diff --git a/CLAUDE.md b/CLAUDE.md
index ab435098..53fcc056 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -32,7 +32,8 @@ Two kinds of document, edited differently:
   first, and a correction is a new entry that names the one it corrects. An
   accepted ADR in `docs/adr/` is immutable: a changed decision is a new ADR
   that supersedes it. An old ADR that says something has not happened yet
-  stays as written.
+  stays as written. An exercise in `docs/exercises/` is a record too: a re-run
+  is a new file.
 - **Current-state documents** must stay true: `README.md`,
   `docs/hardware.md`, `docs/network.md`, `docs/architecture.md`,
   `docs/security.md`, `docs/roadmap.md`, `docs/runbooks/` and
diff --git a/README.md b/README.md
index 01cf6cef..1dbd0d37 100644
--- a/README.md
+++ b/README.md
@@ -224,6 +224,7 @@ inbound passes may still reach it. Data flow and the maintained Mermaid topology
 │   ├── observability.md  security.md  roadmap.md  changelog.md
 │   ├── diagrams/             # the network diagram (SVG) and its predecessors
 │   ├── adr/                  # 94 ADRs — architecture decision records
+│   ├── exercises/            # purple-team exercise records — see its README
 │   └── runbooks/             # 49 runbooks; successor-handover.md is the front door
 └── Makefile                  # make help
 ```
diff --git a/docs/exercises/README.md b/docs/exercises/README.md
new file mode 100644
index 00000000..80dc6d7c
--- /dev/null
+++ b/docs/exercises/README.md
@@ -0,0 +1,47 @@
+# Exercises
+
+One file per purple-team exercise against the lab domain: what was run,
+what the estate was expected to see, what it actually saw, and what changed
+because of it. The point is the gap between *expected* and *observed*. An
+exercise that finds no gap is still worth a file, because it proves a
+detection works on a date.
+
+## The rules
+
+- **One exercise, one file**, named `YYYY-MM-DD-short-name.md`, copied from
+  [`TEMPLATE.md`](TEMPLATE.md).
+- **An exercise file is a record.** Like [`changelog.md`](../changelog.md), it
+  is never rewritten after the day. A re-run is a new file that links the
+  earlier one, so the history shows a detection being fixed, not just the end
+  state.
+- **Every technique carries its MITRE ATT&CK ID**, so exercises can be searched
+  by technique and coverage can be counted.
+- **Only inside the lab.** Every target is on the lab domain or the range
+  ([ADR-0007](../adr/0007-defensive-estate-and-offensive-range.md),
+  [ADR-0014](../adr/0014-put-ifrit-on-imaginationlan-and-give-the-targets-no-route.md)).
+  Nothing here is run against the household tiers or anything outside the
+  house.
+- **Describe the activity, not the payload.** Name the technique and the tool
+  and say what was done. No working exploit code, no captured credentials or
+  hashes, no tokens. This repository is public, and
+  [`security.md`](../security.md#what-this-repository-deliberately-does-not-publish)
+  applies here as everywhere.
+- **A missed or weak detection opens an issue.** The issue is the follow-up;
+  the exercise file links it.
+
+## Where the evidence comes from
+
+| Source | Lives on | What it sees |
+| --- | --- | --- |
+| Wazuh | `odin` ([`stacks/soc`](../../stacks/soc)) | Agent alerts from the domain's hosts, Sysmon included |
+| Velociraptor | `odin` ([`stacks/soc`](../../stacks/soc)) | Endpoint state, collected on demand |
+| Zeek | `fenrir` ([`stacks/sensor`](../../stacks/sensor)) | Network metadata from the mirrored lab bridge |
+| Suricata | `morpheus` | Alerts at the firewall, the chokepoint ([ADR-0006](../adr/0006-detect-at-the-chokepoint.md)) |
+| Loki and Grafana | `alexander` ([`stacks/lab`](../../stacks/lab)) | The lab's logs, searchable by time window |
+
+## Index
+
+Newest first. Add a row with each new file.
+
+| Date | Exercise | Techniques | Result |
+| --- | --- | --- | --- |
diff --git a/docs/exercises/TEMPLATE.md b/docs/exercises/TEMPLATE.md
new file mode 100644
index 00000000..9be98a24
--- /dev/null
+++ b/docs/exercises/TEMPLATE.md
@@ -0,0 +1,50 @@
+# YYYY-MM-DD: Short name of the exercise
+
+**Date:** YYYY-MM-DD · **Operator:** you · **Issue:** #N, if one exists ·
+**Re-run of:** link to an earlier exercise, or *none*
+
+## Objective
+
+One or two sentences: what this exercise tests, and which detection it is
+meant to prove or disprove.
+
+## Scope
+
+- **Attacker:** the host or VM the activity came from.
+- **Targets:** the lab hosts it touched. Lab domain or range only.
+- **Window:** start and end time, with the time zone, so the logs can be
+  pulled for exactly this period.
+- **Lab state:** anything switched on for this run, such as which deliberate
+  weakness tags were enabled
+  ([ADR-0083](../adr/0083-give-the-lab-domain-its-deliberate-weaknesses-as-switchable-tags.md)).
+
+## Techniques
+
+One row per step, in the order run. Describe the activity, not the payload.
+
+| # | Time | ATT&CK ID | Technique | Tool | What was done |
+| --- | --- | --- | --- | --- | --- |
+| 1 | hh:mm | T0000 | Name | Tool | One line |
+
+## Expected detections
+
+Written before the run, so the result cannot shape the expectation.
+
+| # | Source | Expected rule or signal |
+| --- | --- | --- |
+| 1 | Wazuh / Zeek / Suricata / Velociraptor / Loki | Rule ID or query, and what it should show |
+
+## Observed
+
+| # | Detected? | Where, and how long after the step | Notes |
+| --- | --- | --- | --- |
+| 1 | Yes / Partly / No | Source and delay | Alert text summarised, not pasted with secrets |
+
+## Gaps and fixes
+
+For every *No* or *Partly* above: why it was missed, and what changes.
+Each fix links the issue or PR that makes it.
+
+## Lessons
+
+What this exercise taught, in a few lines.