AI outputs change.

Models update. Prompts evolve. Small output shifts can silently break production logic.

If you're extracting structured data (invoices, tickets, reports) from LLMs, a tiny change in model output can cascade into incorrect downstream behavior.

Continuum records a multi-step LLM workflow once, then deterministically replays and verifies it later.

If anything changes — raw model output, parsed JSON, or derived memory — your CI fails.

Example:

1. Run `continuum invoice-demo` 2. It extracts structured fields from an invoice 3. Run `continuum verify-all --strict` → PASS 4. Modify a stored value (e.g., 72 → 99) 5. Run verify again → FAIL

It’s a simple drift guard for LLM pipelines.

No hosted service. No external storage. Just deterministic replay + strict diffing.

Repository: https://github.com/Mofa1245/Continuum

Feedback welcome.

A few clarifications:

- This isn’t trying to make LLMs deterministic. - It records the full workflow output once, then replays and diffs it later. - The goal is CI drift detection, not runtime enforcement.

Curious how others are currently guarding against silent output drift in production.