Rootconf 2026 Annual Conference

One-line summary. Quantisation changes the model, but the decision usually sits with whoever owns the serving cost. I want to compare notes on who actually makes that call across teams, and what evidence they require before it ships.

The problem

Precision is one of the few changes that crosses every team boundary at once. It is a cost decision, a platform decision, a model quality decision and a hardware decision, and in most organisations nobody owns all four. The result is that it either never happens, or it happens without a validation story.

I have measured the technical side extensively and I have very little visibility into how other teams organise the decision. That asymmetry is exactly why this should be a discussion rather than a talk.

Questions for the room

• Who signs off when the served model changes numerically but not by version?
• What evidence does your team require, and has anyone ever rejected a change on that evidence?
• Does your serving platform expose precision as a knob, and should it?
• When a quantised model gets worse in a way your metrics do not show, how would you find out?

Who this is for

Platform engineering, MLOps, SRE, engineering leaders.

Level: intermediate.

Takeaways

1. A comparison of how different teams structure the approval, what evidence each requires, and where it breaks down.
2. A shared list of validation practices worth stealing.

What I will share

About ten minutes of measurements to give the room a common factual footing, including the failure modes that are not obvious, and then I facilitate rather than present.

My experience with this problem

Open source project. Research or investigation.

What failed, disappointed, or created unexpected problems

On my own side, treating precision as a purely numerical question. Almost every real difficulty turned out to be about which artefact you are allowed to change and who has to sign off, which is not something I can answer from my own work.

What I would do differently today

Ask the people who ship this in production rather than infer it from papers.

Trade-offs

Running this as a talk against a BOF. I have enough material for a talk, but the interesting half of the question is the half I have no data on, and a room full of platform engineers does.

How this helps other practitioners

A set of operational practices, and a way of framing a decision that currently falls between teams.

Current state: in progress.

Tags: #bof #mlops #platformengineering #inference #governance #modelserving