FG-0807 Capability Versioning
Description
Models, agents, and prompts change over time, and an undocumented change is an untraceable one. Capability Versioning applies the discipline of software versioning to the components that produce behavior: every model, agent profile, and significant prompt is versioned, reproducible from its inputs, and reversible to a known-good prior state. When behavior changes, versioning is what lets you say exactly what changed, compare it against what came before, and return to the previous version if the new one regresses.
Implementation
Assign explicit versions to models, agent profiles, and significant prompts. Record what each version was built from: base, data, configuration, and parameters. Keep prior versions retrievable so any one can be restored. Tie behavior changes to version changes, so a shift in output can always be traced to a specific, documented change. Pair versioning with a tested rollback path, so reverting to the last known-good version is a routine operation rather than an emergency.
Success Indicators
- Every model, agent, and significant prompt carries an explicit version
- Any version can be reproduced from its recorded inputs
- A behavior change can be traced to the specific version change that caused it
- Reverting to the previous known-good version is a tested, routine operation
Failure Modes
- Changing prompts or configurations in place with no version record, making behavior shifts untraceable
- Versioning the model but not the prompt or agent profile, so a major behavior driver stays uncontrolled
- Retaining no prior versions, so a regression cannot be rolled back
- Recording version numbers without the inputs behind them, so versions are labeled but not reproducible
Sub-Methods 3
Record what each version was built from: base, data, configuration, and parameters. A version number without its inputs is a label, not a build. Reproducibility is what makes a version something you can investigate and trust.
Version the model together with the agent profile and the significant prompts, because each is a driver of behavior. Controlling the model while changing the prompt in place leaves the largest behavior lever uncontrolled.
Keep prior versions retrievable and the rollback path tested, so returning to the last known-good version is routine. This is the Reversibility Guarantee applied to capability: the handle has to work from the other side before you need it.