The 175,568 ZEC leaving the shielded pool is the Orchard result a trader should watch. The -0.162 abnormal return beside it is harder to trade: ZEC had begun falling hours before the unresolved-history disclosure. The authors acknowledge the timing problem in their introduction ("ZEC was already falling before that statement, so timing alone is not causal") and in the Figure 2 caption. Their claim is directional alignment, supported by reconstructed flows. As they put it, "These patterns are directional, not causal." They describe the flow series as the "mechanism-nearest within-event directional test" of their exit prediction.
A repaired circuit, an open history
Cha, Kim, Yeo and their coauthors focus on a limit of private ledgers. Repairing a circuit prevents new counterfeiting, and a supply check can leave the aggregate total looking intact. Past invalid notes inside a shielded pool could still escape detection. "Finding no abuse is therefore weaker than proving that abuse could not have occurred." The paper calls the price reduction attached to that unresolved history the verifiability discount and seeks to measure it separately from reported loss.
Their exit mechanism runs through Zcash's turnstile, which limits transfers out of each shielded pool to its recorded balance. Let θ denote the share of claims exposed to the flaw and q the share of adverse histories excluded by public evidence. The remaining exposure is z = θ(1 - q). Maxmin investors price the worst admissible case. If a mass m exits at par, each holder who stays faces 1 - z/(1 - m); departures concentrate the possible shortfall on those left behind. Exit is zero at z = 0 and rises locally with z on stable branches. The authors make no claim of global uniqueness.
Zcash Orchard in June 2026 provides the test case. The authors mark four times: the initial notice, the NU6.2 repair statement (no evidence of exploitation, supply unaffected), the Foundation's confirmation, and a June 4 21:33 UTC post saying prior exploitation was unlikely but could not be ruled out cryptographically. They estimate hourly ZEC returns against Binance BTC/USDT using 335 pre-discovery hours. The respective 24-hour post-anchor CARs are +0.185, +0.014, +0.016 and -0.162. Their VDISC coding scheme keeps reported loss, exposure, proof, returns and pool flows in separate fields; they apply it to 14 disclosures and a 40-event panel.
When did the market react?
Each 24-hour window excludes its anchor bar. The placebo uses 312 overlapping pre-discovery blocks. For the context post, the one-sided p of 0.003 reaches the floor available from that distribution. Across the 16-test family, Benjamini-Hochberg adjustment puts it at 0.051. The context window therefore clears a conventional 5% threshold before multiple-testing adjustment, though it misses that threshold afterward.
The clock makes the result difficult to read. The same statement appeared in an X article at 21:17, 16 minutes ahead of the forum post. Moving the anchor there changes minute CAR from -0.166 to -0.205. Yet the 24-hour pre-CAR had reached -0.134, and CAR had lost 0.208 since midnight. Of that decline, 0.188 preceded the X release. Turnover during the 16 minutes before the forum post was about four times the June 3 crisis-day level; the 60-minute post-anchor decline was about 1.6 times the pre-anchor decline. Changing clocks preserves the sign. The earlier sell-off rules out a discrete break.
The authors make the same point: checks across clocks and estimators "cannot turn a price decline already under way into a disclosure-time discontinuity." They flag two further complications. The final post combined severity with unverifiability, and Hayes announced an exit whose execution time is unknown. Their market model uses BTC alone. Extending estimation to 60 and 90 days preserves the sign (p≤0.001), but we did not find a second-factor specification. Their peer check gives reason to care: XMR and ETH weakened more than in their pre-windows, DASH improved, and SOL was nearly flat. The authors see the +0.185 after the containment notice as consistent with their prediction, although their two-date model has no channel for a positive move.
Pool exits after the context post
The Blockchair per-block reconstruction agrees with zcashinfo and node snapshots within 0.1%. During the 24 hours after the context post, net unshielding was 175,568 ZEC. The May 15 to 28 baseline was 3,560 ZEC per day, making the later flow about 49 times as large. The NU6.2 window registered 7,573 and the Foundation window 9,052; those windows share 18 hours. The pause registered -7 while spending was disabled. Starting with the first complete 22:00 bin, 48-hour net unshielding came to 243.2k ZEC, or 4.8% of the aggregate shielded balance. Orchard supplied 97.4% of the May 24 to June 7 decline.
One June 5 block matters greatly. At 19:34 UTC, 20 transactions produced 60,324 ZEC of net outflow, or 34% of the 24-hour total. Remove that block and the total remains about 32 times baseline. Net flow subtracts inflows, however. The authors caution that it differs from gross exit m and may reflect delayed transfers or offsetting inflows. Binance open interest increased 21.8% as its dollar value fell with price, while spot order flow remained net selling. More contracts could be consistent with hedging; the paper offers no such interpretation.
The 48-hour figure is unchanged by the two clocks. Both 21:17 and 21:33 fall within the 21:00 hour, leaving 22:00 as the first complete hourly bin.
The cross-event slope changes sign
A daily signal would have to carry across events. The authors' checks leave that prospect weak. Across nine priced events, the calendar-day CAR-z slope is -0.172, with a bootstrap interval of -2.237 to -0.004. Re-time Orchard and the slope becomes -0.097. Re-time all nine and it becomes +0.027. Both re-timed intervals span zero. Orchard's daily CAR under an ETH market model moves from -0.334 to +0.015 when the window shifts to the first three complete post-disclosure days. The authors concede that "the pilot does not identify even the sign of the slope."
Exposure coding presents another difficulty. For 38 events with a loss figure, loss divided by market cap correlates with z at Pearson 0.86 (0.75 without Multichain, Spearman 0.56). In 17 of 31 worksheet-coded events, the exposure numerator is simply the reported loss, as the authors disclose. Proof q averages five yes/no checks collected within two days, partly overlapping the CAR window. The authors therefore treat q as a way to organize evidence, without claiming it was a predetermined treatment. Regress the discount on loss and the nine custody- or privacy-unverifiable events resemble the rest (Welch p = 0.46). Orchard's -0.57σ residual ranks 18 of 38. The event with 49-fold outflow thus lands mid-pack, 18 of 38, on loss-adjusted residuals.
Coding reliability remains open. The two LLMs reached kappa 0.65 on type, against only 0.002 to 0.052 on judgment fields. The authors leave model labels out of θ and q. We did not find reported human inter-coder agreement for those two fields.
We could not backtest this. Incident-level θ and q codes and per-block shielded-pool flows are required; the price and news data we hold have neither. Even before coding, the authors could link token prices to only 151 of 544 DeFiLlama seed incidents.
What would change the trading view?
The staged anchors make this a careful study of one event, with the flow series its most useful evidence. An unresolved-history disclosure would send us to pool exits, while returns alone would leave us unwilling to price the discount. A second privacy-pool case could change that view: the unresolved-history statement would need to arrive on a flat tape, with a discrete negative return and net outflow following it.