Track and Compare Model Versions with oci-bai

This guide walks through a complete evaluation workflow using oci-bai: push a baseline and a candidate, compare them at the package and benchmark depths, and read the verdict back.

Private beta

oci-bai is in private beta. Contact support@briefcaseai.org to request access.

What you need

• A running local stack (make up, make seed)
• The oci-bai CLI on your PATH
• Two image tags to compare (the guide uses the seeded rl-gym-env family)

Workflow

1. Push the baseline

   docker tag my-base:latest localhost:8080/rl-gym-env:cuda-base
   docker push localhost:8080/rl-gym-env:cuda-base
   oci-bai --repo rl-gym-env log cuda-base

2. Push the candidate

   docker tag my-candidate:latest localhost:8080/rl-gym-env:cartpole
   docker push localhost:8080/rl-gym-env:cartpole
   oci-bai --repo rl-gym-env provenance cartpole

3. Compare at package depth

   oci-bai --repo rl-gym-env diff cuda-base cartpole --depth package

   Output shows added, removed, and changed packages between the two versions.

4. Attach a benchmark verdict

   oci-bai --repo rl-gym-env attach-bench cartpole

5. Compare at bench depth

   oci-bai --repo rl-gym-env diff cuda-base cartpole --depth bench

   Output shows verdict deltas between the baseline and candidate.

6. Check impact

   oci-bai --repo rl-gym-env whodepends cuda-base

   Shows every version that descends from the baseline — useful before promoting a base update.

Key commands at a glance

Step	Command
Push	docker push localhost:8080/<repo>:<ref>
History	oci-bai --repo <r> log <ref>
Provenance	oci-bai --repo <r> provenance <ref>
Compare packages	oci-bai --repo <r> diff <v1> <v2> --depth package
Compare verdicts	oci-bai --repo <r> diff <v1> <v2> --depth bench
Find dependents	oci-bai --repo <r> whodepends <ref>
Search	oci-bai search "format==safetensors cuda>=12.4"