LLM Post-Training as Brownfield Maintenance: An Industrial Perspective on Dataware Engineering

Industrial post-training is framed here as a **brownfield regime**. Teams do not start from a clean slate; they inherit a deployed checkpoint and must land targeted improvements under fixed compute and mixture budgets without regressing the rest of the model's behavior. The maintained artifact is increasingly **dataware**: behavior governed by a curated post-training mixture, updated via bounded *mixture patches* rather than clean-slate retraining.

The paper distills this maintainer's perspective from an industrial code-generation improvement effort. Rather than presenting a single winning recipe, the authors identify three recurring challenges that dominate practice:

1. **Zero-sum mixture design** — adding data for one capability tends to displace budget for another, so gains are traded, not simply added.
2. **Yield as the binding metric** — the scarce resource is not raw teacher output but the fraction that converts into usable training data.
3. **End-to-end integration under uncertainty** — improvements must survive the full pipeline and stochastic evaluation before they count.

The central claim is that progress depends less on one-off recipes than on an **engineering discipl

The yield-engineered patch produced measurable, statistically significant gains from a single fixed checkpoint per condition, across 16 stochastic evaluations of each benchmark.

| Benchmark | Metric | Improvement |
|---|---|---|
| CodeForces | pass@1 | **+2.59 points** |
| CodeForces | pass@3 | **+3.11 points** |
| LiveCodeBench v6 (held-out) | pass@1 | **+6.11 points** |
| LiveCodeBench v6 (held-out) | pass@3 | **+8.05 points** |

Two features of these results matter for industrial maintainers. First, the gains are **statistically significant** under the 16-run stochastic protocol, which is the standard the authors argue industrial claims should meet. Second, the held-out LiveCodeBench v6 deltas exceed the in-domain CodeForces deltas, suggesting the yield intervention did not merely overfit the primary benchmark.

Equally important is what did *not* move: internal **AIME and MATH regression suites remained within tolerance**. In a brownfield regime, a patch that improves code generation but degrades mathematical reasoning is not a win — it is a regression that must be caught before landing. The reported tolerance compliance is therefore part of the result, not an afterthought.