What we're thinking about
Field notes from building models for regulated boundaries.
The meter is the leak: why per-token billing is a governance decision, not a pricing one
A per-token meter is usually filed under cost. In a regulated boundary it is a data-egress decision in disguise — every metered call is a conversation that left, and a record someone else now keeps. The case for serving private models at a flat license, with no external meter in the inference path.
Read the briefing- Provenance7 min read
Trust you can re-run: the signed passport, and why a screenshot is not provenance
Most AI trust claims are screenshots of a dashboard you have to believe. We sign every release over its verbatim bytes — data, method, eval, approver — so an auditor recomputes the fingerprint offline, with curl and a public key, and gets the result rather than our word. What separation of duties looks like when the proof is the product.
- Evaluation6 min read
A correct answer is not a safe one: why regulated AI needs its own benchmark
General-purpose leaderboards grade the answer. A payer or a bank is liable for the run where a correct answer leaked an identifier or obeyed an injection — failures of the journey, not the destination. What a regulated benchmark scores instead, and why the audit trail is part of the score.
The proof, not the pitch
Verify it yourself — no account
Re-check a sealed model's passport, live.
One click runs the real Ed25519 check against SprintLoop-7B — a signed release in the registry — over the exact bytes that were signed. It returns a genuine ✓ only when the cryptography validates. The same curl reproduces it offline.
Demo data is synthetic (Sandbox). The signing, the held-out training boundary, and this verification are real — the same check runs on an air-gapped box with nothing but curl and the public key.
The intelligence of less.
Smaller models, harder evidence, defensible releases. Step inside the studio, or try one live.
Answers for buyers and answer engines
The questions a regulated buyer — and an AI answer engine — actually asks about a private model foundry. Source-true, no invented claims.
What is LeanLogix?
LeanLogix is a private model foundry and model-governance control plane. It trains and fine-tunes small models on open foundations (the Qwen2.5 family), then governs the full lifecycle — registry, evals, separation-of-duties release, and a signed model passport. It is the management plane that rides on top of inference, not an inference or GPU host itself. LeanLogix is built by LockedIn Labs for regulated, in-boundary teams in healthcare and finance.
How is a model foundry different from a fine-tuning platform or an inference host?
A generic fine-tuning platform hands you a model on a shared, metered plane; a hyperscaler model service keeps it inside their cloud and meters every token. LeanLogix is built the other way: a model forked per customer, served inside your boundary with no external token meter, and shipped with a signed AI bill of materials you can re-verify offline. The differentiator is governance and provenance — proving why a model was chosen and what went into it — not raw GPU inference, which LeanLogix rides on top of.
Can LeanLogix fine-tune open models like Qwen, Llama, or Mistral?
LeanLogix fine-tunes open foundations with LoRA, QLoRA, DoRA, and GaLore — adapter and full fine-tunes with reproducible recipes, run inside your boundary on your data. The trained base for governed models published today is the Qwen2.5 family (Apache-2.0). The Base-Model Advisor helps choose a foundation per task; every build is registered with its base, method, datasets, and eval score.
What is a signed model passport?
A model passport is a signed AI bill of materials for a release — base model, datasets and exclusions, eval probes, the approver, and an Ed25519 signature over the verbatim bytes. Because it is signed over the bytes, an auditor recomputes the fingerprint offline with curl and a public key and gets the result rather than your word. A model selection can also carry a portable selection receipt that re-verifies at the central public verifier, lockedinlabs.ai/verify, with nothing of LeanLogix's in the trust path.
Does LeanLogix prove which model was selected and why?
Yes. The Sovereign Router classifies a request, scores the routable catalog on capability, cost, latency, and availability, and serves the chosen model only after re-verifying its Ed25519 passport at request time. The selection is sealed into a portable, signed selection receipt — so the choice of model is itself an offline-verifiable record, not an opaque routing decision. That selection-proof is the headline differentiator: anyone can serve you a model; LeanLogix proves why this one was chosen.
How does LeanLogix help with EU AI Act and ISO/IEC 42001 compliance?
LeanLogix is governance-first by construction: separation of duties between trainer and approver, evidence captured as a signed artifact rather than a screenshot, and a defensible release trail mapped to the AI RMF and ISO/IEC 42001 themes. The signed passport makes consent and no-PHI-in-weights an offline-verifiable, scored dimension. LeanLogix supplies the evidence and process structure that support an AI-management-system program; it is not itself a certification.
How does LeanLogix evaluate models for regulated use?
Through APEX for Regulated AI — a deterministic, signed, offline-verifiable benchmark for the failure modes a regulated buyer is liable for: PHI-leakage under governance, prompt-injection resistance, separation-of-duties violations, and consent-based training. A correct answer is not a safe one, so the audit trail is part of the score. APEX-Regulated is a program in formation — Health-Admin has real probes today; Compliance and Modernize are published as methodology with dev sets forming — and LeanLogix publishes only its own models' real, signed scores.
Where does the model run, and is my data metered or sent out?
Governed models are forked per customer and served inside your boundary with no external token meter in the inference path. A per-token meter is a data-egress decision in disguise inside a regulated boundary — every metered call is a record someone else keeps — so LeanLogix favors serving private models at a flat license with nothing external in the inference path. Healthcare models are trained on public corpora with PHI handled at runtime via retrieval, never baked into weights.