Research
Pricing the Unprovable: Ambiguity, Turnstile Runs, and the Verifiability Discount in Crypto Protocols
Prices the value erased when a privacy protocol patches a vulnerability without proving whether it was exploited, backed by Orchard and Zcash flow evidence.

- arXiv
- 2610.21540
- Published
- 2026-10-15
- Authors
- Doohwi Cha, Hanhee Kim, Hyeonjun Yeo, Minsuk Sung, Minjae Lee, Seunghan Son, Donghwa Seo, Juyeong Lee
AI summary
Overview
Research area: Crypto/blockchain economics (q-fin.TR) — the intersection of market microstructure, decision theory under ambiguity, security disclosure, and zero-knowledge/privacy protocol design.
Technical level: Advanced. The paper combines a formal two-date game-theoretic "turnstile" model with maxmin preferences, an hourly/daily event study using BTC- and ETH-adjusted abnormal returns, and reconstructed on-chain shielded-pool flow data.
Scope (one sentence): The paper defines and empirically probes a "verifiability discount" — the value reduction investors apply when a crypto protocol patches a flaw but public evidence cannot rule out that private state was already abused — using the June 2026 Zcash Orchard disclosure sequence, earlier Zcash cases, and a 40-event security panel.
What This Paper Is About
When a crypto protocol is hacked or patched, investors normally ask two questions: how much was lost, and is the flaw fixed? Privacy-preserving protocols such as Zcash add a third: can outsiders verify what happened before the fix? A patch can stop new abuse and a supply rule can protect total supply, yet private transaction history may leave past abuse impossible to rule out — finding no abuse is weaker than proving abuse could not have occurred.
The paper names the market consequence of that gap the verifiability discount, builds a model explaining why unresolved history drives both a price haircut and exit behavior, and introduces a reproducible measurement framework (VDISC) that keeps reported loss, exposed value, public evidence, returns, and pool flows as separate fields.
Key Contributions
-
A new pricing construct. The verifiability discount separates reported loss from adverse histories that public evidence cannot close, treating the latter as ambiguity in the decision-theoretic sense rather than as ordinary volatility or risk.
-
A formal turnstile model. A two-date model in which exposure θ, proof strength q ∈ [0,1], and residual unverifiable exposure z = θ(1−q) determine both a direct price haircut and exit pressure. Early withdrawal shifts part of a possible shortfall onto remaining holders. Proposition 1 shows exit intensity rises locally with z on stable branches; Proposition 2 shows the value of proof scales with exposed value (cross-partial ∂²[ρθ(1−q)]/∂θ∂q = −ρ < 0).
-
The VDISC measurement framework. A reproducible framework that records source and time, reported loss, exposure, public proof, returns, and pool flows as separate fields, with fixed coding rules — including a five-check empirical q averaged over a two-day window (supply proof, containment, replayability, migration proof, official residual-uncertainty statement), each requiring a public URL and quote.
-
A staged empirical design with an explicit identification boundary. Four timestamped Orchard anchors paired with hourly BTC-adjusted CARs and reconstructed net unshielding, plus earlier Zcash cases, 14 coded disclosures, and a 40-event panel that documents why returns alone cannot separate reported loss from unresolved history.
Main Findings
-
Orchard follows the predicted ordering. CAR was 0.185 after the initial containment notice, 0.014 after the NU6.2 repair statement (no evidence of exploitation, supply unaffected), 0.016 after the Foundation's confirmation, and −0.162 only after the context post stating prior exploitation was judged unlikely but could not be ruled out. Pre-window CARs were 0.052, 0.185, 0.148, and −0.134 respectively.
-
The final anchor's negative return is suggestive but not causally clean. For the context post, p_ob = 0.003 and q_BH = 0.051 across a 16-test family (4 anchors × 4 statistics) — clearing a conventional 5% threshold only before multiple-testing adjustment. The pre-anchor CAR was already −0.134, reaching the same placebo floor.
-
Timing is not the story. The statement first appeared as an X article at 21:17 UTC, 16 minutes before the forum post; reanchoring changes minute CAR from −0.166 to −0.205 without changing sign. Same-window CARs on Coinbase, Kraken, OKX, and Bybit range from −0.155 to −0.166. CAR had already fallen 0.208 since midnight (0.188 by 21:17), and the post-anchor 60-minute decline was about 1.6 times the pre-anchor decline.
-
The largest net outflow coincides with the widest acknowledged uncertainty. Net unshielding (ZEC leaving the aggregate Sprout/Sapling/Orchard pools minus entering) was −7 during the soft-fork pause, 7,573 after NU6.2, 9,052 after the Foundation post, and 175,568 ZEC in the 24 hours after the context post — about 49 times the 3,560 ZEC/day May 15–28 baseline.
-
A single block dominates but does not explain the flow. One block on June 5 at 19:34 UTC (20 transactions) nets 60,324 ZEC, 34% of the 24-hour total; without it, the 24-hour outflow is still about 32 times baseline. Over 48 hours from the first complete 22:00 UTC bin, flow totalled 243.2k ZEC, or 4.8% of the aggregate shielded balance. Orchard accounted for 97.4% of the May 24–June 7 aggregate shielded decline.
-
Reported loss and unresolved exposure are confounded. Among the 38 events with a public loss figure, market-cap-normalized reported loss and unresolved exposure co-move (Pearson 0.86, or 0.75 without Multichain; Spearman 0.56) — partly because in 17 of the 31 worksheet-coded events the exposure numerator equals the reported loss. Residuals from a regression of the discount on loss/mcap do not differ between the nine custody- or privacy-unverifiable events and the rest (Welch p = 0.46), and Orchard ranks 18/38 (residual −0.57σ).
-
The cross-event slope's sign is not identified. The CAR–z slope is −0.172 on calendar days, −0.097 with Orchard alone re-timed, and +0.027 when all nine priced events are re-timed, with pairs-bootstrap 95% intervals of [−2.237, −0.004], [−2.103, 0.282], and [−0.402, 0.306].
-
Within-Zcash direction matches the model. The complete-proof Sprout 2026 case (z = 0.000) shows the aggregate pool growing (435 out/day baseline → 14,026 inflow), while two unresolved cases show outflow relative to baseline: Sprout 2019 (z = 0.020, 8.8k/day → 25.4k out, ≈2.9×) and Orchard 2026 (z = 0.162, ≈49×).
-
Model-assisted coding is not yet reliable for judgment fields. In a frozen 50-event audit of GPT-5.4 Mini and Qwen3.7 Plus, 99 of 100 outputs cited a source with an evidence summary, and inter-model Cohen's kappa was 0.65 for type and 0.58 for severity — but only 0.002–0.052 for judgment-dependent fields.
-
Cross-market checks show breadth without uniformity. Binance contract open interest rises 21.8% while its dollar value falls with price and spot market-order flow remains net selling; XMR/ETH weaken more, DASH improves, and SOL is nearly flat.
Methodology in Plain English
The authors work in three layers.
First, a simple model. Imagine a pool of claims where some fraction of hidden claims might be invalid. Exposure θ is the share of the claim base vulnerable to the flaw; proof strength q is how much public evidence rules out bad histories; the leftover z = θ(1−q) is the maximum unresolved claim value. Investors are assumed to evaluate the worst publicly admissible case (maxmin) rather than assigning probabilities. Each holder chooses to exit or stay. Exiting costs an individual friction (tax, lost privacy, inattention), but staying means sharing any shortfall with fewer people. A "turnstile" — the Zcash mechanism limiting transfers out of each shielded pool to its recorded balance — allocates the possible shortfall. The model's key comparative static is local: on a stable branch, more unresolved exposure means more exit.
Second, an event study. Abnormal returns for ZEC are computed by subtracting a BTC-linked market model, using 335 pre-discovery hours ending May 28 at 23:00 UTC, with matched pre/post windows containing 24 complete hourly returns. The authors run four anchors: the public notice (June 2, 06:20 UTC), the NU6.2 repair statement (June 3, 04:05), the Foundation confirmation (June 3, 10:55), and the context post (June 4, 21:33:53). They layer on minute-level reanchoring, cross-venue replication, a placebo test using 312 overlapping pre-discovery blocks, Corrado rank tests, GARCH diagnostics, and log-return and 60/90-day window variants.
Third, on-chain reconstruction and cross-event coding. Per-block changes reconstruct June 1–7 daily and hourly shielded-pool flow, benchmarked against independent snapshots agreeing within 0.1%. Then VDISC's fixed coding rules are applied to 14 disclosures and a 40-event panel (an earlier coverage layer links 151 of 544 DeFiLlama seed incidents to token prices, defining availability rather than an event-study sample). Safety and reproducibility are handled by restricting collection to public timestamped sources and excluding exploit instructions, payloads, keys, and live-system procedures.
Throughout, the authors are explicit that the design tests directions, not structural parameters: net pool flow is a directional proxy for gross exit mass m, and prices, pool flow, and order flow are complementary rather than interchangeable evidence.
Why This Matters
Impact on research. The paper formalizes a category that existing security-disclosure event studies do not have: a patched flaw with protected supply whose private past remains unverifiable. It also documents a methodological trap — when exposure is coded from the loss itself, returns cannot separate the two channels — that should temper a broad class of crypto event studies. The explicit identification boundary (a purposive 14-case sample, a panel that cannot sign the slope, a final anchor that bundles severity with unverifiability) is unusually candid.
Real-world applications:
- Post-incident disclosure practice. Protocols can treat a residual-uncertainty notice as a distinct disclosure item rather than folding "we found nothing" into "nothing happened," keeping proved, unproved, and unknown claims separate.
- Privacy-preserving proof design. The design map points to state replay, aggregate-balance proofs, and containment evidence as ways to raise q without deanonymizing users.
- Risk and treasury management for holders and funds. The turnstile mechanism explains why early exit can be rational even absent confirmed theft, informing withdrawal-queue and liquidity planning around disclosures.
- Exchange and custodian monitoring. Net unshielding around privacy-pool disclosures gives a directional signal of holder movement that is observable even when individual transfers are hidden.
Industry relevance. For privacy-focused protocols, the paper argues that repair should ship evidence alongside corrected code: limit the state one flaw can expose (θ), prepare evidence that bounds past abuse (q), and state what remains unknown. For market participants, it reframes "no theft reported" as a weaker statement than "theft ruled out," which affects how a disclosure should be priced.
Future Directions
-
Building independently sourced exposure measures. Because in 17 of 31 worksheet-coded events the exposure numerator equals the reported loss, the paper's own remedy is a separately sourced exposure field; until then the 0.86 correlation partly reflects construction, not market behavior.
-
Tightening the flow-to-model mapping. The authors note that transaction-level records, protocol timestamps, or cryptographic commitments would map withdrawals to the model's gross exit mass m more tightly, supporting nonlinear or adaptive extensions beyond the current directional proxy.
-
Testing the untested prediction (P3). The paper explicitly sets aside an extension in which AI-assisted vulnerability discovery changes perceived arrival rates — outside the two-date model and not estimated.
-
Improving model-assisted coding for judgment fields. Inter-model kappa of 0.002–0.052 on judgment-dependent fields, versus 0.65 for type and 0.58 for severity, marks where automated annotation remains future work.
Target Audience
Best suited for: researchers in crypto and DeFi market microstructure, financial economists working on ambiguity and disclosure, and protocol designers or security teams at privacy-preserving systems who need to reason about what a post-mortem can and cannot prove.
Also valuable for: quantitative analysts and risk managers at exchanges, funds, and custodians who price disclosure events; and methodologists interested in a worked example of an honestly bounded event study, where the authors repeatedly decline to convert directional evidence into causal claims.
Authors’ abstract
Privacy-preserving protocols can repair a vulnerability without establishing whether it was exploited beforehand, even when no theft is reported and aggregate supply is protected. We call the potential value reduction from this gap the verifiability discount. A turnstile model shows how early withdrawal can shift a possible shortfall to remaining holders and why credible evidence can become more valuable as exposure grows. We introduce VDISC, a reproducible framework that…
The full abstract on arXiv