Background
The client is a foundation-model company offering APIs and fine-tuning services on a self-developed model. With fast iteration and heavy external calls, multiple regulatory obligations triggered at once before the first commercial release, and they are interlinked: filing is “registration” while safety assessment is “risk argumentation”—different procedures, different materials.
Key Challenges
- Procedure confusion: The team treated “LLM filing” and “generative-AI safety assessment” as one action, misaligning the material list and filing scope.
- Missing corpus evidence: The source and authorization chain of training corpora were not systematically recorded, leaving no verifiable evidence when the assessment required a legitimacy statement.
- Overlapping algorithm functions: The model combined generation, recommendation, and retrieval-augmented functions, requiring per-function determination of algorithm-filing triggers.
- Launch-pressure: With a fixed commercialization date, a delayed compliance baseline would mean “launch first, remediate later.”
Geason’s Approach
- Dual-track procedures: Clarified that LLM filing (CAC registration) and generative-AI safety assessment (risk-assessment report) are two independent procedures, prepared and advanced in parallel.
- Corpus compliance record: Built a corpus-source ledger and authorization chain distinguishing owned, purchased, public, and user data, producing a verifiable legitimacy statement.
- Function-split filing: Determined algorithm-filing triggers per function (generation, recommendation, RAG) to avoid under-filing.
- Deliverables: Filing/registration materials, safety-assessment report, and corpus-compliance ledger—three traceable document sets.
The Non-Obvious Judgment
Most foundation-model companies treat “filing” and “safety assessment” as one item and submit a single set of materials. Our view: they differ in nature—filing is registration of the entity and model, while the safety assessment is a risk argument on corpus, labeling, and use cases; the assessment conclusion is what regulators actually examine. Critically, corpus-source compliance is not material patched before launch but an evidence chain recorded at training start; by filing stage it is often impossible to reconstruct. This judgment determines the integrity and verifiability of the compliance baseline.
Key Outcomes
- LLM filing and generative-AI safety assessment materials were complete, producing traceable filing and registration records.
- A corpus-compliance ledger covered authorization chains for owned, purchased, public, and user data, supporting the assessment’s legitimacy statement.
- Algorithm-filing triggers were determined and closed per function, avoiding post-launch regulatory-interview exposure.
Client names and commercial data are anonymized. The rules, regulatory procedures and deliverable types described are as they occurred, not simplified.