diff --git a/.github/workflows/figures.yml b/.github/workflows/figures.yml
new file mode 100644
index 0000000..d72d0f1
--- /dev/null
+++ b/.github/workflows/figures.yml
@@ -0,0 +1,92 @@
+name: Figures
+
+# The front page states four numbers under "Numbers with the conditions they
+# were measured under". Three were typed by hand, and by September 2026 two had
+# gone stale: decisionrl had grown to 32 algorithms across 24 environments while
+# the page said 31 and 22, and coverage read 86% where CI measured 84%. The one
+# organisation-wide rule this site advertises is that a published number is
+# measured on the run that describes it; its own front page was the exception.
+#
+# This reads each figure's source and reports drift into one standing issue.
+# It is a separate workflow rather than a step in CI because it needs the
+# network, and a required check that depends on someone else's uptime teaches
+# people to re-run it rather than to read it.
+on:
+  schedule:
+    - cron: "0 7 * * 1"
+  workflow_dispatch:
+
+permissions:
+  contents: read
+  issues: write
+
+jobs:
+  figures:
+    name: Check the stated figures against their sources
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v7
+
+      - uses: actions/setup-python@v7
+        with:
+          python-version: "3.12"
+
+      - name: Compare
+        id: figures
+        run: |
+          python scripts/check_figures.py . | tee figures.txt
+          echo "drifted=${PIPESTATUS[0]}" >> "$GITHUB_OUTPUT"
+        continue-on-error: true
+
+      - name: Build the report
+        if: steps.figures.outputs.drifted != '0'
+        run: |
+          {
+            echo "A figure on the front page no longer matches the repository it describes."
+            echo
+            echo '```'
+            cat figures.txt
+            echo '```'
+            echo
+            echo "Each line names the page, what it says, and what the source says."
+            echo "Reproduce with \`python scripts/check_figures.py .\`."
+          } > report.md
+
+      - name: Find the standing figures issue
+        id: standing
+        if: steps.figures.outputs.drifted != '0'
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          number=$(gh issue list \
+            --repo "$GITHUB_REPOSITORY" \
+            --state open \
+            --label maintenance \
+            --search 'in:title "A figure on the site has drifted"' \
+            --limit 1 --json number --jq '.[0].number // empty')
+          echo "number=${number}" >> "$GITHUB_OUTPUT"
+
+      - name: Report into the standing issue
+        if: steps.figures.outputs.drifted != '0'
+        uses: peter-evans/create-issue-from-file@v6
+        with:
+          issue-number: ${{ steps.standing.outputs.number }}
+          title: A figure on the site has drifted
+          content-filepath: ./report.md
+          labels: maintenance
+
+      - name: Close the standing issue once every figure matches
+        if: steps.figures.outputs.drifted == '0'
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          number=$(gh issue list \
+            --repo "$GITHUB_REPOSITORY" \
+            --state open \
+            --label maintenance \
+            --search 'in:title "A figure on the site has drifted"' \
+            --limit 1 --json number --jq '.[0].number // empty')
+          if [ -n "$number" ]; then
+            gh issue close "$number" --repo "$GITHUB_REPOSITORY" \
+              --comment "Every checkable figure matches its source again, as of \`${{ github.sha }}\`."
+          fi
diff --git a/index.html b/index.html
index 3bf2165..af7d407 100644
--- a/index.html
+++ b/index.html
@@ -438,7 +438,7 @@ <h2>Numbers with the conditions they were measured under</h2>
       <dl class="figures">
         <div class="figure">
           <dt>Algorithms</dt>
-          <dd>31<span class="sub">In decisionrl, across 22 environments — nine of them applied rather than synthetic.</span></dd>
+          <dd>32<span class="sub">In decisionrl, across 24 environments — nine of them applied rather than synthetic.</span></dd>
         </div>
         <div class="figure">
           <dt>Required dependencies</dt>
@@ -446,7 +446,7 @@ <h2>Numbers with the conditions they were measured under</h2>
         </div>
         <div class="figure">
           <dt>Line coverage</dt>
-          <dd>86%<span class="sub">decisionrl, measured by pytest-cov in CI. The build fails below the threshold rather than reporting it.</span></dd>
+          <dd>84%<span class="sub">decisionrl, measured by pytest-cov on the fast suite in CI. The build fails below 80% rather than reporting the number.</span></dd>
         </div>
         <div class="figure">
           <dt>Retrieval recall@5</dt>
diff --git a/ru/index.html b/ru/index.html
index a4cd84b..6e1169d 100644
--- a/ru/index.html
+++ b/ru/index.html
@@ -436,7 +436,7 @@ <h2>Цифры с условиями, при которых они получе
       <dl class="figures">
         <div class="figure">
           <dt>Алгоритмов</dt>
-          <dd>31<span class="sub">В decisionrl, на 22 средах — девять из них прикладные, а не синтетические.</span></dd>
+          <dd>32<span class="sub">В decisionrl, на 24 средах — девять из них прикладные, а не синтетические.</span></dd>
         </div>
         <div class="figure">
           <dt>Обязательных зависимостей</dt>
@@ -444,7 +444,7 @@ <h2>Цифры с условиями, при которых они получе
         </div>
         <div class="figure">
           <dt>Покрытие строк</dt>
-          <dd>86%<span class="sub">decisionrl, замер pytest-cov в CI. Ниже порога сборка падает, а не сообщает цифру.</span></dd>
+          <dd>84%<span class="sub">decisionrl, замер pytest-cov по быстрому набору в CI. Ниже 80% сборка падает, а не сообщает цифру.</span></dd>
         </div>
         <div class="figure">
           <dt>Поиск, recall@5</dt>
diff --git a/scripts/check_figures.py b/scripts/check_figures.py
new file mode 100644
index 0000000..7a4b747
--- /dev/null
+++ b/scripts/check_figures.py
@@ -0,0 +1,162 @@
+#!/usr/bin/env python3
+"""Hold the figures on the front page to the repositories they describe.
+
+The page states four numbers under "Numbers with the conditions they were
+measured under". Three of them were typed by hand and nothing recomputed them,
+and by September 2026 two had gone stale: decisionrl had grown to 32 algorithms
+across 24 environments while the page still said 31 and 22, and line coverage
+read 86% when CI measured 84%. An organisation whose stated rule is that a
+published number is measured on the run that describes it cannot have its own
+front page be the exception.
+
+Each figure here names its source and is checked against it:
+
+  algorithms / environments  decisionrl's CITATION.cff, which decisionrl's own
+                             test suite pins to the package it ships
+  required dependencies      glia's pyproject.toml, read directly
+  line coverage              Codecov, the run that produced it
+
+recall@5 is not checked yet: praxis states it in prose, with nothing pinning it
+to an eval run. That is a gap in praxis, not something to paper over here, so
+the figure is reported as unverified rather than quietly passed.
+
+Standard library only. Needs network, which is why this runs in its own weekly
+workflow rather than in the required CI check.
+
+Usage:
+    python scripts/check_figures.py [site_root]
+"""
+
+from __future__ import annotations
+
+import json
+import re
+import sys
+import urllib.error
+import urllib.request
+from pathlib import Path
+
+RAW = "https://raw.githubusercontent.com/DrobyshevDev"
+CODECOV = "https://api.codecov.io/api/v2/github/DrobyshevDev/repos/decisionrl/"
+
+PAGES = ("index.html", "ru/index.html")
+
+
+def fetch(url: str) -> str | None:
+    try:
+        request = urllib.request.Request(url, headers={"User-Agent": "drobyshevdev-figures"})
+        with urllib.request.urlopen(request, timeout=30) as response:
+            return response.read().decode("utf-8")
+    except (urllib.error.URLError, urllib.error.HTTPError, TimeoutError):
+        return None
+
+
+def figures_on(page: str) -> dict[str, str]:
+    """The <dt>/<dd> pairs of the figures list, as {label: value}."""
+    block = re.search(r'<dl class="figures">(.*?)</dl>', page, re.DOTALL)
+    if not block:
+        return {}
+    pairs = re.findall(r"<dt>(.*?)</dt>\s*<dd>(.*?)<span", block.group(1), re.DOTALL)
+    return {label.strip(): value.strip() for label, value in pairs}
+
+
+def source_counts() -> tuple[int, int, int] | None:
+    """Algorithms, environments and applied environments, from decisionrl."""
+    citation = fetch(f"{RAW}/decisionrl/main/CITATION.cff")
+    if citation is None:
+        return None
+    match = re.search(r"(\d+) algorithms and (\d+) environments, (\d+) of them applied", citation)
+    return tuple(int(g) for g in match.groups()) if match else None
+
+
+def source_glia_dependencies() -> int | None:
+    pyproject = fetch(f"{RAW}/glia/master/pyproject.toml")
+    if pyproject is None:
+        return None
+    match = re.search(r"^dependencies\s*=\s*\[(.*?)\]", pyproject, re.DOTALL | re.MULTILINE)
+    if not match:
+        return None
+    return len([item for item in match.group(1).split(",") if item.strip()])
+
+
+def source_coverage() -> float | None:
+    body = fetch(CODECOV)
+    if body is None:
+        return None
+    try:
+        totals = json.loads(body).get("totals") or {}
+    except json.JSONDecodeError:
+        return None
+    coverage = totals.get("coverage")
+    return float(coverage) if coverage is not None else None
+
+
+def main(root: Path) -> int:
+    pages = {}
+    for name in PAGES:
+        path = root / name
+        if not path.exists():
+            print(f"  {name}: not found", file=sys.stderr)
+            return 2
+        pages[name] = figures_on(path.read_text(encoding="utf-8"))
+
+    problems: list[str] = []
+    unchecked: list[str] = []
+
+    counts = source_counts()
+    if counts is None:
+        unchecked.append("algorithms/environments — decisionrl's CITATION.cff could not be read")
+    else:
+        algorithms, environments, applied = counts
+        for name, figures in pages.items():
+            stated = next((v for k, v in figures.items() if k in ("Algorithms", "Алгоритмов")), None)
+            if stated is None:
+                problems.append(f"{name}: no algorithms figure")
+            elif stated != str(algorithms):
+                problems.append(f"{name}: algorithms says {stated}, decisionrl ships {algorithms}")
+        words = {9: ("nine", "девять")}
+        for name, page in ((n, (root / n).read_text(encoding="utf-8")) for n in PAGES):
+            for quoted in re.findall(r"(?:across|на) (\d+) (?:environments|средах)", page):
+                if int(quoted) != environments:
+                    problems.append(f"{name}: says {quoted} environments, decisionrl ships {environments}")
+            # The applied count is spelled out in words beside it, which is how
+            # it escaped every check that looked for digits.
+            spelled = words.get(applied)
+            if spelled and not any(word in page for word in spelled):
+                problems.append(
+                    f"{name}: does not say {spelled[0]}/{spelled[1]} applied environments, "
+                    f"but decisionrl ships {applied}"
+                )
+
+    dependencies = source_glia_dependencies()
+    if dependencies is None:
+        unchecked.append("required dependencies — glia's pyproject.toml could not be read")
+    elif dependencies != 0:
+        problems.append(f"glia now has {dependencies} required dependencies; the page claims none")
+
+    coverage = source_coverage()
+    if coverage is None:
+        unchecked.append("line coverage — Codecov has no figure for decisionrl yet")
+    else:
+        for name, figures in pages.items():
+            stated = next((v for k, v in figures.items() if k in ("Line coverage", "Покрытие строк")), None)
+            if stated is None:
+                problems.append(f"{name}: no coverage figure")
+            elif abs(int(stated.rstrip("%")) - coverage) >= 1:
+                problems.append(
+                    f"{name}: coverage says {stated}, Codecov measures {coverage:.1f}%"
+                )
+
+    unchecked.append("recall@5 — praxis states it in prose, with no eval run pinning it")
+
+    for problem in problems:
+        print(f"  DRIFTED  {problem}")
+    for note in unchecked:
+        print(f"  unchecked  {note}")
+    if not problems:
+        print("\n  Every figure that can be checked matches its source.")
+    return 1 if problems else 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main(Path(sys.argv[1] if len(sys.argv) > 1 else ".")))