An engineering notebook. Whatever's on the workbench.

Plan-and-Execute reported success on all 84 gpt-oss trials. Only 48 of its answers held up against the tool evidence.

RAGAS scores 30 naive RAG answers at 0.370 correctness. A gpt-4o-mini judge scores the same answers 0.887.

Dense-only: 0.739 MRR on 30 queries. Hybrid BM25+RRF lifts it to 0.794. FlashRank adds 0.7 s and no ranking gain.

Priority admission keeps 99.7% critical success at 300 rps; blind shedding delivers 32.3%. Background pays with 12.9%.

Three shedding policies, the same 100 rps goodput at twice capacity. Unbounded FIFO falls to 7.8 rps.

Two gates, near-identical goodput at 47.8 and 48.0 rps. The downstream peak differs: 290 rps against 100 rps.

A timeout caps p99 at 200ms yet scores the same SLI as doing nothing. A bulkhead holds 99.6% and keeps the error budget.

Five admission strategies under the same 2x overload. Token buckets hold 10ms p99; a bounded queue is the worst of both.

A closed-loop benchmark reports 10ms p99 while the service froze for 500ms. Open-loop measurement says 460ms. Same run.

ShutdownOnFailure became a Joiner and throwIfFailed disappeared. The two non-mechanical parts of our Java 25 migration.

Outcome counters per scope, propagated deadlines, bulkheads, and JFR pinning watch: four checks before trusting fan-out.

Quorum aggregation, per-tenant bulkheads, and a deadline shape that survives a slow upstream. Cancellation comes first.

Joiner scopes replace ShutdownOnFailure and scoped values leave preview. What moving Loom to Java 25 actually takes.

Adopt the stable Loom pieces now and fence preview APIs behind small helpers. Timing, not features, decides the roadmap.

Carrier counts hide millions of virtual threads, and high CPU usually means pinning. Watch that instead of pool depth.