Imagine you hire a contractor to fix a leaky pipe. They show up, replace a fitting, the water flows, no drips — job done. Except three months later the wall is full of mold because they used a fitting rated for drinking water in a sewage line. The repair passed every obvious check, but the failure was baked in the moment they chose the wrong part from the truck. That is the silent failure problem in LLM-based vulnerability repair: patches that compile, pass tests, and still leave the vulnerability wide open. SAGE is a method for walking backward through the contractor's reasoning to find which decision on the truck was the one that doomed the job. The committed claim: SAGE is the first trace-level attribution method specifically designed for silent security failures — patches that give no observable error signal. Prior failure-attribution work assumes you can see the failure (a crash, a test failure, a timeout). Silent failures are invisible to those methods because the patch looks correct by every automated metric. SAGE fills that gap by combining security-reasoning assessment at each agentic turn with reconstructed code history to pinpoint the earliest divergence from the task's security intent. The evaluation corpus is substantial: 95 confirmed silent failures drawn from 3,684 repair traces across six agent frameworks and six base LLMs, tested on SecurityEval and CVEfixes benchmarks. SAGE assigned an origin turn in 93 of 95 cases. The headline finding is structurally important: most origins were reasoning failures — an unaddressed security requirement or an inadequate defense choice — and only 5 of 93 coincided with the actual code-writing step. When the agent did introduce vulnerable code, the origin preceded the write in 14 of 19 cases. The failure was decided before anyone touched the keyboard. Architecturally, SAGE belongs to the trace-analysis family rather than the runtime-monitoring family. It is a post-hoc diagnostic, not a guardrail. It operates on the serialized reasoning trace of an agentic workflow — the chain of thought, tool calls, and intermediate code states — and applies a structured security-reasoning rubric at each turn. The key structural choice is reconstructing the code history alongside the reasoning trace, which lets SAGE distinguish between a reasoning failure that happened to produce bad code and a coding error with no reasoning antecedent. This is not a gradient-based or attention-based attribution; it is judge-model evaluation over structured traces. Integrity has honest strengths and honest gaps. The 95 silent failures are drawn from real benchmark suites (SecurityEval, CVEfixes), not synthetic examples. Six agent frameworks and six base models provide meaningful diversity. But the grading is done by an LLM judge, and the authors acknowledge the limits: repeated scoring and a second judge reproduced the origin type more consistently than the exact turn, and agreement was lowest for traces that kept only the final file. There is no ground-truth labeling of origin turns by human security experts, which would be the gold standard. The method is internally consistent but not independently validated. The milestone that matters here is not about SAGE itself but what SAGE enables: targeted guardrails. If you know that 74% of silent failures originate in reasoning steps before code is written, you can insert security-reasoning checkpoints at those stages rather than relying solely on post-generation vulnerability scanning. The practical threshold is whether SAGE's origin localization is accurate enough to drive automated intervention — say, triggering a security-focused re-planning step when the reasoning trace shows an unaddressed requirement. That requires human-expert validation of origin labels at scale, which this paper does not provide. The obvious next experiment is a closed-loop intervention study: take SAGE's identified origins, insert a guardrail at that stage, and measure whether silent-failure rates drop. The authors did not run this. The honest read is option (c) — they are saving it for the next paper. SAGE as a diagnostic is a cleaner, more publishable unit than SAGE-plus-intervention, and the intervention study requires engineering an adaptive agent framework, which is a different kind of work. The other missing piece is human expert agreement on origin labels, which likely was not run because it requires recruiting security experts to read hundreds of multi-turn traces — expensive and slow.
