Paper contents
I · Research question
Three clocks must be separated before two venues can be compared.
A source-event clock describes when an exchange says a funding event occurred. A source-record clock describes when a record or snapshot was created. A client-receipt clock describes when a particular consumer received the response. A first-frame duration, a historical created_at field and a settlement timestamp cannot substitute for one another.
The paper asks what public transport documentation and public funding histories can establish about data freshness, timestamp semantics and clock-offset uncertainty. It does not estimate physical clock skew inside an exchange, and it does not convert a single-host connection observation into a latency ranking.
Research 01 and 02 measure and frame net carry. This paper audits the upstream observation layer on which those calculations depend.
Read Research 01 Read Research 02
II · Data & methods
One transport map. One event-level estimand. One explicit missing-data bound.
Transport capability
The 2026-10-09 audit maps funding, price, order-book and all-market capabilities at 22 venues. Price semantics and RFQ or oracle-based designs remain distinct.
Venue-hour offsets
For each venue and nominal UTC hour, the primary offset is the median across BTC, ETH and SOL. This prevents three assets sharing one venue timestamp from becoming three independent clock draws.
Chronological diagnostic
The calibration-period empirical 99th percentile plus 1 ms is frozen before scoring later observations. There is one fixed rule per venue and no holdout tuning.
Partial identification
Missing clocks are not filled with zero and are not allowed to improve the denominator. The sharp interval remains agnostic about missing outcomes.
These are worst-case identification bounds, not a confidence interval. When A = 0, the estimand is not applicable.
III · Figures
Capabilities, offsets, time variation and missingness.
Scroll horizontally or use the arrows. Select a figure to inspect its details.
IV · Evaluation & robustness
Later observations, not a favorable subsample, determine whether the tolerance travels.
Ten ISO weeks plus the calibration/holdout split show level shifts. Hyperliquid remains within tens of milliseconds; Pacifica and Extended change regime.
Asset-specific estimates are nearly identical, confirming venue-level timestamp mechanisms. Record, event and equal-day summaries answer different questions and are all reported.
Three-, seven- and 14-day moving blocks are compared. Pacifica's interval is materially sensitive to regime persistence; the headline seven-day choice is not hidden.
Only explicit edgeX settlement rows enter the settlement sample. Paradex snapshots remain outside settlement-offset estimates. Oracle prices and RFQ tiers are not relabeled as mark prices or order-book levels.
No forecast or trading-performance claim is imported.
The study has no predicted funding rate, fitted return model or model search. Forecast-loss tests, backtest-overfitting adjustments and net-of-cost return calculations are therefore inapplicable rather than omitted.
V · Evidence boundary
What this paper does not claim.
- No multi-location latency estimate. Transport first-frame observations are single-host and single-day. They do not support a venue latency ranking.
- No large-sample Aster or ApeX offset result. Connection resets prevented refresh. Three-row pilot observations are not generalized or replaced with indirect data.
- No physical clock-skew estimate. A timestamp offset can include scheduling, computation and record writing. edgeX's zero applies only to the analyzed settlement field.
- No guaranteed future tolerance. One chronological split and one fixed rule cannot certify future exchange behavior. The Extended anomaly remains unexplained and retained.
- No empirical route-level missingness estimate. Partial-identification figures use synthetic counts only. No production pairs, thresholds, internal fields or product rules appear.
This paper is for research and methodological information only. It does not provide investment, legal, accounting or tax advice; does not recommend any venue, trade or leverage; and does not promise data availability, execution, income or protection from loss.
VI · Conclusion
Report the clock, the anomaly and the unidentified remainder.
The results support a modest measurement practice: document transport capabilities at field level, estimate timestamp offsets by venue, test temporal transfer, and expose missing-clock uncertainty. A universal staleness second count would conceal more than it reveals.
The release provides response-level provenance, normalized data, analysis code, machine-readable tables and vector figures so that each principal number can be recomputed offline.