Skip to content
Multigrid

Compliance and fine-tune migration

Vendor assessments, access control, training configs and what a fine-tune's metrics mean on a new base model.

A provider migration is usually planned as an engineering change: swap the client, map the request bodies, run the regression suite. The parts that actually delay it are elsewhere. Somebody has to redo a vendor risk assessment that was signed off two years ago. Somebody has to work out who can call the fine-tuned model on the new side, and whether the answer is a key, a project or an IAM role. Finance has to explain why the switch month cost more than either the old or the new provider, in two currencies, across two billing cycles that do not line up.

These pages are about that layer. They stay away from what any particular vendor’s contract or audit report says, because those are negotiated per customer and change without notice, and they stay away from quoting prices that go stale. What is left is the part that transfers: what a clause type does, what a scoping model implies, and the arithmetic that makes a before-and-after comparison mean something.

What a Provider Migration Means for an Existing SOC 2 Vendor Assessment

Why a new inference provider triggers a fresh vendor risk assessment even when the outgoing one was approved, and what to request to make the new review short.

9 min read

Migrating Fine-Tuned Model Access Control Between Providers

How access to a tuned model artefact is scoped on different platforms, and which parts of your access story do not survive the move.

9 min read

Recreating a Fine-Tune's Training Config on a New Provider

Which tuning hyperparameters are exposed, what they are called, and why the ones that look portable are the ones that are not.

9 min read

What a Fine-Tune's Evaluation Metrics Mean on a New Base Model

Why training loss and eval scores from a tune on one base cannot be compared to a retune on another, and the four-cell measurement that recovers a usable answer.

11 min read

Mapping Frequency and Presence Penalty Parameters Between APIs

What the two penalties do to the logits, which APIs express them, and what actually substitutes when the target has neither.

9 min read

What to Do When a Migration Target Has No Equivalent for a Sampling Parameter

The rejection errors an unsupported sampling parameter produces, why silent acceptance is worse, and the closest working substitute for each control.

9 min read

Migrating a Cost-Per-Conversation Metric Between Providers

Rebuilding cost per conversation from each provider's own usage fields, including the cache accounting difference that makes a naive sum wrong.

10 min read

Reconciling Historical Cost Data Across a Provider Migration

A common schema and a step-by-step normalisation that makes two providers' usage records and invoices comparable on one axis.

10 min read

What Currency and Billing-Cycle Differences Mean for a Provider Migration

Why the switch month reconciles to nothing sensible, and what to change so the comparison and the accruals still work.

9 min read

Other topics