% ============================================================================
\begin{appendices}
% ============================================================================

\section{Pre-registered design, review, and changes}
\label{app:design}

The analysis plan (research question, data, construction, hypotheses H1--H5, inference, outputs and the re-scoping rule) was written
and frozen in the project repository (tag \texttt{design-frozen}) before any intraday result was computed. The daily factors and the
out-of-sample fidelity of the mimicking portfolios (H1) were, by construction, the first results inspected. A referee-style review of
the frozen design was then obtained before the event-study, periodicity, covariation and forecasting results were examined. The review
and the point-by-point response are archived with the code. The changes made to the plan are the following.

\begin{enumerate}
  \item \emph{Short-end coverage (after H1, prompted by the review).} The CMT target with eleven maturities remains primary, as planned.
  Because the slope and curvature fidelity fell short of the H1 thresholds, we added (i) the cash-yield ceiling---the same rolling
  regressions with the 2-, 5-, 10- and 30-year cash yield changes as regressors---to separate the bill segment from the futures basis;
  (ii) two alternative targets, the GSW zero curve (1--30 years) and CMT factors from maturities of one year and above; (iii) the
  front three-month SOFR future as a seventh instrument from 2022. A fourth target, CMT factors from maturities of two years and
  above, was tested and discarded because with $\lambda$ fixed the slope and curvature are poorly identified from seven points
  (daily-change standard deviations of 15.5 and 28 basis points and a correlation of $-0.75$).
  \item \emph{Minute-level validation (review).} The plan's fidelity test compares daily objects. Two minute-level checks were added:
  an independent construction of the factors from Treasury ETFs by duration inversion and loading regression, with no daily step,
  and a leave-one-contract-out spanning test of the futures returns at one-, five- and thirty-minute sampling.
  \item \emph{Sampling frequency (review).} Five-minute sampling is used for the main covariation and forecasting results; one-minute
  measures are reported as a signature. The tick-size noise benchmark for the level--slope correlation was added.
  \item \emph{Event windows (review).} Placebo windows at the same clock time on matched non-announcement days replace the uniform
  30/1380 benchmark as the main comparison; the benchmark is still reported. Windows of $[-5,+15]$ and $[-5,+90]$ minutes were added for
  FOMC statements, the second to cover the press conference. Confidence intervals resample events rather than days.
  \item \emph{Multiple testing (review).} The nine Diebold--Mariano comparisons of HAR against GARCH are reported with a Holm adjustment.
  \item \emph{Coverage facts.} The plan stated five contracts for 2008--2015; the correct count is four in 2008--2009 (the ultra bond
  starts in 2010) and five from 2010 to 2015. The plan's forecasting section said weights were re-estimated monthly; the weights are
  re-estimated quarterly throughout and the forecasting models monthly.
  \item \emph{Contract inclusion rule (after the first complete draft).} The first draft kept a contract in a quarter's regression if
  it had at least 300 minute returns on 95 percent of the window's days. Applied to the vendor's continuous series, whose bars are
  sparse around rolls, the rule excluded the 2-year note, the ultra 10-year and the ultra bond and left three contracts. The rule
  was relaxed to 100 minutes on 85 percent of days and all results were recomputed.
  \item \emph{Return panel rebuilt from individual contracts (after three referee-style reviews of the first draft).} The reviews
  showed that the vendor's continuous series dropped a third of the 08:30 releases in roll months since 2018 and that the cumulated
  minute series was not the object validated by H1. The panel was rebuilt from the bars of every individual contract, using the
  most-traded contract of each session (Section~\ref{sec:data-futures}); the fidelity of the cumulated minute series itself and of
  the implied yields was added (Section~\ref{sec:results-fidelity}); the short rate was dropped from the headline results because the
  futures do not identify it; the local projections, the five-minute VAR and the volatility connectedness analysis were added;
  and every claim flagged by the reviews as unsupported or inconsistent with the metric files was either computed and documented
  (boundary shift, expanding window, five-minute denominators, release-time check, SOFR liquidity, bounce filter counts, in-sample
  weights) or removed. H5 is now reported separately under QLIKE and MSE.
\end{enumerate}

Not implemented from the reviews: a regression of FOMC-window factor changes on a published monetary-policy surprise series
(the mapping to the target and path factors in Section~\ref{sec:results} is qualitative); Giacomini--White tests and model
confidence sets; noise-robust covariance estimators; historical tick-size changes in the noise benchmark. The design review,
the three referee reports and the point-by-point responses are archived with the code.

\section{Data availability and reproducibility}
\label{app:data}

The one-minute futures and ETF bars are licensed from FirstRate Data for private use and cannot be redistributed; all bar-level
computations run on a single machine and only aggregates (daily factor changes, realized measures, event-window sums, forecast
losses, figure data) leave it. The public inputs---FRED H.15 constant-maturity yields and the effective federal funds rate, the
Federal Reserve Board's GSW parameter file, the FOMC calendars and the ALFRED release-date lists---were downloaded on 9 November
2026 and are stored with the code. The code (Python 3.12; numpy, pandas, polars, scipy, statsmodels, arch) regenerates every
table, figure and number in this paper from the metric files, and the pre-registered design, the design review, the three
referee reports and the responses are archived in the same repository. A researcher with access to the same vendor data can
rerun the pipeline end to end in under fifteen minutes on an Apple Silicon workstation.

\section{Additional tables}
\label{app:tables}

\input{tables/tab_weights}
\input{tables/tab_spanning}
\input{tables/tab_etf}

\end{appendices}