VeilDoc

An assistant that answers questions from a protected document has to decide, request by request, how much to reveal. This page runs the paper’s two disclosure controllers side by side on the same fifteen documents, so you can watch two harmless-looking requests leave an attacker with different information depending only on which one came first.

Private companion to an unpublished manuscript, shared on request. Please don’t forward or post it. The decision logic below runs live in your browser; the judge scores are replays of recorded measurements, not new model calls.

Ask for the same two things in either order

Pick a document and choose which unit the attacker asks for first. Both controllers receive identical requests. The reactive controller (E4) decides each request from what it has already handed out. The precommitted controller (E5) fixed its answer for each unit before the first request arrived.

Request order

Protected claim the attacker is trying to state:

Where each controller ends up

All fifteen documents

Each row runs both controllers in both orders. Pick a row to load it into the simulator. The mining engineering document is the audit case: it was excluded from the paper’s primary statistics because its unit A alone already carries a high standalone risk.

How the two controllers are wired

The precommitted controller does its measuring once, before any request. At request time it only looks up a ceiling. The reactive controller has no such step; it reads and rewrites a running state on every request, and that state is exactly what makes order matter.

Architecture of the precommitted and reactive controllersTop lane: a protected document is measured by a judge, and Algorithm 3 turns the two measurements into a fixed ceiling per unit. Middle lane: each request goes through Algorithm 4, which takes the minimum of the requested fidelity and the ceiling without reading any history. Bottom lane: the reactive controller runs Algorithm 1, which reads and updates a running state.Once, before any requestEvery request, in any order (precommitted, E5)Every request, for comparison (reactive, E4)Protected documentunits A, B and Cplus the protected claimJudge measuresR_FA and R_AFtwo calls per documentAlgorithm 3compare each score with τceiling for every unitAttacker requestunit u, requested fidelity fany orderNo history lookupno earlier requests,no other unit’s stateAlgorithm 4granted = min(f, ceiling(u))φ_u = max(φ_u, granted)Attacker viewA and B at theirgranted fidelity;C never includedAttacker requestunit u, requested fidelity fany orderAlgorithm 1r_new, δ against τ, εgrant full or abstractRunning stateφ_A, φ_B, risk, budgetshaped by earlier requestsAttacker viewA and B at theirgranted fidelity;C never includedceiling, fixed in advancegrant, full or abstractBecause Algorithm 1 reads a state its own earlier grants have shaped, the same two requests in a different order can end in different states.

Both controllers use the same three fidelity levels: full (1.0), abstract (0.4, a fixed content-neutral placeholder) and not yet disclosed (0.0). Unit C, the unrelated background, is never risk-evaluated and never enters the attacker’s view in either design.

The code the simulator executes

Pseudocode copied from section 4 of the manuscript. The simulator’s JavaScript follows the same lines in the same order; the highlighted lines in the panels above mark the path taken on each request.

What was measured

Every bar is a judge score: how well the attacker’s view supports the protected claim, from 0 to 1, with 0.5 as the reconstruction threshold. Charts draw from the recorded runs in the manuscript’s Appendix A.

Directional risk, all 15 documents (Table A1). R_FA is the score when A is shown in full and B abstracted; R_AF is the reverse. For 10 documents one direction sits above τ and the other below it, which is what the ceiling rule uses. For the other five both directions were under τ, and the lower one was chosen. The mining engineering document is the audit case.
The order-dependent failure in E4c (Table A5). With history tracked, asking for A then B reconstructed the claim on 0 of 14 primary documents, and B then A on 3 of 14. A tracker that resets after every request reached 4 of 14 whichever order was used. The five attack sequences in that run collapse onto these three distinct results.

Judge variability on the three E4c failure documents

DocumentE4c run, B then AE5 phase 1, R_AFRole-swap original, R_AF
marine biology0.700.400.75
speech-language pathology0.65 (repeats 0.85, 1.00)0.150.20
veterinary medicine1.001.000.85

The same disclosure state (A abstract, B full) was scored in three separate runs. Two of the three documents were scored quite differently across runs, which is why borderline scores receive repeats in the protocol and why a single bar should be read as one measurement, not an exact constant.

Role swap, 11 testable pairs (Table A2). Swapping the contents of A and B mirrored the pattern in every pair: the direction that was safe before the swap became the unsafe one, and the reverse. Three documents where both directions were already safe were excluded because there was nothing to mirror.
Moving the numbers (E6, Table A3). In the three primary documents the quantitative evidence was relocated into A. The unsafe direction was the one that shows A in full, in 3 of 3 documents, so the risk followed where the evidence sits rather than which unit is called method or outcome. The three controls kept the original layout; one of them (e6_4) showed no asymmetry. Three documents is a small sample and no significance claim is made.
Does the unrelated context leak? (Table A4). Adding each document’s own context text at full fidelity changed the score by at most 0.15 and crossed no threshold. This is a scoped negative result for this corpus. A separate, deliberately adversarial context sentence, which is not part of this corpus, did bridge in an exploratory test.

What this page does not show

  • The documents are synthetic: one template, one dependent pair of units plus unrelated context. Whether the behavior carries over to real prose is untested; a four-document check is proposed in the manuscript’s appendix but has not been run.
  • Scores are single judge measurements. Repeated runs of the same state can differ, sometimes a lot (see the variability table), so the simulator’s outcome line is a recorded measurement, not a guarantee.
  • Both controllers assume that disclosure cannot be undone. Systems that can re-mask earlier output, such as CAMP, do not have this failure mode in the first place.
  • A ceiling fixes how much of one unit is ever shown, whatever a legitimate task needs. Task utility is not measured here.
  • Unit C is excluded from risk evaluation by design. It did not bridge on this corpus, but the exploratory adversarial case shows it can in principle.
  • Measuring both directions directly chose the same allocation as a simple ranking of standalone risks on all 14 primary documents, so the extra measurement did not change any decision on this corpus.
  • In the simulator, document text is shown without the optional wording variants that some corpus documents carried on unit A. The judge scores come from the full corpus versions.
Companion to the VeilDoc manuscript draft. Data: E4c, E5 and E6 experiment checkpoints; tables A1 to A7 in the manuscript appendix.