Answers are grounded in this paper and cite their sources.
We report an empirical evaluation of a reference implementation of Transitive Taint Propagation (TTP) [1], the primitive for tracing and containing the downstream consequences of a poisoned write across a shared field of agent state. The implementation realizes TTP as an extension to the L4 provenance ledger of an existing runtime guard: at write time it records each write's observed-read dependencies as edges in a persistent dependency DAG under a SHA-256 content-and-dependency commitment; on detection of a poisoned write p it computes the forward blast radius B(p) by graph traversal; and it contains B(p) with atomic, append-only quarantine annotations. We measure accuracy, the soundness boundary, and performance against synthetic workloads whose true dependencies are known by construction. Within a session, containment recall is 100%—every true descendant of a poisoned write is contained—at the cost of conservative over-approximation: because dependencies are captured as observed reads, B(p) is an upper bound on contamination, and measured precision falls from 38.7% to 17.6% as incidental reads rise, with the no-clear read scope the dominant source of over-containment. We then characterize the soundness boundary directly: cross-session recall degrades from 100% to 3% as the fraction of dependencies carried across a session boundary without a re-read rises from 0 to 1, establishing that soundness holds only for dependencies that flow through the tracked read path, not for those carried via untracked channels. On performance, the cryptographic commitment is negligible (≈0.002 ms/write); the dominant cost is the write-side accumulation of dependency edges under the no-clear scope, which grows quadratically with session length (≈12× a no-tracking baseline at 400 writes, though under 2 ms in absolute terms and above 500 writes/s); containment is linear in blast-radius size with no observed cliff through 10,000 nodes. We state the limits plainly: ground truth is synthetic, over-containment magnitude is topology-dependent, and the write-side quadratic cost is the primary scaling limit. This work is the Demonstrated and Measured rungs of the primitive's maturity ladder; all figures are reproducible from a seeded harness.
No takes yet. Share an insight, caveat, or question.
Andre Byrd (2026) studied this question.