A software security assurance layer kept its full refresh of 1,000 test contexts below a five-second decision gate in a local benchmark, with the longest one-worker run taking 1.62 seconds. The result is a timing finding, not evidence that the underlying auditor detected security defects.
The work, published as an arXiv preprint dated 28 Sep 2026, examined whether a system called CCPA could keep a current, policy-bound security status as evidence and the number of independent delivery contexts grew. The researchers framed this as a methods question about assurance at software-delivery decision gates.
A separate layer around the auditor
Instead of changing the auditor itself, the proposed design places a separation between policy, evidence and execution around unmodified RepoAudit. It ties each evidence set to the control context in which it belongs and recomputes the assurance status when policy or evidence changes, so an old status is not simply carried forward.
In ordinary terms, a Posture Card is the system's current assurance status for a particular context. The design is meant to keep that status connected to the evidence and rules that produced it.
The benchmark used RepoAudit without modification at a pinned revision on a fixed Python toy program containing a null-pointer dereference. The main evidence set consisted of 80 live RepoAudit executions and 80 run-level Evidence Bundles, split across two model configurations with 40 executions per model.
The timing test
The timing work itself ran as a single-host Python experiment in an isolated sandbox. Provider calls were used only to acquire RepoAudit evidence; evidence translation, context admission, posture computation, replay and scale measurement ran locally. For the single-context timing, the timed path did not include repository scanning, model invocation or evidence collection.
The researchers tested up to 1,000 independent contexts with one worker and with eight workers. They declared a five-second Decision Gateway budget before taking measurements. Each scale condition was repeated 30 times, while the policy sequence used three policy-class cycles in one execution. In the eight-worker series, the lower median was the 15th ordered duration out of those 30, making the summaries descriptive rather than population percentiles or worst-case guarantees.
In the single-context timing, the maximum observed event-admission-to-publication time across the three policy-class cycles was 1.1309 milliseconds, reported as 1.1 ms after rounding. Those cycles came from one sequence, so the result is an observed value, not a guarantee about every run.
In the broader one-worker test, when all 1,000 contexts were affected, full recomputation took at most 1,616.3 ms across 30 repetitions, or 1.62 seconds as reported. That remained below the five-second budget, but it is a single-host sample maximum, not an average per-context time or a promise about a future deadline.
With eight workers, the lower median refresh time rose from 1.3 ms at one Context to 288.9 ms at 1,000 Contexts. Across the reported scale points, maximum observed times were 4.1, 180.0, 35.3 and 317.0 ms, all below five seconds. These figures are descriptive observations from the 30 repetitions.
What the result does not establish
When the active control context changes before compatible evidence is available, the specified response is to reject inherited evidence and publish Inconclusive rather than reuse the earlier Posture Card. This is the design's safeguard against carrying a stale assurance status into a changed context.
That distinction sets the study's boundary. The experiment supports temporal adequacy only for the reported single-host configuration; it does not establish RepoAudit detection accuracy, distributed performance, comparative speedup or production-wide suitability.
No manual, passive-monitoring or competing-posture-engine baseline was included, so the paper reports absolute timing against a gate budget rather than speedup over an alternative.
The experiment was also bounded to a fixed Python null-pointer-dereference toy benchmark and a pinned RepoAudit revision. The result therefore describes the local recomputation path under those conditions, not RepoAudit's detection accuracy.
For readers treating the timing as a claim about security auditing, the distinction is important: the study measures how quickly the assurance status can be recomputed, not whether RepoAudit's security judgments are accurate.
Paper data and sources
Original title: Continuous Assurance of Agentic Security Auditors for Software Delivery Decision Gates
Authors: Guy Lupo, Nguyen Hung Nguyen, Viet Vo et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-28
DOI: Not available
Original paper · Full text