diff --git a/endpoints_rules.md b/endpoints_rules.md
index dfd4ed5..b0eb21b 100644
--- a/endpoints_rules.md
+++ b/endpoints_rules.md
@@ -2550,6 +2550,20 @@ Endpoints submissions must include the following metadata:
 | `config_summary_notes` | Free form field from the submitter to contain information not captured by other fields that concatenate into config_summary. |
 | `tps_utilization` | reported_system_tps / (max of all reported_system_tps for all runs) |
 
+`system_desc.json` does not carry `submission_id`, `submission_date`, or `publish_date`. The
+submission pipeline assigns all three, so submitters do not set them:
+
+- The submission ID ([§8.5](#85-result-id)) is recorded only as the `[submission_id]/` directory
+  level ([§8.1](#81-directory-structure)).
+- The submission date is set when the bundle is submitted
+  ([Submission Rules §4.1](endpoints_submission_rules.md#41-rolling-submission-model)).
+- The publication date is set by the publication cohort or embargo
+  ([Submission Rules §4.2](endpoints_submission_rules.md#42-publication-cohorts-and-embargo)).
+
+Any of these keys left over from an earlier round's template is not a defect: the submission
+tooling drops it when it assembles the bundle, and the compliance validator
+([§9.1](#91-automated-checks)) does not flag it.
+
 #### 8.2.1 Template Structure
 
 `results/<system>/<model_name>/r<N>/system_desc.json` contains the fields defined in the table above.
@@ -2662,7 +2676,9 @@ Two identifiers are attached to every submission, and they serve different purpo
 
 The **submission ID** is generated automatically by the submission pipeline as a hash. It is opaque,
 carries no meaning, and exists so the lifecycle tooling can track a bundle through upload, review,
-and amendment.
+and amendment. It is not a submitter-provided field: it appears in the bundle only as the
+`[submission_id]/` directory ([§8.1](#81-directory-structure)), never in `system_desc.json`
+([§8.2](#82-system-description-system_descjson)).
 
 The **result ID** identifies a single published result and is human-readable. A result is one
 published Pareto curve: one system, one benchmark model, one dataset. It is constructed as: