Registry · Models

The catalog of every model with its evidence attached.

Two tiers, one provenance authority. The models we host carry a real Ed25519-signed passport — base, method, eval, channel, and the adapter hash. The models we don't host carry a signed observation passport — the public facts, dated and cited, no benchmark invented. Both are tamper-evident and re-verifiable offline with the same key.

Open the studioVerify any model

13

Hosted in the registry

1

Served in production

10

Observed · signed, not hosted

1

Provenance key · both tiers

The registry

The register, by module

LeanLogix is a hosting and management platform — it runs many models, not just our own. The catalog separates them into two modules: the LockedIn Labs ecosystem builds the platform was proven on, then the wider library of hosted external and partner models. Both carry the same evidence: base, method, eval, channel, compliance state, and a passport that seals only when a real artifact exists.

Ecosystem models · 8

LockedIn Labs ecosystem models

The models built by the LockedIn Labs ecosystem and LeanLogix’s own platform — SprintLoop, ControlFrame, HealthNext, and the routing/infra layer. These are the builds the platform was first proven on.

ModelBaseParams · SizeVersionChannelEvalCompliancePassport
Portfolio Router
MLX LoRA · rank 16 · 16 layers · 400 iters cosine · promoted val-trough ckpt 200 (val 0.018)
portfolio-router-gold v1 (525 ex · 473 train / 52 valid · synthetic+public · no customer data) · Tiny fast task router: classifies each request to the right portfolio model + reason, in strict JSON. Base Qwen2.5-1.5B-Instruct (Apache-2.0). Held-out router eval 96/100 (base 76); refusal recall 5/5 on adversarial probes. Candidate — NOT certified/production.
Qwen2.5-1.5B-Instruct1.5B · 0.0422 GBv1Candidate96In review
HealthNext-Care-7B
MLX LoRA · rank 16 · 400 iters · cosine
HealthNext-Care corpus: 621 ex (589 train / 32 valid), 6 areas, 0 malformed/PHI/dupes · In-boundary healthcare ops reasoning. Base Qwen2.5-7B-Instruct (Apache-2.0). Val loss 3.671 to 1.108 (best). Serving base for per-customer forks. Domain hard-suite eval pending.
Qwen2.5-7B-Instruct7Bv126In review
HealthNext-Care-32B
MLX LoRA (QLoRA) · rank 16 · promoted val-trough ckpt 125 (val 0.920)
HealthNext-Care corpus: 621 ex (589 train / 32 valid), 6 areas, 0 malformed/PHI/dupes · Quality-tier candidate — wins the hard suite 34 vs 26 over Care-7B, decisive on governance (71.4% vs 39.3%). Promoted from val-loss trough, not last iter. Care-7B remains served default.
Qwen2.5-32B-Instruct32B · 32 GBv1-candidateCandidate34In review
SprintLoop-7B
LoRA · MLX (Apple Silicon)
sprintloop-corpus · 2,800 · no customer data · Governed coding + plan/verify; refuses unsafe ops. Served in production.
Qwen2.5-Coder-7B-Instruct7B · 1.4 GBv6Production89Certified
HealthNext-7B
QLoRA · planned
public CMS corpora + synthetic (RAG; no PHI in weights) · Compliant healthcare ops — tuned on PUBLIC CMS rules; PHI stays in RAG/runtime, never in weights.
Qwen2.5-7B-Instruct7BplannedDevPending
SprintLoop-7B · Edge
GGUF Q4_K_M · quality retention 97%
derived from sprintloop-7b v3 · Lean build for on-device / CPU. Quality-retention report attached.
SprintLoop-7B v37B · 4.1 GBq4-v1Candidate80In review
ControlFrame-Compliance-7B
LoRA · planned
controlframe-gold (in prep) · Compliance-framework interpretation + evidence planning. First governed build.
Qwen2.5-7B-Instruct7BdraftDevPending
SprintLoop-32B
QLoRA · MLX (Apple Silicon)
sprintloop-corpus · 2,800 · no customer data · Scale tier — wins the hard benchmark 90 vs 83. Awaiting promotion.
Qwen2.5-Coder-32B-Instruct (8-bit)32B · 5 GBv1Candidate90In review

Hosted models · 5

Hosted models · the wider library

External, partner, and customer models that LeanLogix hosts and governs but does not own. Same registry, same evidence contract — proof that the platform manages many models, not just our own.

ModelBaseParams · SizeVersionChannelEvalCompliancePassport
healthnext-care-prod-demo
lora-fork
tenant:prod-demo corpus:ab12cd34 · Per-customer in-boundary fork for tenant "prod-demo" — trained on NON-sensitive environment only; PHI/secrets held out to runtime RAG.
healthnext-care-7blorav1In review
healthnext-care-acme-demo
lora-fork
tenant:acme-demo corpus:ab12cd34 · Per-customer in-boundary fork for tenant "acme-demo" — trained on NON-sensitive environment only; PHI/secrets held out to runtime RAG.
healthnext-care-7blorav1In review
healthnext-care-meridian-health
lora-fork
tenant:meridian-health corpus:1d8c8a28 · Per-customer in-boundary fork for tenant "meridian-health" — trained on NON-sensitive environment only; PHI/secrets held out to runtime RAG.
healthnext-care-7blorav1In review
ThirdCandle-Advisor-7B
QLoRA · planned
trade-discipline-narratives (in prep) · Evidence-only trading advisor narratives — no buy/sell directives or predictions; discipline + risk framing.
Qwen2.5-7B-Instruct7BplannedDevPending
Itkalam-7B
QLoRA · planned
itkalam-arabic-corpus (in prep) · Conversational Arabic tutor — MSA + Egyptian/Levantine/Gulf/Maghrebi dialects. First Itkalam build.
Qwen2.5-7B-Instruct7BplannedDevPending

SprintLoop-7B is served in production today; the 32B and Edge builds are candidates in review. Hosted models — Third Candle, Itkalam, and more to come — run on the same registry and the same contract as our own. The passport column reads passport-ready until a build's artifact is stored and hashed, at which point it seals to signed — we don't print a checkmark the registry can't back with a re-checkable artifact. Early builds with no eval and no artifact are shown exactly as they stand; the registry doesn't round drafts up to releases.

Observed · the models we don't host, signed anyway

A signed provenance index for the models everyone runs

We host a handful of models. The industry runs thousands. An observation passport is a tamper-evident, offline-verifiable record of the public facts about a model we do not host — parameters, base family, context window, license, release date, and the source we read them from. It is signed with the same key that signs our hosted passports, so the whole index pins one provenance authority. No benchmark, no quality claim — observations are public facts, dated and cited, never a leaderboard.

ModelParamsBase · contextLicenseObservedPassport
Llama 3.1 70B Instruct
Meta
decoder-only transformer (Grouped-Query Attention; not MoE)
70B128K tokensLlama 3.1 Community License Agreement2026-06-25observed · signed
Qwen2.5-72B-Instruct
Alibaba (Qwen Team)
dense decoder-only transformer (RoPE, SwiGLU, RMSNorm, GQA, 80 layers)
72.7B total · 70.0B non-embedding32,768 native · 131,072 with YaRNQwen License (license_name: qwen)2026-06-25observed · signed
Mistral Small 3 (Mistral-Small-24B-Instruct-2501)
Mistral AI
decoder-only transformer (dense — community-attested, not stated verbatim on the official card)
24B32K tokensApache-2.02026-06-25observed · signed
DeepSeek-V3
DeepSeek-AI
Mixture-of-Experts (Multi-head Latent Attention, DeepSeekMoE, Multi-Token Prediction)
671B total · 37B activated per token128K tokensDeepSeek License Agreement v1.0 (weights) · MIT (code)2026-06-25observed · signed
GPT-4o
OpenAI
single end-to-end multimodal network (internal architecture not publicly disclosed)
not publicly disclosed128K tokens (16,384 max output)proprietary (OpenAI API / commercial terms)2026-06-25observed · signed
Qwen2.5-Coder-7B-Instruct
Alibaba (Qwen Team)
dense decoder-only transformer (RoPE, SwiGLU, RMSNorm, GQA, 28 layers)
7.61B total · 6.53B non-embedding32,768 native · 131,072 with YaRNApache-2.02026-06-25observed · signed
Llama 3.3 70B Instruct
Meta
decoder-only transformer (Grouped-Query Attention; not MoE)
70B128K tokensLlama 3.3 Community License Agreement2026-06-25observed · signed
Gemma 2 27B Instruction-Tuned
Google DeepMind
decoder-only transformer (interleaved local/global attention, logit soft-capping, GQA)
27.2B8,192 tokensGemma Terms of Use2026-06-25observed · signed
Phi-4
Microsoft Research
dense decoder-only transformer (trained heavily on synthetic + filtered data)
14B16,384 tokensMIT2026-06-25observed · signed
Mixtral 8x22B Instruct v0.1
Mistral AI
Sparse Mixture-of-Experts (8 experts, top-2 routing)
141B total · 39B active per token64K tokensApache-2.02026-06-25observed · signed

All 10 observation passports re-verified server-side as you loaded this page. These are observed, not hosted + signed — we did not train or evaluate them; we signed the public facts so the record is tamper-evident. Open any one to re-verify it yourself, or paste the receipt into the public verifier at lockedinlabs.ai/verify. Signing key a97122c37f…97fac1e1.

Datasets

The corpora behind the builds

Every dataset is registered, PII-scanned, and marked for customer-data status before a run can use it. Across the whole registry there are zero customer rows — 'No' under customer data is the state we want, and the one we report.

DatasetRowsStatusPII scanCustomer data
sprintloop-corpus2,800 verifiedapprovedclean · 0 findingsNo
controlframe-goldin prepregisteredpendingNo

The live library

This catalog is the public read of the registry

The full model library — version history, eval runs, release records, and the signed passports themselves — lives inside the gated studio. Open it to follow any build through dev → candidate → approved → production and re-verify its evidence.

Open the registry where these builds run

13 models on the public catalog today — 8 ecosystem builds and 5 hosted models — and the platform is built to manage many more. Step into the studio and follow one through every stage of its evidence.

Open the studioHow the registry works