elbow v2 · HEAD-TO-HEAD
the honest head-to-head · v2

SILO vs PLAIN

v1 asked it; v2 measures it. The silo (cluster the context to K intents, then transform) against the plain transformer (attend over all N tokens). These are untrained toys, so there's no honest "accuracy" to score — instead, the industry-standard tests for compression: the elbow method (variance explained by K) and rate vs distortion (attention N²→K² saved, against fidelity lost). A trade-off, not a winner.

The comparison

Sweep K · read the elbow

The plain transformer is the K = N end (no clustering, 100% variance, output agreement 1.0 — a self-test anchor). Everything at K < N is the silo trading fidelity for compute. Type a context; the sweep, the elbow, and the verdict update live and deterministically.

◎ sweep + verdict — the elbow method + rate–distortion, computed client-side (no network)
keep ≥
90% variance
variance explained (fidelity) attention cost K²/N² (rate) the elbow (recommended K)
Variance explained = 1 − WCSS/TSS (the elbow-method / k-means metric). Rate = self-attention pairs, N² (plain) vs K² (silo); the silo also pays a one-time N·K·spins clustering pass, disclosed in the table. Output agreement = cosine between the silo's pooled output and the plain transformer's. At K=N the silo is the plain model (agreement 1.0). Untrained toys — no accuracy claim; just the trade-off, measured right.
The colophon

What this settles — and what it doesn't

What it settles

The silo's cost/benefit is a genuine rate–distortion trade-off: it buys a K²/N² attention reduction for (1 − variance explained) lost. The elbow shows where the curve turns. python selftest.py proves the metrics and the anchors (K=N ≡ plain, agreement 1.0) with no network (14 checks).

What it doesn't

It does not claim the silo is better — the self-test literally forbids the verdict from saying "beats" or "wins." Whether the compression helps a trained model on real data is a different, empirical question. That would be v3: real training, a real benchmark, honestly reported.

SOURCE · github.com/DavidWise01/transformer-silo-v2 · bench.py (explained_variance · attention_pairs · compare · sweep · verdict) · silo.py (v1 engine, vendored) · selftest.py (metrics + honest framing). Head-to-head of transformer-silo v1.
⧉ the code (GitHub) ⌗ silo v1 ▲ v3 · trained + tested \o/ the forward pass ⌂ ud0 · the atlas