Separate the estimate from the evidence.

A mention market asks whether a designated speaker will say a word or phrase during a specified event. Its exact contract rules determine which statements and events qualify; a word appearing in a transcript is not, by itself, a settlement decision.

A 70% model estimate means an estimated 0.70 chance that the contract resolves YES under its exact rules. It is not a guarantee, an accuracy score, or a return.

The operational pilot is Federal Reserve press conferences. Contracts from the same physical event are dependent; ten words in one speech are not ten independent events. Unsupported special structures and missing timing remain visible as exclusions.

The declared horizon uses a corroborated scheduled start, S. The information cutoff C is 65 minutes before S. The publication deadline D is 60 minutes before S.

At or before C

Capture admitted source versions and baseline quotes. An old page date cannot excuse a later capture.

Strictly before D

Observe the original forecast’s actual public bytes. A timestamp equal to the deadline is too late.

After publication

Keep the original probability, inputs and baseline. Later quotes and page edits do not replace them.

Scheduled start is not actual-start evidence. If the event starts earlier, append a timing qualification correction; never backdate a forecast. Unknown timing is a visible limitation.

On an event page, Open captured artifact points to the retained bytes. Original URL · may have changed opens today’s source page. They are not interchangeable. A retained excerpt is not necessarily a full source archive.

The market baseline is the midpoint of the latest admitted, non-crossed two-sided YES quote at or before C. The declared policy requires a quote no more than five minutes old and a spread no larger than $0.10. Otherwise, the market baseline is missing.

Source quality is not a reason to buy

Synthetic teaching example: a 55% model estimate with a YES ask of 56¢ (bid 48¢) does not recommend buying YES—even if the evidence quality is high. The model’s estimate is below the executable ask before fees. Evidence quality describes support, not price advantage.

The ledger does not establish order-book depth, fills, fees, execution delays, or profitable trading. Current executable prices must not be inferred from a frozen quote.

An event displays the achieved tier, not the tier we would like it to have. A hash identifies bytes; it does not, by itself, prove when they existed publicly.

1. 1

   Internally recorded

   internal_recorded

   Self-recorded metadata describes claimed execution order. It does not independently prove timing or public availability.

2. 2

   Platform reader receipt

   platform_reader_receipt

   The operator’s anonymous reader observed matching public bytes. Retained response headers and observation times are not independent clock proof.

3. 3

   External pre-event witness

   external_pre_event_witness

   An actual independent receipt must bind the exact core hash before D. A self-authored date, a generic repository timestamp or an after-event archive is insufficient.

Finalized binary exchange results are the scoring labels: YES = 1, NO = 0. For each contract, Brier error is (p − y)². Lower is better.

Average contract errors within each physical event, then weight events equally. Compare paired model-minus-baseline event errors on the same declared cases. A negative paired difference means lower model error on that cohort—not a proven trading edge.

• Pending: no finalized label, so no invented result. Primary scores wait for complete supported, on-time issued event outcomes.
• Abstained or missed: retain coverage failures. An unpublished fallback does not become a headline forecast.
• Missing baseline: disclose the unmatched case. Keep it out of the paired comparison, not out of the opportunity ledger.
• Excluded: apply predeclared support and timing rules, not retrospective removal of bad predictions.
• Synthetic: exercise arithmetic and record handling; never count as real performance evidence.

Calibration, probability skill, arithmetic reproduction, public timing and trading value are different claims. Reliable model promotion needs a registered forward comparison with enough independent events for the declared effect and uncertainty rule. A small sample may remain inconclusive.

Append exchange corrections and score revisions. Keep the old observation and original model probability. Transcript review can explain support or ambiguity; it cannot silently remove an unfavorable finalized label.

The reproduction package contains forecast opportunities and dispositions, append-only outcome observations, the protocol and artifact digests. Reproduction checks membership, corrections, exclusions, denominators and metric arithmetic without private database or LLM access.

Retained artifacts are necessary to inspect exact source bytes. A dead link, unretained source or missing version limits reproduction. Identical arithmetic does not establish calibration, forecasting skill, independent timing, or successful unattended operation.