Key takeaways
- A clarification can affect the order book as well as the eventual payout. Polymarket International's published policy says a clarification clears resting orders. An advance announcement also clears the book when the announcement is made. A rule event therefore has an immediate operational consequence, even before the outcome is known.1
- An effective date is a contract-selection rule. Kalshi's weekly natural-gas series notice assigns markets settling from 4 September 2026 to a Pyth index and earlier settlements to the prior front-month methodology. Applying the latest series description to every historical contract would erase that distinction.2
- A timestamp establishes when a message was recorded; its authority must be established separately. The public bulletin-board implementation in Polymarket's UMA adapter stores updates by question and sender. Its design permits anyone to post, while directing readers to the question's initializer. Counting messages without authenticating their authors would count the wrong population.3
- Five observations can contain two texts and two changes. In our worked example, the sequence A, A, B, B, A has three periods of unchanged text. The first period is a baseline, and the return to A is a second transition. Neither a count of observations nor a count of distinct hashes measures the number of changes.
- The unit of review should be the affected conclusion. A new threshold, source or fallback can change a payout, a cross-venue relationship or a derived figure. Preserving the old terms is necessary; identifying which prior conclusions depended on them is the next step.
1. A market can keep its name while its reference changes
Kalshi's weekly natural-gas series carries a notice that makes a precise division. Markets settling from Friday, 4 September 2026 use the Pyth Natural Gas 24/7 index, identified as Index.NATGAS/USD. Markets settling before that date continue under the front-month natural-gas futures methodology. The notice was present in the series data retrieved on 31 October.2
The title remains Natural Gas Weekly. A reader following that title could reasonably think they were following one continuous question. The notice instead describes two groups of contracts with different reference sources. The boundary is their settlement date, not the date a data vendor happens to download them.
That distinction matters when a research platform rebuilds a chart, matches contracts across venues or checks an old decision. The current description answers what the series says now. To recover what an earlier contract referenced, the platform also needs the notice's applicability rule and the terms attached to that contract. A catalog update is the beginning of that investigation.
Our first research paper examined how similar questions can settle differently across venues. The same problem runs through time. A relationship that was reviewed against one set of terms may need another review after a source, deadline or contingency changes. The event still has the same public name; the reason two instruments were considered comparable may have changed.
There are legitimate reasons for maintaining terms. A data source can stop publishing, a reference can become unreliable, or an unforeseen event can expose an ambiguity. Kalshi's April 2023 FEDDECISION filing expressly reserves source-substitution authority in specified reliability or transparency circumstances. It separately says that instructions for locating the underlying are non-binding and may be clarified.4
Those two provisions illustrate why a list of changed words is insufficient. Replacing a broken link to the same Federal Reserve release and replacing the source used to determine the outcome are different events. A useful history preserves both and explains which one occurred.
2. What exactly changed?
The phrase rule change compresses several kinds of evidence into one label. Some records establish that the governing terms changed. Others establish that a venue explained its interpretation, that a webpage changed or that a collector produced different output. These deserve separate classifications.
For US registered entities, the regulatory process already distinguishes substantive amendments from specified non-substantive revisions. Under the ordinary self-certification route in 17 CFR 40.6, a submission includes marked additions and deletions and the intended implementation date. The usual review period is 10 business days; emergency provisions, stays and specified exceptions make the filing date alone insufficient to establish applicability.5
Polymarket International describes a different process. Its resolution documentation allows additional context after trading begins for unforeseen circumstances, says that such context cannot change the question's fundamental intent, and says it is published onchain for consideration during resolution. That is the venue's stated clarification policy. Whether an individual notice follows it requires examination of the actual question and notice.6
The clarification policy also affects the tradable book. Polymarket says it clears resting orders when a clarification is issued and, when there is an advance announcement, at that announcement too.1 A before-and-after quote comparison can therefore span both interpretive information and the removal of existing orders. A researcher assessing the episode needs to retain the notice and book-clearing times, then distinguish that mechanical removal from subsequent repricing. A disappearing quote alone does not establish that participants changed their beliefs about the event. The notice can matter operationally even when a reviewer finds that the underlying payout condition is unchanged.
The distinction between an amendment and a clarification should survive collection. Calling every notice an amendment risks overstating what the venue has done. Calling every economically consequential notice a harmless explanation risks understating its importance to holders.
| Record | What the evidence establishes | What a reviewer still needs to decide |
|---|---|---|
| Formal amendment with an applicability provision | The publisher states a change and identifies its intended scope | Which listed contracts and periods fall within that scope |
| Additional context or clarification | An authorized publisher has supplied an interpretation | Which contingencies it addresses and whether prior comparisons remain supported |
| Source or benchmark notice | A reference input, method or publication process has changed | Whether the contract incorporates that change or supplies a fallback |
| Display or access-instruction edit | Text shown to users has changed | Whether the governing economic definition changed with it |
| Collector or parser revision | The local representation changed | Whether the source changed at all |
Source: Predict Here classification, informed by 17 CFR 40.6, Kalshi's FEDDECISION filing and Polymarket International's resolution documentation. These are analytical categories, not legal determinations.
Economic significance depends on the payoff, not the number of edited characters. A source URL might change completely while continuing to identify the same release. Changing above to at or above adds only a few words but changes who is paid when the fixing equals the threshold.
A reviewer can make that question concrete: identify a possible set of observations under which the old and new readings would produce different payouts. Finding such a case establishes a consequential distinction between the readings. It does not establish how likely that case is, whether the change is legally effective, or whether any holder has suffered a loss.
This is also a useful way to avoid false precision in automated scores. A long textual difference need not imply a large financial effect. A short difference can change the entire payout in a narrowly defined state.
3. Publication, applicability and observation run on different clocks
A change record needs more than one timestamp. We propose recording four: publication, applicability, first observation and review. They answer different questions.
Publication concerns when the authorized source issued the statement. Applicability concerns the contracts or periods to which it applies. First observation concerns when the collecting system captured it. Review concerns when a person or defined process assessed its consequences. A dated notice can establish one of these without establishing the others.
The natural-gas record provides a compact example. The retrieved series metadata includes an update timestamp of 29 October 2026, while the notice's boundary is markets settling from 4 September. The series timestamp is not labeled as the first publication time of that particular notice. Assigning it that meaning would be an unsupported inference.2
| Time or scope field | Evidence in the 31 October capture | Proper interpretation |
|---|---|---|
| Applicability | Markets settling from 4 September 2026 use the new reference; earlier settlements retain the prior method | A boundary selecting the covered contract iterations |
| Series metadata update | 29 October 2026, 20:10:20 UTC | The supplied update time for the series record |
| Research observation | 31 October 2026 | The notice was present when this research captured the record |
| First publication of this notice | Not established by this capture | Requires a prior authenticated notice or historical record |
Source: Kalshi Trade API v2, KXNATGASW series response. The companion commodity paper examines the economic differences between the references. This exhibit concerns chronology and scope.
Preserving these distinctions changes what a historical review can say. A later download may establish that a venue now describes a rule as applicable to an earlier period. It does not establish that a participant could have read that description at the time. An assessment of information available on a particular day needs evidence available by that day.
The same applies to market APIs. Kalshi's market response exposes an update timestamp and primary and secondary rule fields. Polymarket's market schema includes a description and an update timestamp among many other fields. Those interfaces are useful snapshots. A record-level timestamp, by itself, does not identify the field that changed or preserve its earlier contents.78
There is a practical consequence for alerts. Suppose an analyst observes a new notice after its stated applicability boundary. The alert should retain both dates. It should not relabel the earlier boundary as the detection time, or rewrite an old review as though the new notice had already been in its evidence set.
History must retain what the source said and what the observer knew. Collapsing those two histories creates a clean-looking record at the cost of losing the question the record was meant to answer.
4. An onchain message still needs an author
Public transaction records can make a notice's text and recording time independently inspectable. They do not eliminate the need to establish who had authority to issue it.
The BulletinBoard component in Polymarket's public UMA adapter repository illustrates the distinction. The reviewed implementation allows any caller to submit an update. It stores updates under a key combining the question identifier and the sender, records the block timestamp and emits an event containing the question, sender and text. Its documentation directs readers to updates from the question's initializer.3
The sender is therefore part of the evidence, not incidental metadata. In an analysis of that implementation, a query for every update mentioning a question would collect a broader population than authorized question updates. The correct filter requires the applicable contract deployment, the question identifier and the relevant authorized address.
This is a property of the public code reviewed here, not evidence that unauthorized notices were accepted in any live market. The code deliberately preserves the distinction an analyst needs to make. A production study must also verify which deployed adapter and authority rules applied to the market being studied.
An authenticated notice also has to be linked to the right instrument. One event may contain several dated markets, and one question may sit inside a wider outcome structure. Matching a notice to a familiar title can attach it too broadly. The relationship should be established through the venue's identifiers and the notice's stated scope.
These requirements are not unique to onchain records. A PDF obtained from an official exchange host still needs its version, filing status and covered contracts. A screenshot preserves what was visible on a page, but may omit a linked amendment. Every medium solves part of the provenance problem. None substitutes for the complete record.
The useful standard is consequently modest and demanding: preserve the document, establish its author, identify its authority and retain the basis for assigning it to the contract. Only then classify the economic change.
5. Why counting alerts can miscount changes
Even with authenticated sources, a frequency statistic needs a defined unit. Observations, distinct texts, version periods and transitions are different quantities.
Consider a deliberately small example. A collector checks one contract every five minutes and records A, A, B, B, A. A and B stand for complete captured texts. The last A is byte-for-byte identical to the first.
| Observation time | Captured text | New period of unchanged text? | Change relative to the previous observation? |
|---|---|---|---|
| 09:00 UTC | A | Yes: initial baseline | No prior observation |
| 09:05 UTC | A | No | No |
| 09:10 UTC | B | Yes | A to B |
| 09:15 UTC | B | No | No |
| 09:20 UTC | A | Yes: return to earlier text | B to A |
Source: Predict Here illustration. These are invented records, not a venue sample. There are five observations, two distinct texts, three version periods and two observed transitions.
A system that emits one alert for each new version period produces three alerts, including the initial baseline. Describing all three as changes would overstate the two observed transitions by 50%. A system that globally deduplicates by text hash has the opposite problem: it can suppress the return to A because it has seen that text before. Both errors arise before anyone interprets the words.
The observation schedule also limits timing precision. This example locates the first observed transition between 09:05 and 09:10 and the second between 09:15 and 09:20. It does not show the exact publication time, or any intermediate text that appeared and disappeared between checks.
Now add meaning. If A and B differ only in formatting, the two observed transitions are textual changes without an established payoff change. If B changes the threshold's treatment of equality, both transitions have an identifiable economic consequence. The observation counts remain identical.
A responsible frequency study would therefore report at least three separate measures: authenticated textual transitions, reviewed changes affecting settlement interpretation, and the proportion of monitored contracts affected. It would also publish the observation window and coverage, separating first captures from transitions and reversions from duplicates.
The denominator matters. A contract observed once cannot establish that it remained unchanged over its life. A feed outage is not a period of confirmed stability. A monitor that begins after a contentious notice will record the resulting text as its baseline, not recover the earlier transition. Those contracts may belong in a catalog count, but cannot all support the same longitudinal claim.
This paper does not publish a venue-wide change rate. The examples instead establish the rules a defensible rate would need to follow. An impressive numerator without those rules can be precise and still measure the wrong thing.
6. The consequence belongs to a particular state
To see why a changed word can matter, take an illustrative $1 contract with a threshold of $80. Under version A, YES pays when the reference is strictly above $80. Under version B, YES pays when it is at least $80. Assume the source, observation time and measurement precision are identical.
| Recorded reference | Version A: above $80 | Version B: at least $80 | Difference for 100,000 YES contracts |
|---|---|---|---|
| $79.99 | $0 per contract | $0 per contract | $0 |
| $80.00 | $0 per contract | $1 per contract | $100,000 |
| $80.01 | $1 per contract | $1 per contract | $0 |
Source: Predict Here arithmetic on hypothetical terms. The table reports conditional settlement receipts, not a change in market value, a profit estimate or an observed venue amendment.
The $100,000 is the difference in receipts if the fixing is exactly $80. It is not the economic value of the wording change today. Under a simple undiscounted expected-payout model, that value depends on the probability assigned to the exact boundary. A 5% illustrative boundary probability would produce a $5,000 expected-receipt difference for these 100,000 contracts. The assumed probability is not a market estimate.
The example gives a reviewer two separate tasks. The first is logical: establish the states in which the readings differ. The second is empirical: establish whether those states are relevant to the actual measurement process. Rounding, discrete publication increments and data revisions can all matter to that second task.
It also explains why an old cross-venue comparison may need to change. Suppose a second venue's contract still uses strict inequality. A prior claim of equivalence depended on both sides treating the boundary the same way. Once one side uses an inclusive threshold, the previous justification no longer covers every possible fixing.
The correct response is to review that relationship against the new evidence, preserving the earlier decision and its supporting version. A price series can remain numerically continuous while the set of observations that produces a payout changes underneath it.
7. A rule history should follow the decisions that depend on it
The practical purpose of a change ledger is to identify what needs attention. A notice about a settlement source may affect a contract description, a venue comparison, a scenario calculation and a published research table. The same record can matter to all four for different reasons.
We propose linking each reviewed change to the conclusions that use the affected terms. This is a dependency record: the contract version supports a relationship; that relationship supports a derived result; the result supports a particular review or publication. A change can then be assessed at the point where it alters the reasoning.
| Stage | Record to preserve | Question it answers |
|---|---|---|
| Capture | Original content, source identifier and observation time | What did the system receive? |
| Authenticate | Publisher, authority and relevant venue or deployment | Whose statement is it? |
| Classify | Before-and-after reading, change category and rationale | What changed, and why does it matter? |
| Apply | Contract scope, effective boundary and fallback provisions | Which instruments are affected? |
| Trace | Referenced mappings, calculations and review versions | Which conclusions relied on the previous terms? |
| Resolve | New review, retained prior record and correction where required | What is the supported conclusion now? |
Source: Predict Here proposed review model. The table describes the complete model, not a claim that every stage is available in the current product.
Predict Here's current public governance statement draws a narrower implementation boundary. It describes append-only rule and metadata hash versions, raw source payload retention where collected, and signals for newly observed hashes. It explicitly distinguishes source-change detection from legal interpretation. The full revision and retirement workflow remains incomplete as a customer-facing process.9
That distinction is useful. A change alert is evidence that something merits examination. It should not become an automatic declaration that the contract was unlawfully altered, that every mapping is invalid, or that a previous relationship remains safe to use. Those are separate conclusions with separate evidentiary requirements.
The opposite failure is to leave the alert detached from the user's work. A history containing every observed version but no link to the comparisons built from it becomes an archive that someone must search manually. The larger opportunity is to preserve the path from terms to conclusion, so a reviewer can see exactly which assumption now needs attention.
8. Our take
Contract maintenance is part of market structure. Sources fail, ambiguities surface and venues issue clarifications. The quality question is whether the resulting history makes scope, authority and consequence legible.
Record the clocks separately. A first observation is not a publication date. A filing date is not automatically an effective date. A current description is not proof of what a participant could have known earlier.
Count transitions before interpreting them. Baselines, duplicates, reversions and collection gaps determine what an alert statistic actually measures. Semantic review then determines which observed transitions matter to settlement.
Preserve the old conclusion and review the new one. Event identity needs a time dimension. A comparison should remain tied to the terms that supported it, with later evidence producing a visible new assessment rather than a silent rewrite.
That is the role of a useful contract-rule ledger. It turns a stream of notices into a history that can explain why a contract was understood one way at one time, why that understanding was revisited, and what the evidence supports now.
Methodology and limitations
Scope. This is a source-based study of change records and their interpretation, with original illustrative calculations. It is not a census of amendments, a study of realized holder losses or an assessment of the legal validity of a particular notice.
Natural-gas notice. We inspected the KXNATGASW series response captured on 31 October 2026. It establishes the notice present in that response and its stated applicability boundary. It does not recover the notice's first publication time, prove when each participant saw it or establish a retrospective edit to an already open contract.
Public code. The bulletin-board analysis concerns the public source file linked below. We preserved the retrieved source and its hash. We did not match a complete historical set of deployed adapters, question authorities and update transactions. No frequency or misuse claim is inferred from the implementation's ability to accept messages from multiple senders.
Illustrations. Exhibits 3 and 4 use invented records and terms. The optional 5% boundary probability is an assumption chosen to distinguish conditional payout from expected receipts. The arithmetic excludes fees and discounting and is not a pricing model for any identified contract.
Historical coverage. No production customer records or complete signal-history dataset were used. Observed API fields and a changed hash cannot establish an exhaustive amendment history. The proposed workflow in Exhibit 5 extends beyond the product capabilities described in the current public governance statement.
Regulatory sources. The discussion of 17 CFR 40.6 describes selected procedural distinctions relevant to recordkeeping. Actual applicability depends on the rule, filing route, notices and circumstances. It is not legal advice. Public help pages and API documentation can change; retained captures identify what was checked for this paper.
Notes and references
- Polymarket Help Center, How Are Markets Clarified?, dated 29 June 2026; checked 31 October 2026. The policy describes cancellation of resting orders at clarification and, when applicable, at advance announcement. It does not say that already executed positions are cancelled. ↩
- Kalshi Trade API v2, KXNATGASW series record, captured 31 October 2026. Relevant fields:
product_metadata.important_info.markdown,settlement_sourcesandlast_updated_ts. The exact returned series timestamp is 2026-10-29T20:10:20.059516Z. The source notice divides contracts by settlement date. ↩ - Polymarket, BulletinBoard.sol in the UMA CTF adapter repository, checked 31 October 2026. Functions
postUpdate,getUpdatesandgetLatestUpdatepreserve the question/sender distinction. The file's latest commit returned by the repository API was dated 31 May 2023. The analysis is of this source implementation; deployment-specific authority must be verified separately. ↩ - KalshiEX LLC, FEDDECISION product certification, 5 April 2023, PDF pages 3, 7 and 8. Source-substitution authority is conditional on circumstances affecting reliability or transparency. Access instructions are expressly distinguished from binding terms. This historical filing is cited for those provisions, not as proof of every current series specification. ↩
- Electronic Code of Federal Regulations, 17 CFR 40.6, Self-certification of rules, checked 31 October 2026. See (a)(3), (a)(6), (a)(7)(ii)–(iii), (b), (c) and (d)(2)(i). See also the CFTC's Rules and Rule Amendments overview. Ordinary self-certification, emergency rules and non-substantive notifications have different procedural conditions. ↩
- Polymarket International, Resolution documentation, “Clarifications” section, checked 31 October 2026. The discussion concerns the International venue's documented clarification process and does not extend that process to Polymarket US. ↩
- Kalshi, Get Market API documentation, checked 31 October 2026. The response schema contains
updated_time,rules_primaryandrules_secondary; their coexistence does not identify which field caused a record update. ↩ - Polymarket International, Get market by ID API documentation, checked 31 October 2026. Relevant fields include
descriptionandupdatedAt. The historical-inference limits discussed in the text are our analysis of snapshot data. ↩ - Predict Here, Methodology governance, checked 31 October 2026, especially “Rules, methodology, and historical state” and “Implemented review controls.” This is the source for present product boundaries. The broader review model in this paper is a proposal. ↩
Disclosures
This paper is general research published for information only. It is not investment, legal or tax advice, and it does not recommend any position, venue or strategy. References to other organizations describe their published work and do not imply endorsement of, or a relationship with, Predict Here. Predict Here is operated by Odyssea Research LLC. It does not list markets, hold client funds, take the other side of trades, or execute or route orders.
Persons associated with Predict Here may hold positions in event contracts, including on venues discussed in this report.