vahid-ahmadi · Nov 6, 2026 · Nov 6, 2026 · Nov 6, 2026 · Nov 6, 2026
diff --git a/integration/src/policyengine_macro/capabilities.py b/integration/src/policyengine_macro/capabilities.py
@@ -255,6 +255,8 @@ def _quality(level: str, evidence: str, next_gate: str) -> dict:
             "real data to numerical precision.",
             "Keep all exact invariants hard-gated for every specification.",
         ),
+        # Figures in this entry and the next are from boe-var-model 845e2a4
+        # (data through 2026Q1), the run vintage cited on /svar/validation.
         "predictive_validation": _quality(
             "weak",
             "No demonstrated skill against a random walk WITH DRIFT on any "

diff --git a/papers/boe-svar/figures/make_comparison.py b/papers/boe-svar/figures/make_comparison.py
@@ -101,7 +101,8 @@
 # Fast fixed-seed configuration, identical to tests/conftest.py.
 N_DRAWS, SAMPLE_SEED, IDENT_SEED, LAGS = 600, 1, 2, 4
 
-# Production artifact (results/summary.md): 10,000 draws, importance-weighted,
+# Production artifact (results/summary.md at boe-var-model 845e2a4, data through
+# 2026Q1 -- the run vintage the site and paper cite): 10,000 draws, importance-weighted,
 # sample 1992Q1-2025Q1 (2026Q1 vintage). Posterior MEAN of the per-draw group
 # share at 4 quarters — the paper-comparable statistic.
 PROD = {"gdp": 37.4, "cpi": 42.3}

diff --git a/papers/boe-svar/index.html b/papers/boe-svar/index.html
@@ -83,7 +83,7 @@ <h2 class="paper-title">An open replication of the Bank of England structural VA
 
     <p class="chip-row"><span class="chip chip-brand">Working paper</span><span class="chip">No DOI</span><span class="chip">Vahid Ahmadi &middot; PolicyEngine</span></p>
     <p class="paper-result" style="max-width:78ch;">The paper&rsquo;s claim in one line: an open replication that matches the Bank&rsquo;s GDP variance decomposition, falls short on CPI, and shows no forecast skill over a random walk with drift once the tests are adjusted together.</p>
-<details class="about"><summary>More detail</summary><p>Current forecast: data through <strong>2026Q1</strong>, outlook <strong>2026Q2&ndash;2029Q2</strong>. Every figure is on the <a href="/svar/validation">validation tab</a>.</p></details>
+<details class="about"><summary>More detail</summary><p>Current forecast: data through <strong>2026Q1</strong>, outlook <strong>2026Q2&ndash;2029Q2</strong>. Every figure is on the <a href="/svar/validation">validation tab</a>. The replication and evaluation figures come from the model repository as committed at <a href="https://github.com/PolicyEngine/boe-var-model/tree/845e2a4">845e2a4</a> (data through 2026Q1).</p></details>
     <details class="expander">
       <summary>What the paper covers</summary>
       <ol>

diff --git a/papers/boe-svar/main.pdf b/papers/boe-svar/main.pdf
diff --git a/papers/boe-svar/sections/implementation.tex b/papers/boe-svar/sections/implementation.tex
@@ -16,7 +16,7 @@ \subsection{Code architecture}
 
 \subsection{Estimation and identification runs}
 
-The production configuration takes \textbf{10{,}000} posterior draws, of which \textbf{744} survive the sign-restriction accept/reject step (a 7.4 per cent acceptance rate), with $p=4$ lags and hyperparameters $\lambda = 0.2$, $\mu = 1$. Zero restrictions are satisfied by construction on every draw; sign restrictions are enforced by rejection, so they too hold exactly on every accepted draw --- neither fact is a test (Section~\ref{sec:validation}). Because the \citet{arias2018} zero-restriction sampler does not draw $Q$ uniformly from the restricted set, the slow validation path additionally computes the algorithm's importance weights (each requiring a numerical Jacobian per accepted draw); the importance-weighted effective sample size is \textbf{356.1}, and weighted and unweighted results are compared in Section~\ref{sec:validation}. The forecast-revision exercise re-estimates through 2024Q2 with 6{,}000 draws (487 accepted, weighted ESS 207.5). The hosted interactive tools default to a smaller 2{,}000-draw run on the extended sample; the 10{,}000-draw run described here is the publication artifact.
+The production configuration takes \textbf{10{,}000} posterior draws, of which \textbf{744} survive the sign-restriction accept/reject step (a 7.4 per cent acceptance rate), with $p=4$ lags and hyperparameters $\lambda = 0.2$, $\mu = 1$. Zero restrictions are satisfied by construction on every draw; sign restrictions are enforced by rejection, so they too hold exactly on every accepted draw --- neither fact is a test (Section~\ref{sec:validation}). Because the \citet{arias2018} zero-restriction sampler does not draw $Q$ uniformly from the restricted set, the slow validation path additionally computes the algorithm's importance weights (each requiring a numerical Jacobian per accepted draw); the importance-weighted effective sample size is \textbf{356.1}, and weighted and unweighted results are compared in Section~\ref{sec:validation}. The forecast-revision exercise re-estimates through 2024Q2 with 6{,}000 draws (487 accepted, weighted ESS 207.5). The hosted interactive tools default to a smaller 2{,}000-draw run on the extended sample; the 10{,}000-draw run described here is the publication artifact. Its figures, and those of the rolling-origin evaluation in Section~\ref{sec:validation}, are the ones committed in the model repository at revision \texttt{845e2a4} (22 September 2026; data through 2026Q1, estimation sample 1992Q1--2025Q1).
 
 \subsection{Hosted MCP tools}
 

diff --git a/svar/methodology/index.html b/svar/methodology/index.html
@@ -230,7 +230,7 @@ <h4>Corrections</h4>
 
       <div class="meth-step" id="meth-4" data-step="3">
         <div class="meth-step-head"><span class="meth-step-num mono">4</span><h3>Sampling: draws and acceptance</h3></div>
-        <p>This is accept/reject sampling. The production run drew <strong>10,000</strong> posterior <mjx-container data-tex="(\Pi, \Sigma)" data-mjx="df92f8bdd622c0fb" class="MathJax" jax="CHTML" style="font-size: 123.6%; position: relative;"><mjx-math class=" MJX-TEX" aria-hidden="true"><mjx-mo class="mjx-n"><mjx-c class="mjx-c28"></mjx-c></mjx-mo><mjx-mi class="mjx-n"><mjx-c class="mjx-c3A0"></mjx-c></mjx-mi><mjx-mo class="mjx-n"><mjx-c class="mjx-c2C"></mjx-c></mjx-mo><mjx-mi class="mjx-n" space="2"><mjx-c class="mjx-c3A3"></mjx-c></mjx-mi><mjx-mo class="mjx-n"><mjx-c class="mjx-c29"></mjx-c></mjx-mo></mjx-math><mjx-assistive-mml unselectable="on" display="inline"><math xmlns="http://www.w3.org/1998/Math/MathML"><mo stretchy="false">(</mo><mi mathvariant="normal">Π</mi><mo>,</mo><mi mathvariant="normal">Σ</mi><mo stretchy="false">)</mo></math></mjx-assistive-mml></mjx-container> draws, of which <strong>744 passed</strong> the sign restrictions (7.4%), for an importance-weighted <a class="gloss-link" href="/glossary#ess">effective sample size</a> of <strong>356.1</strong> (<a href="https://github.com/PolicyEngine/boe-var-model/blob/main/results/summary.md">results/summary.md</a>, seed 20260812) — so medians and bands are correct but noisier than the paper's.</p>
+        <p>This is accept/reject sampling. The production run drew <strong>10,000</strong> posterior <mjx-container data-tex="(\Pi, \Sigma)" data-mjx="df92f8bdd622c0fb" class="MathJax" jax="CHTML" style="font-size: 123.6%; position: relative;"><mjx-math class=" MJX-TEX" aria-hidden="true"><mjx-mo class="mjx-n"><mjx-c class="mjx-c28"></mjx-c></mjx-mo><mjx-mi class="mjx-n"><mjx-c class="mjx-c3A0"></mjx-c></mjx-mi><mjx-mo class="mjx-n"><mjx-c class="mjx-c2C"></mjx-c></mjx-mo><mjx-mi class="mjx-n" space="2"><mjx-c class="mjx-c3A3"></mjx-c></mjx-mi><mjx-mo class="mjx-n"><mjx-c class="mjx-c29"></mjx-c></mjx-mo></mjx-math><mjx-assistive-mml unselectable="on" display="inline"><math xmlns="http://www.w3.org/1998/Math/MathML"><mo stretchy="false">(</mo><mi mathvariant="normal">Π</mi><mo>,</mo><mi mathvariant="normal">Σ</mi><mo stretchy="false">)</mo></math></mjx-assistive-mml></mjx-container> draws, of which <strong>744 passed</strong> the sign restrictions (7.4%), for an importance-weighted <a class="gloss-link" href="/glossary#ess">effective sample size</a> of <strong>356.1</strong> (<a href="https://github.com/PolicyEngine/boe-var-model/blob/845e2a4/results/summary.md">results/summary.md</a> at commit 845e2a4, data through 2026Q1, seed 20260812) — so medians and bands are correct but noisier than the paper's.</p>
         <div class="meth-inline">
         <div class="meth-sec-title">Accept/reject sampling</div>
         <div class="meth-eq"><div class="meth-eq-label">Production run</div><div class="meth-eq-body">10,000 draws &rarr; 744 accepted (7.4%)</div></div>

diff --git a/svar/validation/index.html b/svar/validation/index.html
@@ -162,6 +162,7 @@ <h2>Where this departs from the Bank&rsquo;s model</h2>
     <div class="band-head" data-icon="chart">
       <h2>Replication, one frozen origin, 49 origins, the benchmark</h2>
     </div>
+    <p class="section-intro">Run vintage: every replication and evaluation figure on this tab comes from the model repository as committed at <a href="https://github.com/PolicyEngine/boe-var-model/tree/845e2a4">845e2a4</a> (22 Sep 2026), with data through 2026Q1 and the estimation sample 1992Q1&ndash;2025Q1. A later data refresh would move them.</p>
     <details class="expander dim-fold">
       <summary><span><strong>Evidence by dimension:</strong> implementation fidelity strong; identification, uncertainty and vintage reproducibility moderate; predictive validation weak; policy counterfactuals not applicable</span></summary>
     <!-- Levels and next gates: integration/src/policyengine_macro/capabilities.py MODEL_QUALITY["boe-svar"]. -->
@@ -458,7 +459,7 @@ <h2>Replication, one frozen origin, 49 origins, the benchmark</h2>
 </ul>
 <p class="vc-notes">Vertical bars span the best and worst of the eight variables; lowest cell 29% (CPI energy, 68% band, h4).</p>
 </div><!-- /vc-fig -->
-                <figcaption>Empirical interval coverage across 49 expanding-window origins, mean across the eight variables, by horizon. The 68% band <em>over</em>-covers at one quarter (71%) and falls below nominal from two onward; the 90% band <strong>under-covers</strong> at every horizon. Both worsen with horizon — the opposite of the frozen-origin reading above, and the one to weight. These figures are the evaluation run without Covid dummies. The same evaluation with the six Covid dummies (2020Q1&ndash;2021Q2) that the published fans use covers about the same: 69% at one quarter and 56% at eight (<a href="https://github.com/PolicyEngine/boe-var-model/blob/main/results/coverage_evaluation.json">boe-var-model <code>production_spec</code></a>). Source: <code>papers/boe-svar/figures/coverage_evaluation.json</code>.</figcaption>
+                <figcaption>Empirical interval coverage across 49 expanding-window origins, mean across the eight variables, by horizon. The 68% band <em>over</em>-covers at one quarter (71%) and falls below nominal from two onward; the 90% band <strong>under-covers</strong> at every horizon. Both worsen with horizon — the opposite of the frozen-origin reading above, and the one to weight. These figures are the evaluation run without Covid dummies. The same evaluation with the six Covid dummies (2020Q1&ndash;2021Q2) that the published fans use covers about the same: 69% at one quarter and 56% at eight (<a href="https://github.com/PolicyEngine/boe-var-model/blob/845e2a4/results/coverage_evaluation.json">boe-var-model <code>production_spec</code></a>). Source: <code>papers/boe-svar/figures/coverage_evaluation.json</code>.</figcaption>
               </figure>
         <p class="kv-title">49 expanding-window origins</p>
         <dl class="kv kv-text"><div><dt>Run</dt><dd>evaluation model without Covid dummies</dd></div><div><dt>68% band, h=1</dt><dd>71%, just above nominal</dd></div><div><dt>68% band, h=8</dt><dd>57%</dd></div><div><dt>90% band</dt><dd>under-covers at every horizon</dd></div><div><dt>Unemployment satellite</dt><dd>beats no-change at h=1&ndash;4 only, not significant</dd></div></dl>
@@ -479,7 +480,7 @@ <h2>Replication, one frozen origin, 49 origins, the benchmark</h2>
                 not significant), so the published path is capped there. Its
                 specification is on the <a href="/svar/methodology#meth-5">methodology
                 tab</a>; skill table:
-                <a href="https://github.com/PolicyEngine/boe-var-model/blob/main/results/unemployment_satellite_validation.json">unemployment_satellite_validation.json</a>.
+                <a href="https://github.com/PolicyEngine/boe-var-model/blob/845e2a4/results/unemployment_satellite_validation.json">unemployment_satellite_validation.json</a>.
               </p>
         </details>
               </details>