diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 03c671a5a..f4265a777 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -426,6 +426,27 @@ jobs:
       # correct everywhere else). Run the table; do not re-read the regex.
       - name: "`apr`-pinning guard must still turn RED (case table + surfaces)"
         run: bash scripts/check_apr_bin_pinned.sh --self-test
+      # deny.toml is 7.2 KB of licence / banned-crate / source-allowlist policy
+      # that ran in NO workflow -- only `make deny`, which is not a prerequisite
+      # of any tier, so nothing but a human typing it ever invoked it. Meanwhile
+      # cargo-deny 0.20.2 ships INSIDE the CI image already.
+      #
+      # Because nothing ran it, the policy drifted out of sync with the repo's
+      # own deliberate architecture and now fails on main:
+      #     advisories ok, bans FAILED, licenses FAILED, sources ok
+      # Both failures were config drift, not defects (see the deny.toml comments).
+      # Fixed there; wired here so it cannot drift unnoticed again.
+      #
+      # `ci / security` from the reusable workflow runs `cargo audit`, which does
+      # NOT read deny.toml -- so this covers licences, bans and sources, which
+      # nothing else does.
+      - name: cargo-deny (licences, bans, sources, advisories)
+        run: |
+          docker run --rm \
+            -v "${GITHUB_WORKSPACE}:/workspace" \
+            -w /workspace \
+            "$IMAGE" \
+            cargo deny check
       # Poka-yoke: `set` in a SOURCED file mutates the caller's shell. apr_bin.sh
       # opened with `set -euo pipefail`; qwen-story.sh sources it and had chosen
       # `set -uo pipefail` deliberately (it must run every beat and tally the
diff --git a/Makefile b/Makefile
index cd3ccdebe..5fd2858c5 100644
--- a/Makefile
+++ b/Makefile
@@ -214,6 +214,8 @@ tier3:
 	@bash scripts/check_include_files.sh
 	@echo "Checking publish safety (symlinks, companion lookups)..."
 	@bash scripts/check_publish_safety.sh
+	@echo "Checking cargo-deny policy (licences, bans, sources, advisories)..."
+	@$(MAKE) --no-print-directory deny
 	@echo "Checking exclude patterns are root-anchored (CB-510 class)..."
 	@bash scripts/check_exclude_anchored.sh
 	@echo "Checking build.rs crate-root escapes (v0.31.1 yank class)..."
diff --git a/deny.toml b/deny.toml
index 26820aa32..9562f5088 100644
--- a/deny.toml
+++ b/deny.toml
@@ -92,9 +92,40 @@ allow = [
 ]
 confidence-threshold = 0.8
 
+# NCSA, scoped to the ONE crate that carries it rather than added to `allow`
+# above. Provenance:
+#
+#     libfuzzer-sys (MIT OR Apache-2.0) AND NCSA
+#       <- rav1e  <- ravif  <- image  <- aprender-test-lib
+#
+# i.e. it arrives through an AV1 encoder pulled in by `image`, not through
+# anything this project chose directly. NCSA is OSI-approved and FSF Free/Libre
+# and is permissive, so it is acceptable here -- but as an EXCEPTION, so that a
+# future NCSA dependency arriving by some other path is still flagged for a
+# decision instead of being silently admitted.
+[[licenses.exceptions]]
+name = "libfuzzer-sys"
+allow = ["NCSA"]
+
 [bans]
 multiple-versions = "warn"
 wildcards = "deny"
+# Sibling dev-dependencies in this monorepo are DELIBERATELY path-only, with no
+# `version` field:
+#
+#     aprender-core     -> aprender-train, aprender-profile      (dev)
+#     aprender-compute  -> aprender-core, aprender-simulate, ... (dev)
+#     aprender-serve    -> aprender-cuda-edge                    (dev)
+#
+# That is how the crates.io publish cycle is broken: `cargo publish` strips
+# dev-dependencies, so a path-only dev-dep cannot create a circular version
+# requirement. Giving them a `version` would reintroduce the cycle.
+#
+# cargo-deny reports a path dep with no version as a WILDCARD, so without this
+# it flags all 10 as banned -- every one of them a deliberate, load-bearing
+# choice. `wildcards = "deny"` still applies to real (non-path) dependencies,
+# which is the case worth banning.
+allow-wildcard-paths = true
 highlight = "all"
 
 [sources]