Quickstart in four commands.

From pip install to a signed, verifiable adapter in under five minutes. No keys leave the box. No telemetry. No account.

§ 01Prerequisites

You need python 3.10 or later and roughly 8GB of free disk for the default Hardened 7B model. Anything ARM64 (Apple Silicon, Snapdragon X) or x86_64 Linux works out of the box. Windows users should run under WSL2.

• Python ≥ 3.10 (3.11 / 3.12 supported)
• pip ≥ 23 (older pip can't resolve the cryptography wheel)
• 8GB free disk · 16GB RAM recommended
• Optional: CUDA 12.1+ or Metal — Lattice falls back to CPU otherwise

§ 02Install ltce

One command. The package vendors its Rust crypto core, so there's no separate compile step.

$pip install ltce
→ collecting ltce-0.1.4
→ resolved cryptography, fastapi, uvicorn, ed25519 · 12 deps
→ installed in 4.2s · python 3.11

Verify the install:

$ltce --version
→ ltce 0.1.4 · python 3.11 · darwin/arm64
→ crypto core: BLAKE3 1.5 · ed25519 1.7
→ 79/79 tests passing

§ 03Pull a signed model

Models live at huggingface.co/lattice-ai. Pulling fetches both the weights and the certificate, then verifies the BLAKE3 hash beforethe file is unpacked. If it doesn't match, the pull aborts — there is no fallback.

$ltce pull gpt-private-7b
→ fetching from huggingface.co/lattice-ai/gpt-private-7b
→ progress: ████████████████████ 4.1 GB · 28s
→ verifying BLAKE3 signature · ✓ valid
→ certificate: ε ≤ 4.8 · δ = 1e⁻⁶ · accountant=RDP
→ ready · ~/.lattice/models/gpt-private-7b/

The certificate is a JSON blob alongside the weights. Inspect it any time:

$ltce verify gpt-private-7b
→ model · gpt-private-7b · v0.1.4
→ blake3 · a8f3c7e2b4d6...91 ✓
→ signature · ed25519:7c4... ✓
→ ε ≤ 4.8 · δ = 1e⁻⁶ · steps = 14,720
→ verified · true

§ 04Train with differential privacy

Fine-tune the pulled model on your sensitive dataset. Lattice wraps the training loop with DP-SGD — per-sample gradient clipping plus calibrated Gaussian noise. The privacy budget (ε) is tracked automatically.

$ltce train ./medical-records --model gpt-private-7b --epsilon 4.8 --method qlora
→ loading base: gpt-private-7b · Q4_K_M
→ dataset: 12,847 samples · encrypted vault
→ DP-SGD: ε target 4.8 · δ 1e-6 · σ 1.1
→ epoch 1/3 ████████████████████ 100% · ε spent 1.62
→ epoch 2/3 ████████████████████ 100% · ε spent 3.24
→ epoch 3/3 ████████████████████ 100% · ε spent 4.81
→ adapter saved: ./output/adapter
→ certificate signed: BLAKE3 + ed25519

§ 05Verify and share

Every adapter ships with a tamper-evident certificate. Verify it locally, then share — publish to HuggingFace, send to a colleague, or deploy to production. The certificate travels with the weights.

$ltce verify ./output/adapter
→ ✓ JSON schema valid
→ ✓ BLAKE3 hash matches weights
→ ✓ ed25519 signature valid
→ ✓ ε ≤ 4.8 · δ = 1e-6
→ ✓ issuer: lattice-ai
→ certificate: PASS (5/5 checks)

$ltce push ./output/adapter --repo lattice-ai/medical-7b-private
→ uploading to huggingface.co/lattice-ai/medical-7b-private
→ certificate included · weights + cert.json
→ published ✓

§ 06Where to go next

You've trained and certified a private adapter. Here's what's worth reading next, ranked by how often people ask about it.