A Kalshi contract may become tradable on settled public facts before Kalshi declares the result. Money depends on the length of that window and what prices do inside it. Nechepurenko's paper supplies the framework for measuring both. It also reports candidly that the historical record cannot support the measurement.
The tradable clock
For market i, the paper applies a rule graph at time t. The graph contains contract and series terms, along with threshold and time-zone predicates. Combine it with the admissible public evidence available by t, and it returns a set of permissible settlement values. The market becomes first-decidable when that set contracts to one element and every contract-required future condition has been satisfied.
The second clock, stable decidability, can only be identified retrospectively. It marks the earliest time after which the mapping continues to return the same final value through every observed rule and source revision up to finalization.
This separation matters. A BLS release by itself does not settle a contract, since the same print can produce different admissible sets under different strikes, exception clauses and time windows. Kalshi's determination timestamp comes later, after review queues and waiting clauses. The paper expressly bars venue determination or final outcome from construction of the upstream clock. Otherwise, measured latency would be driven toward zero by construction.
Once first decidability is measured, the trade is easy to frame. Examine the order book between first decidability and determination, then ask whether the contract already trades at 99 or remains at 80. Two related definitions narrow the search. The bounded share Q_post is capped at 1 because decidability may arrive before market close. Stable decidability that lands after determination goes into a separate conflict class instead of becoming negative delay. Such cases mean Kalshi settled before the sequence preserved in the public archive supported the final value.
What the archive could recover
The retrospective study is complete, and it failed. Nechepurenko began with a frozen 25-market blind pilot. Identities were auditable for 25/25, blinding checks passed for 25/25, and current rule text was recovered for 25/25. Exact or bounded historical rule versions were available for 0/25. Exact or bounded official source-release objects were also available for 0/25.
The full recovery covered 152,694 exact ordinary tickers and 11,530 exact event identities through 6,540 read-only official requests. It found zero historically eligible events and zero historically eligible tickers. Those zero counts appear directly in the abstract.
The revealing comparison is 25/25 for current rule text against 0/25 for historical versions. We interpret that gap as evidence that rule pages were overwritten in place. The paper does not make that claim, and its stated limitation concerns source pages. Either way, the historical binding required by the method cannot be reconstructed from the available public record at the paper's accepted evidence grade.
Zero of 11,530.
Identification failed completely.
The paper draws the boundary carefully: "This is an observability result, not a claim that no market was decidable or that public evidence never existed." Its finding concerns retrieval. The record showing what the rules said at the relevant time is unavailable at the accepted evidence grade. Anyone hoping to search Kalshi history for pre-determination edges now has an inexpensive warning that the rule side of the join is missing. The prospective answer is already underway: production evidence enrollment is active, with the target sample size selected mechanically when enrollment closes.
Seventy-two hours of capture, no prices
The forward arm currently demonstrates infrastructure. From 2026-09-01T22:31:38Z to 2026-09-04T22:54:26Z, the system logged 260,568 seconds of valid observation and 781,266 lifecycle frames. Coverage spanned three source programmes, BLS CPI, FOMC target rate and Federal Register executive orders, across three events and 25 markets. Full integrity revalidation passed. Twenty-two reconnect receipts were classified as closed lower bounds, with no unresolved reconnect gap and no due-but-missed official release. Sustainability checks passed at 30, 42 and 60 days.
These engineering results answer a necessary question. Versioned rule objects and exact release objects can be archived quickly enough, with gap accounting sufficient for later blind adjudication. The evidence supports that conclusion for three days, on three US macro and regulatory feeds.
The same 72-hour run collected concurrent public Polymarket metadata. It produced 1,405 preliminary cross-venue candidates and 0 semantically accepted matches. Its scope remains three days and 25 markets. We reached a compatible conclusion from the Polymarket side in our note on Polymarket oracle events, where the adapter census was exact while the decidability clock remained absent.
Prices were missing. The paper labels the programmes observation capability only, while the Federal Register subset is conformance-only. The run therefore says nothing about mispricing during the decidability window. Three source programmes, three events and 25 markets also make for a narrow cross-section.
Will enrollment clear the bar?
The questions are registered, but the paper reports no answers: no clock values, no latency estimates, no adjudication outcomes and no price results. The prospective sample will be formed only when enrollment closes. Registered thresholds then determine its size mechanically. Full support requires at least 500 blind-adjudicable markets, 100 independent release/event clusters, 10 rule/series families, 3 source classes and 80% exact-or-bounded rule/source support, producing n = 500. Reduced support requires 150 markets and 40 clusters. Exploratory support requires 50 markets and 20 clusters. Below that threshold, n = 0 and the study stops.
With n* = 500 under simple random sampling, the worst-case 95% margin of error for a proportion is about 0.044. The paper warns that stratification, clustering, weights and adjudication attrition may widen it.
Attrition deserves the closest attention. Each packet must move from sample to named source, then through release object, publication, first clock, stable clock and latency. A packet can disappear at every rung. Effective n could finish well below the frozen target while the headline sample size remains unchanged. Reliability is assessed on at least 20% of the sample, with a target minimum of 100 packets. Because the paper's operational formula caps that count at the realised sample, 100 remains a target rather than a result.
Two numbers would change my assessment. One is the weighted share of sampled markets that yield an exact first clock instead of an interval. Interval-only clocks merely bound latency, and the primary analysis excludes their midpoints. The other is the frequency of a positive stability gap, where the singleton changes before finalization.
Why we did not test it
We lack Kalshi contract prices, historical contract identities, an archive of versioned market rules and the official release objects named by those rules. Historical rule versions are the binding constraint. The paper's own 0/25 result shows that they cannot be recovered from the public record at its accepted evidence grade. Equity, ETF or futures data offer no valid substitute because the mechanism resides in contract text.
Nechepurenko has delivered a precise definition and a capture system that operated for three days. The paper's own line stays with me: a clock may be economically well-defined yet retrospectively unobservable. Anyone funding the data collection should know that in advance. No price result is reported here, and the price layer stayed empty, leaving nothing to trade against yet.