For visitors who'd rather read than walk.
# Mr. Vega and the closet Mac mini
You're Mr. Vega, school district administrator. 4:37 PM Tuesday.
Buses left at 3:45. There's a Mac mini in a closet doing nothing
useful. The board wants AI in classrooms. A cloud subscription
would eat 40% of your discretionary tech budget.
The hook on Merkle Trust's landing: there is a third option, and
the math behind it is something a high schooler can understand.
The walk for this card is not really an install conversation. It's
a sizing conversation. The deployment shape is local — your closet,
your hardware, your district. The four canonical paths exist for
the substrate; for a district running classroom AI, the practical
order is GitHub-first because the district's IT team already
controls the closet machine.
The interesting decision is which model fits, not which operator
to call.
BitNet uses ternary weights — each weight is +1, 0, or 1. That's
three states, which means each weight carries log2(3) ≈ 1.585 bits
of information. Call it 1.58 bits per digit.
The arithmetic that comes out of this is unusual:
```
═══════════════════════════════════════════════
TERNARY MULTIPLICATION — the four cases
═══════════════════════════════════════════════
+1 × +1 = +1 (no surprise)
1 × 1 = +1 (negative × negative)
1 × +1 = 1 (sign flip)
0 × anything = 0 (sparse path)
No floating-point multiply.
Add, subtract, or skip.
Runs on a CPU.
═══════════════════════════════════════════════
```
Standard 7B model: 14–28 GB on a GPU. BitNet 7B: ~1.4–2.5 GB on a
CPU. The closet Mac mini has 16 GB of RAM and an M2 chip. The
sizing panel returns:
- BitNet 3B: ~6 tokens/second, fits comfortably with headroom.
- BitNet 7B: ~3 tokens/second, fits with headroom.
- Standard 7B: will not fit.
Annual electricity to run BitNet on the M2 mini: single dollars.
Annual cloud-LLM subscription cost the district avoids: roughly
$48,000.
Yes — and the math is checkable. The BitNet paper is public, the
ternary-quantization implementation is open, the runtime isbitnet.cpp. The card links the math directly: log2(3) is
1.58496…; the four-case multiplication is the entire kernel; the
sparsity pathway (0 × anything = 0) is what makes the inference
cheap on CPU.
The substrate's role here is not the model. The substrate is what
the wrapper around the model anchors to: the model's binary hash
sealed at install, the wrapper's configuration sealed, the
inference logs sealed, the soul-chain seven-file SHA verified at
boot. The model runs locally; the attestation that it ran
correctly anchors to the chain.
The .md button puts the sizing math summary into your
tag-along bundle. Comment field routes a board-budget-style
question to your own claude.ai session.
There is no traditional ceremony in this card. The ceremony shape
for a district AI deployment is the wrapper handoff: the substrate
hands the model + the runtime + the seven-file soul chain to the
two Garrison Node council members who do this work — Ember and
Hearth, council seats 16 and 17. Their job is to seal the
deployment shape and report on its state.
The ticker streams the seven SHAs as they verify at boot. The
wrapper reports: i2_s quantization confirmed, bitnet.cpp notllama.cpp, ~6 tokens per second on the local M2.
The most useful close for a district administrator is to take three
documents:
Faculty primer — a one-page explainer for teachers. What ternary
math is. Why the model fits the closet machine. What it can and
cannot do.
Board budget memo — the math the board needs to see. $48K
saved versus single dollars on electricity, with the deployment
shape spelled out plainly.
Intern deployment runbook — what a high-school intern can do on
a Saturday with the M2 mini and adult supervision. The deployment
is not magic. The decisions belong to the district.
A school newsletter announcement rides along for the families.
<!-- finish_text -->
That was the simulated path through Mr. Vega's third option. The
district owns the deployment, the data, and the decision. Garrison
Node supplies the tools — the wrapper that turns a generic local
LLM into something a school can be accountable for. Ember and
Hearth are the council members who do this work. The full card
breaks out the savings, the deployment runbook, and a
single-Tuesday-afternoon prediction that's yours to test.