When a journal handles manuscripts on a rolling basis, authors can see how long their paper spends in each step: "under review," "reviews completed," "awaiting editor decision." Security research on shared computer queues has shown that the order and timing of service events leak the private state that drives them. We propose to measure whether the same holds for editorial boards. If a handling editor settles consensus papers quickly and holds split-decision papers for extended deliberation, the timestamps an author sees become a readout of confidential information, such as disagreement among reviewers or the likely decision. Our central claim is that this leak comes mostly from one or two specific tier transitions in the board hierarchy, and is not spread evenly across the workflow. If so, delaying and batching only those transitions should remove most of the leakage at a small fraction of the delay a journal would pay by batching everything or releasing decisions on a fixed date.
This matters because editorial institutions work hard to cut turnaround time. Queueing theory says that exactly this pressure maximizes timing leakage. Quantifying the Information Leakage in Timing Side Channels in Deterministic Work-Conserving Schedulers proves that any deterministic scheduler that never sits idle while work is waiting leaks at least half of a user's traffic pattern at low arrival rates. Common policies such as first-come-first-served and round robin leak nearly all of it. Nobody has checked whether editorial queues behave this way, how large the effect is on real data, or how a multi-tier board should trade speed for confidentiality.
Core Claim
We treat an editorial board as a series of queues, one per tier: reviewers, then an area chair or handling editor, then a senior or chief editor, then public release. Some of the transitions between tiers are exposed to the author as status changes. The secret $S$ is the confidential internal state of a paper, for example the eventual decision or whether reviewer scores diverge. The observable $Y$ is the vector of exposed transition times.
We will not attempt to estimate the mutual information $I(S;Y)$ directly, since that is unreliable at these sample sizes. Instead we use a standard lower bound that any trained predictor $q$ provides:
$$I(S;Y) \ge H(S) - \mathbb{E}\left[-\log q(S \mid Y)\right],$$
Here $H(S)$ is the entropy of the secret under its base rate, and the expectation is the predictor's held-out cross-entropy. A stronger predictor gives a tighter bound, so every leakage number we report is conservative.
The claim that goes beyond prior work has two parts. First, the bound is dominated by a small subset of tier transitions. Second, applying Accumulate-and-Serve batching (holding events for a window $T$ and releasing them together) only at those transitions gives a better leakage-versus-delay frontier than batching uniformly or releasing on a fixed date.
Relation to Prior Work
The mechanism comes from the shared-scheduler literature. Quantifying the Information Leakage in Timing Side Channels in Deterministic Work-Conserving Schedulers establishes the half-leakage lower bound and ranks queueing policies by how much they leak. Mitigating Timing Side Channel in Shared Schedulers introduces Accumulate-and-Serve and proportional time-slot scheduling and shows that both keep full throughput. Both papers assume Poisson arrivals and roughly uniform job sizes, and both study a single queue. Editorial work violates all three assumptions: submissions are bursty, review effort varies widely, and the queues are chained across tiers. That makes the analysis of how leakage composes across a hierarchy a real open problem, and it is the problem this note takes on.
The concurrency-theory work motivates the framing but does not supply the method. Epistemic Strategies and Games on Concurrent Processes shows that a scheduler restricted to knowing only its own history cannot leak hidden choices. IFC Inside shows that whether an observer can learn from termination depends on the scheduling discipline. Our analogue of termination is the number of revision rounds a manuscript goes through. We borrow the question these papers ask, not their formal proofs. An area chair who has read the reviews cannot be made blind to them, and that is the tension the file identifies between aggregation and blinding. What can be controlled is when the area chair's actions become visible to authors.
The editorial side supplies data and a testbed. ReviewArena covers 51,529 papers across seven venues and includes TMLR, a rolling venue. FirstPass provides 3,668 multi-round Nature Communications dialogues and predicts from content whether a paper needs standard or extended revision. AgentReview is a five-phase multi-agent simulation of review. It treats its orchestrator as a neutral synchronous engine and never asks what the orchestration exposes. We make that question the object of study.
Research Program
The first stage is measurement on real data. For rolling venues we reconstruct per-paper timelines of exposed status transitions. For TMLR we use the ReviewArena records; if those lack modification times, we recover them from the public OpenReview history of the same papers. For Nature Communications we use the FirstPass round structure, plus dates in the transparent review files where they exist. We then train timing-only predictors of the secret. These include gradient-boosted trees and a small sequence model over inter-event gaps, with the venue and the submission month as nuisance controls. From these predictors we compute the leakage bound for the full set of transitions and for each transition on its own, using drop-one and keep-one attribution to locate the tiers that leak.
The second stage is a controlled simulation where the true secret is known. We run the AgentReview pipeline and attach a tandem queue model whose service times depend on the internal state. We calibrate that dependence to the timing distributions measured in stage one, with bursty non-Poisson arrivals and heavy-tailed effort per paper. This lets us vary reviewer disagreement, area chair style (the authoritarian and conformist personas from AgentReview), and the release policy independently. Real data cannot separate these factors.
The third stage is policy design. We compare release policies per tier on the frontier of leakage bound against added median time to decision. We also check whether the batching windows prescribed by Kadloor et al. under Poisson assumptions remain adequate under the calibrated bursty traffic.
Baselines
The naive reference point is the fully work-conserving workflow that exists today: every status change is exposed the moment it happens. On real data this gives the empirical leakage of current practice at zero added delay. It anchors the right-hand end of every frontier.
The strongest existing mitigation is Accumulate-and-Serve applied uniformly to all exposed transitions, as described in Mitigating Timing Side Channel in Shared Schedulers. That paper reports that a buffering window of five times the observation window attains over 80% of the theoretical maximum privacy under Poisson traffic. We have found no reported result on Lacuna for editorial timing leakage under any policy. We will sweep $T$ over the same grid for uniform and targeted batching and evaluate both with the same held-out predictors, trained with the same tuning budget. Targeted batching beats the uniform version if, at matched leakage, it adds meaningfully less delay. We will require the frontier gap to exceed the bootstrap spread across splits.
A second strong competitor is fixed-date release, in which everything is held until a calendar date, as conferences do. It can reach near-zero timing leakage, and we will report its delay cost honestly. It is a real point on the frontier, not a straw man.
We include time-slot scheduling in the style of the work-conserving TDMA analyzed in Quantifying the Information Leakage in Timing Side Channels in Deterministic Work-Conserving Schedulers. That paper reports a leakage ratio of about 0.5, the lower bound, at low arrival rates. It is the best policy that never idles.
Two further references complete the comparison. The first is random-delay injection with the same mean added delay. The second is the key ablation: targeted batching applied to transitions chosen at random rather than by attribution. This isolates whether locating the leaky tier is what produces the gain.
As a content-side reference we use FirstPass, which reports 80.5% accuracy and 78.2% macro-F1 in predicting revision trajectory from manuscript and review text. Comparing timing-only predictors against this figure tells us how much of what the text reveals is already exposed by the clock.
Why not just…?
Why not just hide timestamps or release everything on a fixed date? For fixed-deadline conferences this is essentially the answer, and we expect our experiments to confirm it. For rolling journals the cost is the whole point: fixed-date release adds months to decisions that could otherwise come out in weeks. Hiding timestamps from the portal also does not hide when the author receives an email. The proposal places fixed-date release on the measured frontier. If targeted batching does no better, the result reduces to "use fixed dates," and we will report that.
Why not just add random delays? Random noise is the cheapest mitigation, and it is a serious competitor. A delay-dependent secret, however, still shifts the mean of the observed timing, so with enough papers an observer can average the noise away. Batching instead quantizes time and removes the within-window ordering. The matched-mean-delay comparison in our baselines settles which approach is better on real traffic.
Why not just train a classifier and report its accuracy, without any queueing theory? We do train classifiers; they are the measurement instrument. The theory adds two things a classifier cannot. It says which policies can drive leakage below a given level at all, since the work-conserving bound rules out zero-delay fixes. It also indicates where batching should go. A classifier alone yields a leakage number, but not the policy to reduce it.
Why not rely on double-blind review? Blinding hides the author from reviewers. Our secret runs in the other direction, from the board to the author, so textual anonymization does not touch the channel.
Decisive Checks
The idea is worth pursuing if three results hold. First, on TMLR or Nature Communications timelines, timing-only predictors beat the base rate for decision or reviewer divergence, with a leakage bound whose bootstrap interval excludes zero. Second, attribution concentrates most of that bound in at most two transitions. Third, targeted batching at those transitions dominates uniform batching and random delay on the leakage-versus-delay frontier, both on real data and in the calibrated simulation, and the random-transition ablation loses that advantage.
If timing leakage turns out to be negligible compared with the content-side reference, the proposal stops there with a useful negative result: current editorial portals do not meaningfully leak through time.
References
- Quantifying the Information Leakage in Timing Side Channels in Deterministic Work-Conserving Schedulers
- Mitigating Timing Side Channel in Shared Schedulers
- Epistemic Strategies and Games on Concurrent Processes
- IFC Inside: Retrofitting Languages with Dynamic Information Flow Control
- ReviewArena: A Large-Scale Cross-Conference Dataset and Benchmark for LLM Peer Review
- FirstPass: Grounding AI Scientific Judgment in Multi-Round Editorial Outcomes
- AgentReview: Exploring Peer Review Dynamics with LLM Agents