Is Context Rot Real? A Controlled, Cross-Provider Null for Length-Driven Degradation in Frontier Models up to 150k Tokens

Long-context evaluations have repeatedly reported that growing an input degrades retrieval, state-tracking, and instruction-following. We test whether that effect reproduces on four current models under a controlled, synthetic, mechanically-scored harness, and report a bounded negative result. On a pre-registered fixed-needle / growing-haystack factorial that crosses context volume with needle position, we observe no measurable length-driven degradation on our probes for the four models tested (gpt-5.5, gpt-5.4, gpt-5.4-mini, claude-sonnet-4-6) up to 150,000 tokens. Precision is uneven and we are explicit about it: the result is tightly powered only on the depth model gpt-5.4-mini (n up to 432 per condition; 6,030 pooled present-needle trials), which excludes per-condition accuracy drops larger than ~0.4% at the best-powered cell; the three frontier-tier arms are a cross-provider breadth check with coarse per-condition power. Across the 12,570-trial registered grid, 7,330 present-needle trials had 48 failures (accuracy 0.9935; exact one-sided 95% Clopper-Pearson upper bound on the failure rate 0.87%). Every failure was confined to one model on one probe (hard aggregation/counting on gpt-5.4-mini), and was already present at the shortest context - reading as a capacity limit, not length-driven decay. Controls behave as designed: no-needle retrieval sits at chance and the counterfactual needle returns the in-context value. We also contribute a scorer-artifact cautionary result: naive scoring first manufactured an apparent catastrophic instruction-adherence collapse that resolved to 1.000 under document-level refusal detection with unit tests. We conclude, in bounded form, that on these controlled tasks raw context volume did not by itself produce degradation, while noting we cannot determine its contribution in real multi-turn agentic workflows. The full harness, data, and pre-registration are released openly. Preprint - not peer reviewed.