Refresh OBR figures that had been stale since the ONS anchor, and name the two levers that do not work

economy/topics/public-finances/index.html

@@ -124,7 +124,7 @@ <h2>The model view, and the same breath its limits.</h2>
       </ul>
       <p class="callout">
         <strong>Predictive validation: weak.</strong>
-        Free-running GDP and consumption MAPE are 5.75% and 9.56%; the anchored fit is by construction.
+        Free-running GDP and consumption MAPE are 4.48% and 7.49%; the anchored fit is by construction.
         <strong>Policy counterfactuals: weak.</strong>
         One income-tax costing is independently compared with HMRC; trade, labour, prices and parts of household income remain constrained.
         <a href="/obr/validation">Read the validation page →</a>

index.html

@@ -210,7 +210,7 @@ <h2>What the evidence shows—and what it doesn't.</h2>
       </p>
       <div class="verification-grid" aria-label="The verification gradient, model by model">
         <article class="verification-card"><div><strong class="mono">pe-microsim</strong><span class="verification-tag verification-tag-statute">Statutory checks</span></div><p class="verification-what">PolicyEngine&rsquo;s own tax&ndash;benefit microsimulation for the UK and US &mdash; the engine behind policyengine.org.</p><p>Rules are tested against statute; population totals inherit <strong>survey uncertainty</strong>. It cannot answer GDP, inflation, interest rates or macro feedback &mdash; that is what the models below add.</p></article>
-        <article class="verification-card"><div><strong class="mono">obr-macro</strong><span class="verification-tag">Anchored replication</span></div><p class="verification-what">Emulator of the OBR&rsquo;s UK macroeconometric forecast model.</p><p>Tracks the March 2026 EFO to <strong>0.15% GDP MAPE</strong> with published anchors; free-running error is <strong>5.75%</strong>.</p></article>
+        <article class="verification-card"><div><strong class="mono">obr-macro</strong><span class="verification-tag">Anchored replication</span></div><p class="verification-what">Emulator of the OBR&rsquo;s UK macroeconometric forecast model.</p><p>Tracks the March 2026 EFO to <strong>0.15% GDP MAPE</strong> with published anchors; free-running error is <strong>4.48%</strong>.</p></article>
         <article class="verification-card"><div><strong class="mono">boe-svar</strong><span class="verification-tag">Forecast evaluation</span></div><p class="verification-what">Bank of England structural VAR separating global and domestic shocks.</p><p>Reproduces the published decomposition for GDP (<strong>37.4% vs ~40%</strong>) and falls about <strong>8pp short on CPI</strong> (42.3% vs ~50%) once measured the way the paper defines it. Against a random walk <em>with drift</em>, and after adjusting for the 64 tests run, <strong>no variable at any horizon shows skill</strong>. <a href="/svar/validation">See the evaluation →</a></p></article>
         <article class="verification-card"><div><strong class="mono">frb-us</strong><span class="verification-tag">Software replication</span></div><p class="verification-what">The Federal Reserve&rsquo;s large-scale model of the US economy.</p><p>Matches the Fed&rsquo;s pyfrbus to within the Fed&rsquo;s own two releases&rsquo; disagreement (<strong>~1&times;10⁻⁸</strong>); no predictive claim.</p></article>
         <article class="verification-card"><div><strong class="mono">us-hank</strong><span class="verification-tag">Published replication</span></div><p class="verification-what">Heterogeneous-agent New Keynesian model of US households.</p><p>Hits every Auclert et al. (2021) calibration target; responses are <strong>first-order and stylized</strong>.</p></article>

integration/modal_app.py

@@ -129,7 +129,7 @@
 BOE_URL = "https://github.com/PolicyEngine/boe-var-model"
 FRB_URL = "https://github.com/PolicyEngine/us-frb-model"
 HANK_URL = "https://github.com/PolicyEngine/us-hank-model"
-OBR_REVISION = "85ada9be070a34a4a4a0a65d92f8f148a476ccb1"
+OBR_REVISION = "9c4b595f3b60b305459ba8c91eac5bc69fd0ffc8"
 BOE_REVISION = "42cbcbb728b55b3f416764568fa8f3aef68a6730"
 FRB_REVISION = "2508a2d3ee7d078fd82f38bc17567e4433489913"
 HANK_REVISION = "19a8e5e5cd8e7120fbbd2e3de3c5b63509824f3c"

integration/pyproject.toml

@@ -49,7 +49,7 @@ Issues = "https://github.com/PolicyEngine/macro/issues"
 # particular, frbus includes model.xml and LONGBASE.TXT inside its wheel, so a
 # clean non-editable installation is a supported access path.
 models = [
-    "obr-macro-model @ git+https://github.com/PolicyEngine/obr-macroeconomic-model@85ada9be070a34a4a4a0a65d92f8f148a476ccb1",
+    "obr-macro-model @ git+https://github.com/PolicyEngine/obr-macroeconomic-model@9c4b595f3b60b305459ba8c91eac5bc69fd0ffc8",
     "boe_var @ git+https://github.com/PolicyEngine/boe-var-model@42cbcbb728b55b3f416764568fa8f3aef68a6730",
     "frbus @ git+https://github.com/PolicyEngine/us-frb-model@2508a2d3ee7d078fd82f38bc17567e4433489913",
     "us-hank-model @ git+https://github.com/PolicyEngine/us-hank-model@19a8e5e5cd8e7120fbbd2e3de3c5b63509824f3c",

integration/src/policyengine_macro/capabilities.py

@@ -204,7 +204,7 @@ def _quality(level: str, evidence: str, next_gate: str) -> dict:
         ),
         "predictive_validation": _quality(
             "weak",
-            "Free-running GDP and consumption MAPE are 5.75% and 9.56%; the "
+            "Free-running GDP and consumption MAPE are 4.48% and 7.49%; the "
             "anchored fit is by construction.",
             "Pass rolling-origin historical-vintage tests against simple "
             "benchmarks and first-release outturns.",

integration/src/policyengine_macro/core.py

@@ -136,12 +136,33 @@ def _import_boe_var():
         "description": "Corporation tax (main) rate",
         "units": "rate change in decimal (e.g. -0.05 = 5pp cut from 25% to 20%)",
         "investment_closure": True,
+        # The investment closure's stabiliser anchors the LEVEL, not the
+        # deviation, and the deviation does not converge at any horizon: for a
+        # sustained +5pp rise |delta_IF| runs £97m at q3, £2,876m at q12 and
+        # £43,387m at q25, growing 1.21-1.27x every quarter for 25 quarters.
+        # Truncating the shock from 12 quarters to 8 barely moves the q12
+        # response, so the number is carried by accumulated drift rather than
+        # by the tax rate. The 12-quarter default is not where this settles;
+        # it is where the magnitude still looks plausible.
+        "caveat": (
+            "the response does not converge: it compounds at roughly "
+            "25%/quarter with no steady state, so read the sign and the "
+            "first few quarters, not the level or the cumulative total"
+        ),
     },
     {
         "var": "CGIPS",
         "description": "Nominal central government investment (feeds real GGI via the GGIPS/GGIDEF chain)",
         "units": "£m nominal per quarter (e.g. 3000 ≈ £2.5bn real per quarter)",
         "investment_closure": False,
+        # Measured: delta_IF is exactly 0.0 in all twelve quarters and the
+        # residual GDP effect goes negative (-0.047 by q12) against the OBR's
+        # published 1.0. The channel is not weak, it is absent.
+        "caveat": (
+            "this channel is dead under the demand closure: business "
+            "investment does not respond at all and the residual GDP effect "
+            "is wrong-signed. Do not use it to score capital spending"
+        ),
     },
 ]
 
@@ -203,7 +224,16 @@ def obr_shock(
     rows = _obr_result_rows(df)
     shocked = rows[: int(periods)]
     peak = max(rows, key=lambda r: abs(r["pct_gdp"]))
-    return {
+    # A lever whose channel is dead or non-convergent must say so in the
+    # payload, not only in a docstring nobody reads over MCP. Of the four
+    # fiscal instruments here exactly one contains behaviour: CGG is an
+    # accounting identity (multiplier exactly 1.0000, flat, against the OBR's
+    # published 0.6), CGIPS is dead, TCPRO never converges, and the
+    # household-income lever is the only one with a response worth reading.
+    caveat = next(
+        (v.get("caveat") for v in OBR_VARIABLES if v["var"] == var), None
+    )
+    result = {
         "name": name,
         "provenance": _provenance(
             model_id="obr-emulator",
@@ -222,6 +252,9 @@ def obr_shock(
         "peak_pct_gdp": peak["pct_gdp"],
         "peak_period": peak["period"],
     }
+    if caveat:
+        result["caveat"] = caveat
+    return result
 
 
 # ---------------------------------------------------------------------------

integration/tests/test_wiring.py

@@ -404,3 +404,22 @@ def test_cli_score_still_rejects_a_genuinely_unknown_model(runner):
     assert res.exit_code != 0
     assert "Traceback" not in res.output
     assert "model must be one of" in res.output
+
+
+def test_obr_shock_carries_the_lever_caveat_in_the_payload():
+    """Two of the four OBR levers do not work, and the payload must say so.
+
+    Measured in the model repo: CGIPS moves business investment by exactly
+    zero in all twelve quarters and leaves a wrong-signed residual, and TCPRO
+    compounds at roughly 25%/quarter with no steady state. A docstring does
+    not reach an MCP caller reading a JSON result, so the caveat travels with
+    the numbers.
+    """
+    from policyengine_macro import core
+
+    levers = {v["var"]: v for v in core.obr_list_variables()}
+    assert levers["TCPRO"]["caveat"], "the non-convergent lever declares no caveat"
+    assert levers["CGIPS"]["caveat"], "the dead lever declares no caveat"
+    # CGG has its own documented "~1 by construction" reading on the site and
+    # is not caveated here; asserting that keeps this test honest about scope.
+    assert "caveat" not in levers["CGG"]

obr/code/index.html

@@ -170,8 +170,8 @@ <h2>Shock a variable, read the deviation.</h2>
           <div class="dial-card">
             <div class="dial-name mono">levers — set these (exogenous)</div>
             <div class="dial-choice"><code>CGG</code><span class="dial-tradeoff"><strong>Government consumption</strong> — spending shocks, £m per quarter</span></div>
-            <div class="dial-choice"><code>CGIPS</code><span class="dial-tradeoff"><strong>Central government investment</strong> — public investment, £m per quarter (nominal)</span></div>
-            <div class="dial-choice"><code>TCPRO</code><span class="dial-tradeoff"><strong>Corporation tax rate</strong> — rate, e.g. &minus;0.05 for a 5pp cut</span></div>
+            <div class="dial-choice"><code>CGIPS</code><span class="dial-tradeoff"><strong>Central government investment</strong> — public investment, £m per quarter (nominal). <strong>Dead channel:</strong> business investment does not respond at all and the residual GDP effect is wrong-signed, against the OBR&rsquo;s published 1.0. Do not score capital spending with it</span></div>
+            <div class="dial-choice"><code>TCPRO</code><span class="dial-tradeoff"><strong>Corporation tax rate</strong> — rate, e.g. &minus;0.05 for a 5pp cut. <strong>Does not converge:</strong> the response compounds at roughly 25% a quarter with no steady state, so read the sign and the first few quarters, never the level or the cumulative total</span></div>
             <div class="dial-choice"><code>HHDI_ADDFACTOR</code><span class="dial-tradeoff"><strong>Externally costed household reform</strong> — virtual instrument, quarterly £m; positive revenue lowers HHDI via a held add-factor</span></div>
             <div class="dial-choice"><code>R / GILT / RX</code><span class="dial-tradeoff"><strong>Bank Rate &middot; gilt yields &middot; exchange rate</strong> — monetary &amp; market assumptions</span></div>
           </div>
@@ -218,8 +218,8 @@ <h2>Shock a variable, read the deviation.</h2>
           <div class="dial-card">
             <div class="dial-name mono">levers — set these (exogenous)</div>
             <div class="dial-choice"><code>CGG</code><span class="dial-tradeoff"><strong>Government consumption</strong> — spending shocks, £m per quarter</span></div>
-            <div class="dial-choice"><code>CGIPS</code><span class="dial-tradeoff"><strong>Central government investment</strong> — public investment, £m per quarter (nominal)</span></div>
-            <div class="dial-choice"><code>TCPRO</code><span class="dial-tradeoff"><strong>Corporation tax rate</strong> — rate, e.g. &minus;0.05 for a 5pp cut</span></div>
+            <div class="dial-choice"><code>CGIPS</code><span class="dial-tradeoff"><strong>Central government investment</strong> — public investment, £m per quarter (nominal). <strong>Dead channel:</strong> business investment does not respond at all and the residual GDP effect is wrong-signed, against the OBR&rsquo;s published 1.0. Do not score capital spending with it</span></div>
+            <div class="dial-choice"><code>TCPRO</code><span class="dial-tradeoff"><strong>Corporation tax rate</strong> — rate, e.g. &minus;0.05 for a 5pp cut. <strong>Does not converge:</strong> the response compounds at roughly 25% a quarter with no steady state, so read the sign and the first few quarters, never the level or the cumulative total</span></div>
             <div class="dial-choice"><code>HHDI_ADDFACTOR</code><span class="dial-tradeoff"><strong>Externally costed household reform</strong> — virtual instrument, quarterly £m; positive revenue lowers HHDI via a held add-factor</span></div>
             <div class="dial-choice"><code>R / GILT / RX</code><span class="dial-tradeoff"><strong>Bank Rate &middot; gilt yields &middot; exchange rate</strong> — monetary &amp; market assumptions</span></div>
           </div>

obr/index.html

@@ -82,7 +82,7 @@ <h2>Microsimulation supplies the direct cost; this model returns the feedback.</
       <div class="qa-facts">
         <div class="qa-fact"><span class="icon">&rarr;</span><span><strong>Connection.</strong> Static-costing bridge, both hosted: <a href="/pe">PolicyEngine</a> computes the direct Exchequer effect of a statutory reform; <code>score_reform</code> injects it through the declared <code>HHDI_ADDFACTOR</code> point and this model returns the economy-wide second round.</span></div>
         <div class="qa-fact"><span class="icon">&rarr;</span><span><strong>Worked example.</strong> 1p on the basic rate from April 2026: direct yield &pound;6.46bn in 2026; GDP &minus;0.020% on impact, deepening to &minus;0.058% by 2027Q4 (full walkthrough &rarr; <a href="/models#score">score a reform</a>).</span></div>
-        <div class="qa-fact"><span class="icon">&rarr;</span><span><strong>Validation.</strong> <a href="/obr/validation">Anchored to the March 2026 EFO</a> (0.15% GDP MAPE) &mdash; but anchored accuracy is a <strong>by-construction invariant</strong>, not a forecast claim: free-running, the same equations miss GDP by <strong>5.75%</strong>. Checked against HMRC&rsquo;s ready reckoner (&pound;6.46bn vs &pound;6.9bn).</span></div>
+        <div class="qa-fact"><span class="icon">&rarr;</span><span><strong>Validation.</strong> <a href="/obr/validation">Anchored to the March 2026 EFO</a> (0.15% GDP MAPE) &mdash; but anchored accuracy is a <strong>by-construction invariant</strong>, not a forecast claim: free-running, the same equations miss GDP by <strong>4.48%</strong>. Checked against HMRC&rsquo;s ready reckoner (&pound;6.46bn vs &pound;6.9bn).</span></div>
         <div class="qa-fact"><span class="icon">&rarr;</span><span><strong>Results.</strong> Run it from <a href="/connect">any connected client</a>; every score carries its assumptions and data vintage.</span></div>
       </div>
     </div>