diff --git a/services/hermes/skills/master-hermes-on-atlas/references/two-hour-proof-sprint.md b/services/hermes/skills/master-hermes-on-atlas/references/two-hour-proof-sprint.md
new file mode 100644
index 000000000..98f7bbba1
--- /dev/null
+++ b/services/hermes/skills/master-hermes-on-atlas/references/two-hour-proof-sprint.md
@@ -0,0 +1,82 @@
+# Atlas two-hour proof sprint
+
+The sprint establishes usable, defensible evidence for the current Hermes
+claims. It does not award mastery and it cannot retroactively prove historical
+OpenClaw use. Brad performs the work; Hermes gives one task at a time, checks
+the evidence, and withholds later answers until Brad commits to a result.
+
+## 0–15 minutes — control surface and request path
+
+1. Have Brad select `openai-codex/gpt-5.4` for the training session.
+2. In the web UI, have him locate Chat, Sessions, Files, Models, Logs, Skills,
+   Plugins, MCP, Channels, Webhooks, Cron, Profiles, Config, Keys, System, and
+   Documentation.
+3. Have him explain browser → operator agent → model gate → Ollama, including
+   Codex fallback and the separate consumer instance.
+
+Evidence: Brad can identify persistent, shared, and instance-local state and
+can name the GPU owner check without changing it.
+
+## 15–35 minutes — permissions and safety
+
+1. Have Brad inspect the operator ServiceAccount, ClusterRole, NetworkPolicy,
+   and approval deny list using read-only commands.
+2. Have him produce an allow/deny matrix for pod/log reads, Secret reads,
+   exec, deployment mutation, Flux reads, Flux reconcile, and internal
+   evidence endpoints.
+3. Require him to explain why prompt instructions and skills are not the
+   primary security boundary.
+
+Evidence: correct identity/RBAC/network/approval layering with no Secret values
+read and no mutation attempted.
+
+## 35–75 minutes — perform the claimed triage path
+
+Invoke `$triage-titan-test-failures`. Brad must:
+
+1. check Ariadne diagnosis and bundle freshness;
+2. identify one current or retained failed suite/build;
+3. correlate Jenkins log/artifact evidence;
+4. correlate Git/Flux revision and Kubernetes environment health;
+5. query the relevant Pushgateway/VictoriaMetrics quality signal and identify
+   the matching Grafana context;
+6. produce Finding, Confidence, Evidence, Likely cause, Blast radius, Next
+   checks, Repo-side fix, and Approval required.
+
+Hermes may suggest a read-only command after Brad explains why it is needed.
+Do not let the built-in diagnosis replace Brad's own evidence correlation.
+
+Evidence: an 8/10 or better result under the triage rubric with no invented
+facts and no unacknowledged mutation.
+
+## 75–105 minutes — build the reusable workflow
+
+1. Have Brad inspect the mounted `triage-titan-test-failures` skill and explain
+   its trigger, fixed procedure, variable evidence, and security boundary.
+2. Have Brad choose one genuinely repeated sub-workflow from the triage.
+3. Guide him to create a new skill in the operator's writable skill area. Do
+   not edit the mounted Flux skill and do not write the skill for him.
+4. Test one expected trigger and one adversarial/non-trigger case.
+5. Have Brad make one improvement based on the test results.
+
+Evidence: Brad can find the skill in the Skills UI, explain every instruction,
+and show a passing and adversarial evaluation.
+
+## 105–120 minutes — supervised proposal and claim audit
+
+1. Have Brad turn the triage into the smallest repo-side proposal without
+   applying it.
+2. Have him identify every step requiring approval: file edit, commit/push,
+   pipeline trigger, Flux reconcile, credential/environment change, or manual
+   backfill.
+3. Audit the interview answer sentence by sentence using `mastery-rubric.md`.
+4. Offer to record progress only after Brad approves the write.
+
+Minimum honest outcome:
+
+- current Hermes deployment and supervised triage use are demonstrated;
+- the established Atlas automation path and read-only boundary are explained;
+- one reusable writable skill is created, tested, and improved by Brad;
+- autonomous production mutation is not claimed;
+- historical OpenClaw replacement remains unproven unless independent retained
+  evidence is produced.