them by scripts. Model revisions are pinned by commit hash in \code{benchmarks/configs/models.yaml};
weights are never stored in the repository. A node is prepared with \code{scripts/cluster\_setup.sh} (rsync of the code,
pinned virtual environment, import check), and the sanitized inventory \code{artifacts/environment.json} is generated from
raw probes that are kept out of version control.