For a financed event product, 182 seconds is the figure that matters. It is the median interval between an observed protocol payout record and a first redemption across 99,283 exact-linked Polymarket conditions. The result argues against pricing multi-day holder-delay carry. Two caveats come with that reading. By the snapshot, 7,125 conditions, or 6.56% of the linked cohort, show no redemption at all. And the statistic belongs to the condition, rather than serving as a clock for any particular holder.
What Nechepurenko actually built
Polymarket settles markets through the Gnosis Conditional Tokens Framework. Each market is registered as a condition that binds an oracle, a question identifier and an outcome-slot count. Collateral is split into complementary ERC-1155 outcome tokens. After an oracle answer is consumed, an adapter writes ordered payout numerators into the CTF contract. Holders can then call redeem and exchange their tokens for collateral.
In the paper's terminology, the numerator write marks protocol finality, when positions become redeemable. Four things happen around that point and are commonly compressed into one word: "resolved". The paper's spine is the separation among oracle finality, protocol finality and holder realization. None identifies the others.
The paper measures each stage from contract logs. Three event families are selected by topic hashes derived from the pinned contract's application binary interface over a fixed block interval: ConditionPreparation, ConditionResolution, PayoutRedemption. The upstream cohort is frozen at Polygon block 79,721,080. It inherits 185,550 initialized adapter-question instances and 185,549 unique question IDs from Part I, previously reviewed as an exact request-generation census with no decidability clock. Follow-up ends at block 90,114,204, timestamp 2026-07-12T17:11:41Z. Questions are linked to conditions solely through protocol identifiers. Title similarity and nearest-time matching appear only in the negative controls.
The resulting bridge contains 108,638 unique linked conditions, with 108,636 mapping one-to-one. An observed ConditionResolution exists for 99,283, or 91.4%, by the snapshot. Within that resolved risk set, 92,158 have a redemption of any amount, leaving 7,125 censored. A strictly positive payout appears for 91,817, leaving 7,466 censored.
Measured from first protocol resolution, the Kaplan-Meier medians are 182 and 200 seconds. Their interquartile ranges are 52 to 544 and 58 to 720 seconds. The linked cohort's payout vectors comprise 53,847 canonical (0,1), 45,024 canonical (1,0), 410 fifty-fifty, two other valid, and 9,355 with no observed resolution. The total target-event union contains 1,857,116 resolved conditions. The paper never uses that union as a substitute for the cohort.
The paper reports no returns; its product is a settlement-state surface for collateral accounting.
Three minutes, and what the three minutes is measuring
Only 341 of 108,638 conditions, or 0.31%, finish with a zero-only first redemption and no later positive event. The 18-second gap between the any-redemption and positive-redemption medians deserves attention. For this cohort, the multistate machinery in section 6.5 mainly fixes the zero-only class at 341.
The survival curve cannot supply a per-holder clock. Time zero is the condition-level first resolution, while the event is the first redemption by any address at the condition level. So 182 seconds measures when any address first redeems on that condition. A per-account clock would require per-position redemption histories, and those histories do not appear in the tables.
The 7,125 non-redeemers
The 7,125 conditions without a redemption by the snapshot represent 6.56% of the linked cohort. Treating them as right-censored is correct. A reserve model still needs to know whether they consist of dust or real stranded collateral, and redemption events alone cannot answer that question. The paper says so directly.
Theorem 6 formalizes the limitation. Identical redemption histories can coexist with different outstanding balances of winning positions, leaving the entitlement-completion ratio unidentified without a balance-complete reconstruction. G-VALUE, the paper's label for the gate that requires a balance-consistent entitlement denominator, therefore remains blocked. No value for unredeemed collateral is reported. The abstract makes the same disclosure.
One number sharpens the concern. Exact-linked resolved conditions have median available follow-up of 25,254,353 seconds, roughly 292 days, and a maximum of 142,006,845 seconds. For the typical condition, censoring is no short-window artifact. We did not find a separate exposure distribution for the 7,125 censored cases. That cut would show whether the non-redeemers are recent or genuinely aged.
The ordering claim exceeds its evidence
The paper's architectural centrepiece concerns adapter-terminal and protocol-resolution logs emitted within the same transaction. Canonical log indices determine their order; timestamps cannot. Proposition 1 establishes that point cleanly.
The conservative classification across 108,638 linked conditions yields 823 interval-qualified terminal generations. Another 48 have multiple candidates, 91,638 have no compatible terminal generation in the frozen evidence, and 16,129 are right-censored or unevaluable.
Hypothesis 2, the atomic ordinary path, is therefore untestable on this release, for reasons the paper spells out. The abstract gives all four counts. The conclusion treats non-evaluability as a finding: timestamps and identity joins cannot repair missing cross-contract ordering keys. The paper advertises the architectural claim and concedes on the same page that its ordering evidence is absent. Epistemically, that is the right answer. For replication, it leaves an implementer needing adapter/protocol emission order to re-acquire receipts at log-index grade. The 823 qualifying cases provide too thin a base for generalisation.
Where the specification ends
Of 185,549 unique question IDs, 76,912, or 41.5%, remain unmatched in the exact condition mapping. Section 6.1 defines the attrition states that the mapping report must separate, while Table 7 supplies the aggregate. Whether the 41.5% reflects pre-deployment truncation or linkage failure determines whether cohort-level coverage can be quoted at all. Table 7 leaves that unresolved.
A redemption appears without an observed resolution for 3,669 conditions, or 3.38%. Assumption 1 says an accepted redemption requires a previously recorded payout rule for the same condition. From that assumption, Lemma 8 derives the ordering: every accepted redemption occurs at or after protocol finality. The class is consequently an observation-boundary effect by construction. Section 5.11 gives every non-canonical path an explicit disposition, including left truncation, snapshot censoring, and five others. The lifecycle table supplies the count. The implementer must assign those 3,669 cases.
One more issue matters for anyone seeking an outcome base rate. The payout taxonomy deliberately ignores orientation, and the table caption says human-readable YES/NO labels are never used to infer slot order. Without the slot-order mapping, the 53,847 versus 45,024 split, about 54.5/45.5, cannot be interpreted as a YES frequency. Across the whole target-event union, neutral resolutions account for 29,609 of 1,857,116, or 1.6%. The linked cohort has 410 of 99,283, or 0.4%. That composition gap has to be understood before the universes are pooled.
We could not reproduce any of this. Reproduction requires Polygon event logs for the pinned CTF contract at block, transaction and log-index grade, together with the adapter-question to condition identity bridge. We hold crypto OHLCV and news. Substituting price bars would test nothing about redemption. The paper also withholds holder-level records, leaving the deterministic rebuild claim independently uncheckable from outside.
I would carry the 182-second median into a conversion or debt-extinction design. It argues against pricing much holder-delay carry. The leverage ceilings in an earlier note on conditional debt extinction point the other way. The censored tail could change my mind. If the 7,125 unredeemed conditions contain outstanding winning-token balances rather than dust, the median describes the easy cases and the reserve question reopens. This release contains no such figure, and the paper explicitly says it lacks the balance ledger needed for that test.