diff --git a/bench/cdeb/AUTHORIZATION.md b/bench/cdeb/AUTHORIZATION.md
new file mode 100644
index 00000000..e7ebdd47
--- /dev/null
+++ b/bench/cdeb/AUTHORIZATION.md
@@ -0,0 +1,59 @@
+# Source authorization for the CDEB corpus
+
+CDEB reads real decision records out of real repositories and publishes what it
+finds. §3.3 requires repository ownership and decision authorship to be
+disclosed, and CDEB-10 lists privacy authorization among the things that must
+exist before the corpus is sealed. This file is that authorization, and the
+`source_authorization_id` in each repository's freeze entry (§6.1) points here.
+
+A public repository is not consent to be studied. Visibility and authorization
+are different things, and the study needs the second one named by the person who
+holds it — not inferred from the first.
+
+## Granted
+
+The owner of the repositories below authorized their use as CDEB source data,
+and authorized publishing the result transparently, including repository
+ownership and decision authorship, on 2026-08-11.
+
+| authorization_id | repository | owner |
+|---|---|---|
+| `auth-owner-2026-08-11` | agent-operator-score | MongLong0214 |
+| `auth-owner-2026-08-11` | logic-pro-mcp | MongLong0214 |
+| `auth-owner-2026-08-11` | stock-ai-newsletter | MongLong0214 |
+| `auth-owner-2026-08-11` | hermes-agent (fork) | MongLong0214 |
+
+## What the authorization does not extend to
+
+**Commits authored by other people.** `hermes-agent` is a fork of an upstream
+project, and the substantial majority of its recent history is upstream authors'
+work pulled in — 1,052 of the commits in one recent week. The owner can
+authorize the use of their own repository and their own decisions; they cannot
+authorize on behalf of upstream contributors. Any task drawn from that
+repository must rest on a record the owner authored, and §3.2's candidate
+registry field `decision_source_refs` is where that is checked rather than
+assumed.
+
+**The CommitLore repository itself.** §3.3 excludes it from the primary corpus,
+and no authorization changes that. It is excluded because the product's own
+decisions are not independent of the product being measured, which is a
+methodological exclusion rather than a permission one.
+
+## Consequence for what the result may say
+
+Every authorized repository is operated by the same person who authored this
+benchmark and the product it measures. That is §3.3's **Tier B**, and the rule
+attached to it is not optional: the result must read *"five author-operated
+repositories"* and must not describe itself as independent external validation.
+
+Tier A would require at least two repositories where the decision author or
+accepting reviewer differs from the benchmark and product author. Nothing here
+reaches that, and no wording may imply otherwise.
+
+## Corpus arithmetic, stated plainly
+
+§3.3 requires five repositories, and the exclusion of the CommitLore repository
+leaves four authorized ones. A fifth is needed before the corpus can be sealed,
+and §3.2's rule 5 is explicit about what may not be done to close that gap:
+when candidates are short, the study stops rather than lowering the bar or
+creating records for it.