v0.1.4 · MIT · pip install ltce

Train private.
Prove it. Share safely.

Fine-tune on sensitive data with differential privacy. Every adapter ships with a signed certificate — verifiable by anyone, tamper-evident by design. Your training data stays protected even if the weights leak.

ε≤8
Budget
79/79
Tests
10
Signed Models
0
Training Data Leaked
εδσPr[M]
§ 01 — The three pillars

Train. Certify. Share.

Three primitives, one pipeline. Each is a node in the lattice — useless alone, definitive together. Privacy you can verify.

// node_01

Train
with noise.

DP-SGD with Rényi accountant. Gaussian noise calibrated to your ε. Every gradient step accounted for, every dataset bounded.

μσ
// node_02

Certify
the weights.

BLAKE3 hash every weight tensor. ed25519 sign the certificate. Record ε, δ, noise multiplier, training steps. Tamper-evident — one flipped bit and verification fails.

cert
// node_03

Share
without leaking.

Publish your adapter to HuggingFace, share with a colleague, deploy to production. The certificate travels with the weights. Anyone can verify. Your training data stays protected.

π
§ 02 — The math

Privacy is an inequality.

A model is (ε, δ)-differentially private if the probability of any output changes by less than eε when a single training example is added or removed. We make ε small. We make δ tiny. We prove it.

Pr[M(D) ∈ S] eε · Pr[M(D′) ∈ S]+ δ

For all neighboring datasets D, D′ and all output sets S.

// 01
Clip

Per-sample gradients bounded to L2 norm C. Hard cap.

// 02
Noise

Gaussian noise σ added per batch. Calibrated to ε.

// 03
Step

Noisy gradient → optimizer. Update happens.

// 04
Account

RDP accountant tracks budget. Bounds compose tightly.

// privacy_budget · liveε = 4.96 / 8.00
δ = 1.0e⁻⁶RDP α = 32steps 14,720
§ 03 — Why it matters

Share the model, not the data.

Every team fine-tunes on proprietary data. The problem isn’t training — it’s what happens after. Lattice lets you ship the adapter without shipping the dataset.

// 01

Codebase fine-tuning

500K lines of internal code. You want a copilot that knows your patterns, conventions, and architecture. Fine-tune with DP-SGD — even if the adapter leaks, no one extracts your source. Share across the team with a signed certificate.

$ ltce train ./src --epsilon 4.8
// 02

Customer data models

Support tickets, user feedback, sales calls. Train a model that understands your customers without leaking individual records. The ε guarantee means no single customer’s data can be extracted from the weights. GDPR-compatible by construction.

$ ltce train ./tickets --epsilon 3.2
// 03

Internal knowledge base

Docs, wikis, runbooks, post-mortems. Fine-tune a model on institutional knowledge and deploy to the whole org. The certificate proves the model can’t regurgitate confidential specifics — auditors can verify without seeing the data.

$ ltce train ./docs --epsilon 6.0
// 04

Medical& legal data

Patient records, case files, privileged correspondence. Regulated industries need provable privacy — not just promises. The Lattice certificate is a HIPAA artifact: here’s the ε, here’s the δ, here’s the signature. Verify it.

$ ltce verify ./adapter --strict
§ 03b — Web3 applications

On-chain models need on-chain proof. Lattice certificates are verifiable by smart contracts — not just humans.

// 05

DeFi alpha protection

Train a model on proprietary trading signals — MEV strategies, arb patterns, liquidity analysis. Sell the model, not the alpha. The ε guarantee means buyers can’t reverse-engineer your edge from the weights.

$ ltce train ./alpha --epsilon 2.0
// 06

DAO governance privacy

Train analytics on member voting patterns without exposing individual votes. The model predicts proposal outcomes — no single member’s history can be extracted. Verifiable privacy for on-chain governance.

$ ltce train ./votes --epsilon 1.8
// 07

Audit models as public goods

Security firms train on private vulnerability databases and publish the model for the ecosystem. The certificate proves it doesn’t leak the vuln DB — clients verify before trusting. Open security, closed data.

$ ltce push ./audit-model --public
§ 04 — Model catalog

Every model ships with a certificate.

Ten models on huggingface.co/lattice-ai. Each carries a privacy tier and a signed certificate. Pull one, train on it, verify it.

// hardened
gemma-4-e4b-uncensored
8B params · uncensored
Gemma 4 · 16 GB
ε ≤ 4.8
// hardened
dolphin-llama-3.1-8b
8B params · uncensored
Dolphin · 16 GB
ε ≤ 6.2
// hardened
dolphin-mistral-nemo-12b
12.2B params · reasoning
Nemo · 24 GB
ε ≤ 5.2
// hardened
qwen-3-8b-private
8B params · multilingual
Qwen 3 · 16 GB
ε ≤ 4.8
// hardened
deepseek-coder-v2-lite
15.7B params · MoE code
338 langs · 32 GB
ε ≤ 3.8
// hardened
phi-4-mini-private
3.8B params · reasoning
Phi-4 · 7.6 GB
ε ≤ 3.1
// wrapped
mistral-7b-private
7.3B params · general
Mistral v0.3 · 14 GB
tier-1
// wrapped
qwen-2.5-coder-7b-private
7.6B params · code
92 langs · 15 GB
tier-1

▸3 more on the registry · all signed · all MIT

§ 05 — Quickstart

Four commands. Zero data leaked.

One pip install, one pull, one train, one verify. Private training from byte one. The CLI is the contract.

~ / lattice · zsh
install the runtime
$pip install ltce
pull a base model from the registry
$ltce pull lattice/llama-3.1-8b-private
train with differential privacy
$ltce train ./medical-data --epsilon 4.8 --method qlora
→ ε budget: 4.8 / 8.0 · 14,720 steps · cert signed
verify the output adapter
$ltce verify ./output/adapter
✓ certificate valid · BLAKE3 match · ε ≤ 4.8
$
§ 06 — Attestation

Every weight ships with a certificate.

A tamper-evident JSON blob, BLAKE3-signed at the registry. Read it before you load the model. Reject if it doesn't match.

Training params · ε, δ, σ, steps, clip-norm
Hash of every weight tensor
Registry signature · independent verifier
Groth16 ZK proof · v0.2
{
  "model": "gemma-4-e4b-uncensored",
  "version": "0.1.4",
  "privacy": {
    "epsilon": 4.84,
    "delta": 1e-6,
    "accountant": "RDP",
    "steps": 14720
  }
  "blake3": "a8f3c7e2b4d6...91",
  "signature": "ed25519:7c4...",
  "verified": true
}
§ 07 — Privacy tiers

Three densities. One lattice.

Pick the tier that matches your threat model. Each is a denser lattice of guarantees — more nodes, more proofs.

Wrapped

Encrypted runtime around any model. No retraining.

AES-256-GCM at rest
Local-only inference
BLAKE3 signed
no DP guarantee
no ZK proof

Native

Trained from scratch on your private data. Maximum density.

Everything in Hardened
Your data, your weights
Custom architecture
Per-tenant ε
Groth16 ZK proof

Train. Certify. Share.

Install ltce in your next terminal. Train on your sensitive data. Ship a signed certificate with every adapter. Verify anywhere.