A procurement system of record for open-weight and hosted model choice: one catalogue entry per model with licence terms, training-data disclosures, jurisdiction, weights hash, benchmark provenance, and the incidents attached to it — plus a policy engine that blocks a model from a project whose rules it violates (no ambiguous licence, no undisclosed provenance, no EU-hosted dependence). Designed to be the document a founder shows a client, an auditor, or a lawyer when asked which models the product uses.

Market Panorama

Competitive Intelligence

Users & Demand

Business Model

Market Sizing

Tech Solution & Build Plan

Go-to-Market Strategy

Risk Assessment

Action Recommendations

### Market definition and scope

The model sourcing register is a maintained record of what each model is and what choosing it obliges you to do. One entry per model, covering licence terms with a plain-English summary, hosting options and jurisdictions, weights availability and hash where published, context and pricing where disclosed, training-data statements, and dated links to incidents that touched the model.

On top of the record sits a policy engine. A founder defines rules — no ambiguous training-data provenance, EU data residency required, licence must permit commercial redistribution — and the engine returns a verdict per model per project. The output is a procurement packet: a generated document listing the models in use, their terms and the rationale for each choice.

The scope boundary excludes three adjacent things. It is not a model router; routers optimise price and latency and carry no compliance metadata. It is not an inference provider. And it is not an enterprise governance platform, which is a different product for a different buyer at a different scale.

Its unit of value is the document a founder hands to a client, an auditor or a lawyer when asked which models the product uses and under what terms.

October made the question concrete. Mistral released weights for Large 4 (1,726 points / 440 comments) weeks after the hosted launch at 2,044 points and 1,216 comments — the same capability under two completely different obligation regimes. Aleph Alpha positioned Kolibri explicitly as a sovereign open-weight model (713 / 336), making jurisdiction a first-class product claim. Cloudflare shipped Clef decision models with an RL fine-tuning platform attached (642 / 218), and Reflection shipped Beam at 501B parameters (556 / 172).

### Key trend signals with data citations

The volume is the signal. Eight or more model releases reached the front page in a single month, spanning weight releases, hosted launches, decision models and time-series foundation models. That volume of attention on licence and hosting distinctions means buyers are actively comparing, which is the precondition for a register.

The provenance dimension arrived separately and is more consequential for a small vendor. Anthropic was reported to have trained Claude on illegally obtained Stanford professors' books (606 points / 298 comments). A founder selling AI features to business customers now faces a question about training data that a closed-model vendor cannot answer on the founder's behalf.

Sovereignty formalised the jurisdiction axis. Kolibri's framing as a sovereign open-weight model is a marketing claim, and it is also a purchasing category: some buyers must know where inference happens and which legal regime governs.

### Why now

Two forces converged. Open-weight releases created obligations where a vendor API imposed only a contract, and an obligation requires a record. Simultaneously, enterprise procurement questions about AI vendors became routine, converting an intellectual interest into a deadline.

The position stays unoccupied for a mundane reason: the work is curation, and curation does not look like a startup. Routers have the traffic but no incentive to publish compliance metadata that might make their catalogue look risky. Governance platforms have the depth but sell to enterprises. Nobody has taken the middle, which is a maintained record at founder prices.

### Market maturity

Early. There is no standard licence summary format, no shared provenance disclosure structure, and no accepted way to compare jurisdiction. The first register to establish a citation-grade format defines what the question looks like for everyone after.

### What this means for a solo founder

Build the free register to accumulate citation value, and sell the packet to the founder who has just received a security questionnaire. The build is four weeks, which makes this one of the few opportunities in this dossier where a wrong bet costs very little.

Mistral released weights for Large 4 (1,726 points / 440 comments) weeks after the hosted launch (2,044 / 1,216), making licence terms a live decision.

Open weights as a release strategy

Aleph Alpha shipped Kolibri as a 'sovereign open-weight model' (713 points / 336 comments), formalising jurisdiction as a purchasing dimension.

Sovereignty as a product claim

Anthropic was reported to have trained Claude on illegally obtained Stanford professors' books (606 points / 298 comments).

Provenance becomes a legal story

Cloudflare released Clef decision models with an RL fine-tuning platform (642 points / 218 comments), blurring model and infrastructure provenance.

Model Provenance and Sourcing Register