RichardoC · RichardoC · Jul 24, 2026 · Jul 24, 2026
diff --git a/docs/COURSE_SPEC.md b/docs/COURSE_SPEC.md
@@ -13,8 +13,7 @@ Model, version 1.2)** package with **LMS (Learning Management System)-reported s
 
 - **Course license: CC BY-SA 4.0** (ShareAlike, required by the Agentic PDF).
 - Attribution + "changes were made" + "not endorsed by OWASP" on the About page.
-- **Provenance**: generated with the **pi coding-agent harness v0.79.8** using LLM (Large Language
-  Model) models **glm-5.2** and **kimi k2.7-coder**. Not endorsed by OWASP.
+- **Provenance**: generated with the assistance of an AI coding agent. Not endorsed by OWASP.
 - All prose is written in our own words as an adaptation; quiz questions are original.
 
 ## Conventions (apply on every page and in every quiz)
@@ -38,6 +37,8 @@ Model, version 1.2)** package with **LMS (Learning Management System)-reported s
 - **Scoring**: a **Configure Scores** page sets project scoring; theme quizzes (25%) + final quiz
   (75%) **all count** toward the LMS score. Final quiz pass mark **80%**.
 - **SCORM 1.2** export → `cmi.core.score.raw`, `cmi.core.lesson_status`, `cmi.interactions.n`.
+- **Option order**: every quiz question sets `answerOrder="random"` so the answer options
+  shuffle on each attempt (Nottingham per-question "Answer Order" property).
 - **Question types used** (all selection-based, no free text): Multiple Choice (single),
   Multiple Response (select all), True/False, Matching, Categorise.
 - Editor opened directly at `/edit.php?template_id=<id>`; **Publish** before export.
@@ -74,7 +75,7 @@ Real-world incident callout (where Appendix D has one) · Secure-coding mitigati
 - Attribution to both OWASP sources (name, author, source link, license link).
 - "Changes were made: this course is an adapted summary/derivative."
 - "Not endorsed by OWASP."
-- Provenance: pi v0.79.8, models glm-5.2 and kimi k2.7-coder.
+- Provenance: generated with the assistance of an AI coding agent.
 
 ### Page 3 — Introduction to secure code development  [Content]
 - What secure code development is; why it matters; cost of defects left vs right of the
@@ -128,7 +129,7 @@ Real-world incident callout (where Appendix D has one) · Secure-coding mitigati
   *An authenticated user edits a URL parameter to view another user's account* (correct).
   Distractors: an attacker sends a crafted SQL (Structured Query Language) string; a server uses
   a deprecated hash; an app ships with default admin credentials.
-- **Multiple response**: "Which are effective mitigations for authentication failures?" ✅ MFA
+- **Multiple response**: "Which are effective mitigations against authentication attacks?" ✅ MFA
   (Multi-Factor Authentication); ✅ hashed + salted passwords; ✅ lockout on repeated failures;
   ❌ plaintext password reminders by email; ❌ client-side authorization checks.
 - **True/False**: "A delegated agent should reuse the delegating user's full permissions when
@@ -157,8 +158,9 @@ Real-world incident callout (where Appendix D has one) · Secure-coding mitigati
   XPath (XML Path) injection, expression/language injection.
 - Examples: SQLi via concatenated strings, command injection, expression injection.
 - Mitigations: parameterised queries / prepared statements, allowlist input validation,
-  escaping/encoding, ORM (Object-Relational Mapping) safe APIs, least privilege on DB
-  (database) accounts.
+  escaping/encoding, parameterised ORM (Object-Relational Mapping) query methods (an ORM is
+  not inherently injection-proof — never concatenate raw input into its query strings), least
+  privilege on DB (database) accounts.
 
 ### Page 12 — Agent Goal Hijack (Agentic, ASI01)  [Content]
 - Overview: manipulating an agent's objectives, task selection, or decision pathways via prompt
@@ -560,14 +562,20 @@ All framed with full descriptive names (no item codes as the question subject).
   Memory & Context Poisoning = Agentic; Cryptographic Failures = Classic; Rogue Agents = Agentic;
   Security Misconfiguration = Classic; Cascading Failures = Agentic.
 
+### Page 44 — Course complete  [Content]
+- A closing page so it is unambiguous that the training is over: congratulates the learner,
+  reminds them their result is recorded by the LMS (Learning Management System) and that the
+  80% pass mark / last-attempt rule applies, and tells them they may close the window. Plain
+  Text page type.
+
 ### [Scores] — Configure Scores
 - Project scoring enabled; theme quizzes (25%) + final quiz (75%) all count; final pass mark
   80%; per-question weighting set at build time.
 
 ---
 
 ## Totals
-- 43 learner-visible pages + 1 Configure Scores page.
+- 44 learner-visible pages + 1 Configure Scores page.
 - 20 item pages (10 classic + 10 agentic), 7 theme intros, 7 theme quizzes, 1 final quiz,
   4 front-matter + 5 back-matter pages.
 - All quiz questions selection-based; auto-graded; LMS-reported via SCORM 1.2.

diff --git a/docs/COURSE_VERIFICATION.md b/docs/COURSE_VERIFICATION.md
@@ -46,7 +46,7 @@ Glossary/abbreviations, References & further reading — all present.
 CC BY-SA 4.0 (Creative Commons Attribution-ShareAlike 4.0); both OWASP sources
 attributed (Agentic 2026 / genai.owasp.org / CC BY-SA 4.0; Top 10:2025 /
 owasp.org/Top10/2025/ / CC BY 3.0); "changes were made"; "not endorsed by OWASP";
-provenance (pi v0.79.8, models glm-5.2 and kimi k2.7-coder).
+provenance (generated with the assistance of an AI coding agent).
 
 ### Acronym expansions ✅ (sample all present)
 IDOR (Insecure Direct Object Reference), RCE (Remote Code Execution),
@@ -206,3 +206,75 @@ empty options, 27 `delaySecs=0`, autosave script present). A Publish
 round-trip preserved the `{{BUILD_VERSION}}` placeholder intact in
 `data.xml` (CKEditor does not mangle the braces). See `PROJECT_CONTEXT.md` §
 "Build version stamping" for full detail.
+
+## Feedback pass: option randomisation + content fixes (post-0.0.6)
+
+A round of review feedback was applied to `source/data.xml` +
+`source/preview.xml` (kept byte-identical). All changes are content-level and
+use engine-supported attributes, so they survive the release-workflow XOT
+import/export and the editor round-trip.
+
+1. **Option order randomised per attempt** ✅ — every `<question>` now carries
+   `answerOrder="random"` (45/45). Nottingham's HTML5 quiz model
+   (`models_html5/quiz.html` ~line 165) shuffles each question's options on
+   quiz load/restart; scoring is computed from the shuffled array, so SCORM
+   tracking is unaffected. This removes the previous bias where the correct
+   answer was frequently the first option (most visible in the final quiz).
+   `answerOrder` is a declared wizard property (`quiz.xwd` "Answer Order"),
+   so the editor round-trips it natively (like `trackingWeight`).
+2. **Theme 1 Quiz Q2 wording** ✅ — "Which are effective mitigations for
+   authentication failures?" → "Which are effective mitigations against
+   authentication attacks?" (clearer; the mitigations defend against attacks,
+   not "for" failures).
+3. **Injection (A05:2025) page — ORM bullet** ✅ — the line "Use safe ORM
+   (Object-Relational Mapping) APIs and stored procedures correctly." was
+   replaced with "Use parameterised query APIs and stored procedures
+   correctly; an ORM (Object-Relational Mapping) is not automatically safe —
+   use its parameterised methods and never concatenate raw input into
+   queries." An ORM is not inherently injection-proof, so it should not be
+   listed as a standalone mitigation.
+4. **Theme 7 Quiz Q2 — logging tool calls** ✅ — the correct answer was
+   inverted (False was marked correct, implying logging tool calls is *not*
+   recommended). Swapped so **True** is correct, matching the Theme 7 content
+   page ("You should log tool calls, inter-agent messages, goal state and
+   action sequences") and the final-quiz question that treats missing agent
+   logs as a Security Logging and Alerting Failure. We must be able to prove
+   what an agent did and why; privacy is handled by privacy-aware logging,
+   not by not logging.
+5. **Final quiz — over-scoped CRM tool question** ✅ — the question whose
+   correct answer was "Tool Misuse" (the over-scoped CRM summariser) was
+   reframed to ask for the *principle that prevents* the risk, with
+   "Scope each tool to least privilege: only the data and actions the task
+   requires" as the correct answer. (Over-scoped tools do map to Tool Misuse
+   / ASI02 per the course's own page, but the lesson the scenario teaches is
+   least-privilege scoping; the reframe makes that the answer.)
+6. **"Course complete" page added** ✅ — a new `<text>` page
+   (`linkID="PG1781901000001"`, name "Course complete") was added after the
+   final quiz so it is unambiguous that the training is over. It is a Plain
+   Text page (not Bullets), so the `delaySecs="0"` count is unchanged at 27.
+   Top-level page count is now 44 (43 learner content/quiz pages + this
+   closing page). Per the SCORM completion analysis in `PROJECT_CONTEXT.md`,
+   this page becomes the untracked last content node (no score, no
+   interaction — harmless), and the Final quiz's completion is now tracked.
+
+### Local checks (run against `source/data.xml`; XOT re-export still required per convention 9)
+
+- 45 `<question>` elements, 45 with `answerOrder="random"`.
+- 8 `<quiz>` nodes; `trackingWeight` = 1×7 + 21; `trackingMode="full"`;
+  `trackingPassed="80%"`; `judge="true"` ×8.
+- `unmarkForCompletion="true"` ×1 (Welcome); `xPersistProgress` ×1 (autosave
+  script); `{{BUILD_VERSION}}` ×1 (placeholder, substituted at build time).
+- `delaySecs="0"` ×27 (unchanged — the new page is Plain Text).
+- 0 empty option nodes; every question has ≥2 options and ≥1 correct.
+- `tools/render_preview.py` renders 44 pages (final page = "Course complete")
+  without error.
+- `source/data.xml` and `source/preview.xml` are byte-identical.
+
+**Pending (convention 9 — MANDATORY before merge)**: push the updated
+`source/data.xml` + `source/preview.xml` into a running XOT instance,
+confirm `play.php` returns 200, open the HTML5 editor to confirm a Publish
+round-trip preserves `answerOrder="random"` on all 45 questions (and the
+other tracking attrs), export a SCORM package, and grep the exported
+`template.xml` for the checklist values in `PROJECT_CONTEXT.md`. Then walk a
+quiz end-to-end (answer → Check → Next → complete → Restart) and confirm
+options reorder between attempts.
diff --git a/docs/PROJECT_CONTEXT.md b/docs/PROJECT_CONTEXT.md
@@ -47,9 +47,13 @@ with agents and building agents.
 
 ## Current state of the course (as last exported)
 
-- **48 learner pages** in spec order: front matter (4) → 7 themes (each: intro +
-  classic item page(s) + agentic item page(s) + theme quiz) → back matter (4:
-  secure-SDLC integration, mapping, glossary, references) → final quiz.
+- **44 top-level pages** in `source/data.xml` (43 learner content/quiz pages + 1
+  final **"Course complete"** page added after the final quiz so it is clear the
+  training is over): front matter (4) → 7 themes (each: intro + classic item
+  page(s) + agentic item page(s) + theme quiz) → back matter (4: secure-SDLC
+  integration, mapping, glossary, references) → final quiz → Course complete.
+  (Earlier docs/reports may say 43 or 48; the authoritative count is whatever
+  is in `source/data.xml`.)
 - **Content page types**: pages 1–8 are `text` (Plain Text); pages from Theme 2
   onward are `bullets` (Bullets/Timed Content) with **`delaySecs="0"`** so they
   do **not** auto-play (a deliberate fix — see "Conventions" below).
@@ -59,6 +63,14 @@ with agents and building agents.
   nested `<question>` children. **45 scored questions total** (23 theme +
   18 final + 4 … see `COURSE_VERIFICATION.md`). Every question has ≥2 options
   and ≥1 `correct="true"`; **no empty option nodes**.
+- **Option order is randomised per attempt**: every `<question>` carries
+  `answerOrder="random"` (the Nottingham wizard's per-question "Answer Order"
+  property; see `quiz.xwd` and `models_html5/quiz.html` line ~165). The HTML5
+  player shuffles each question's options on quiz load/restart, and scoring is
+  computed from the shuffled array so SCORM tracking stays correct. This
+  removes the previous bias where the correct answer was frequently the first
+  option (especially in the final quiz). The attribute is editor-declared, so
+  the XOT HTML5 editor round-trips it natively (like `trackingWeight`).
 - **Final quiz** has **18 distinct questions** (fresh scenarios, not duplicates
   of the theme-quiz questions), mix of single-answer risk-identification and
   true/false.
@@ -84,8 +96,7 @@ These were agreed with the course owner and must be preserved:
    Applications 2026, genai.owasp.org, CC BY-SA 4.0; OWASP Top 10:2025,
    owasp.org/Top10/2025/, CC BY 3.0); a "changes were made" notice; a
    "not endorsed by OWASP" notice; and **provenance**:
-   *"Generated with the pi coding-agent harness v0.79.8 using LLM models
-   glm-5.2 and kimi k2.7-coder."*
+   *"Generated with the assistance of an AI coding agent."*
 2. **Acronym expansion**: the first time a security acronym is used **on a
    page**, give the full form in brackets, e.g. `TOCTOU (Time-of-check to
    time-of-use)`. Reuse the acronym thereafter on that page; expand again on the
@@ -125,6 +136,14 @@ These were agreed with the course owner and must be preserved:
    confirm `trackingMode`, `trackingWeight`, `trackingPassed`, the question
    count, and the empty-option count are all as expected). Do not merge a
    `source/data.xml` change that has not passed this validation.
+10. **Option order must be randomised per attempt**: every `<question>` must
+    carry `answerOrder="random"` (Nottingham wizard's per-question "Answer
+    Order" property, `quiz.xwd`; honoured by `models_html5/quiz.html`). This
+    shuffles each question's options on quiz load/restart and removes the
+    pattern of the correct answer always being the first option. Scoring is
+    computed from the shuffled array, so SCORM tracking is unaffected. When
+    adding a new question, set `answerOrder="random"` on it. Keep `data.xml`
+    and `preview.xml` byte-identical.
 
 ## Known deviations (acknowledged, not to "fix" without asking)
 
@@ -153,7 +172,9 @@ being sent (75, then 100). `cmi.core.lesson_status` was never reported as
 **Root cause** (a Xerte Nottingham engine bug, worked around at the content
 level): with `navigation="Menu with Page Controls"`, Xerte inserts a built-in
 Table-of-Contents page as `x_pages[0]` **after** `toCompletePages` is built
-from `currActPage` over the 43 content nodes. So every content page's
+from `currActPage` over the content nodes (43 when this fix was designed;
+44 now that a **Course complete** page follows the final quiz — the analysis
+is identical, only the last `page_nr` shifts). So every content page's
 `page_nr` (used by `exitInteraction` to mark `completedPages[i]`) is
 `currActPage + 1` — a systematic **off-by-one**:
 
@@ -179,16 +200,19 @@ against `trackingPassed="80%"`. Verified locally with a SCORM 1.2 mock API:
 a completed run sends `cmi.core.lesson_status="passed"`, `cmi.core.exit=""`,
 `cmi.core.score.raw`, `score.min`, `score.max`, and `cmi.core.session_time`.
 
-**Trade-off**: the Final quiz (`page_nr == 43`) is still not in
-`toCompletePages` (there is no index 43), so its *completion* is not tracked
-in `completedPages` — but its *score* is still tracked via its interactions
-and `trackingWeight="21"`, so the LMS grade and pass/fail are correct. The
-Welcome page is no longer "required" for completion, which is acceptable for
-an intro page. A proper fix would be in the Xerte engine (`xenith.js`: build
-`toCompletePages` from `page_nr` after the menu is inserted, or do not insert
-the menu into `x_pages` for tracking); this content workaround is used because
-the release workflow builds from the XOT engine and an engine patch would not
-persist across XOT rebuilds.
+**Trade-off** (with the Course complete page now present): the *last* content
+node — the **Course complete** page (`page_nr == 44`), a static
+acknowledgement page — is not in `toCompletePages` (there is no index 44), so
+its *completion* is not tracked; this is harmless (it carries no score and no
+interaction). The Final quiz (`page_nr == 43`) **is** now in
+`toCompletePages`, so its completion is tracked, and its *score* is tracked
+via its interactions and `trackingWeight="21"`. The Welcome page is no longer
+"required" for completion, which is acceptable for an intro page. A proper
+fix would be in the Xerte engine (`xenith.js`: build `toCompletePages` from
+`page_nr` after the menu is inserted, or do not insert the menu into `x_pages`
+for tracking); this content workaround is used because the release workflow
+builds from the XOT engine and an engine patch would not persist across XOT
+rebuilds.
 
 **Do not remove `unmarkForCompletion="true"` from the Welcome page** without
 a replacement fix — re-introducing the off-by-one will silently break LMS
@@ -392,13 +416,15 @@ unzip -p Secure_code_development_scorm.zip template.xml | grep -oE 'trackingPass
 unzip -p Secure_code_development_scorm.zip template.xml > /tmp/c.xml
 grep -c "Trial MCQ" /tmp/c.xml                 # 0
 grep -oE '<question' /tmp/c.xml | wc -l           # 45 (23 theme + 18 final + 4 …)
+grep -oE 'answerOrder="random"' /tmp/c.xml | wc -l  # 45 (every question shuffles options per attempt)
 grep -oE 'trackingPassed="[^"]*"' /tmp/c.xml   # 80%
 grep -oE 'judge="true"' /tmp/c.xml | wc -l     # 8
 grep -oE 'trackingWeight="[0-9]+"' /tmp/c.xml   # 7x "1" + 1x "21"
 grep -oE 'trackingMode="[a-z_]+"' /tmp/c.xml   # "full"
 grep -oE 'delaySecs="0"' /tmp/c.xml | wc -l    # 27 (all bullets pages)
 grep -c 'xPersistProgress' /tmp/c.xml        # 1 (SCORM autosave script present on root)
 grep -oE 'unmarkForCompletion="true"' /tmp/c.xml | wc -l  # 1 (Welcome page)
+grep -oE 'name="Course complete"' /tmp/c.xml | wc -l  # 1 (final "Course complete" page after the final quiz)
 # build version: in a release export {{BUILD_VERSION}} is substituted; in source/preview it is the placeholder
 grep -oE 'Version:</strong> [^<&]*' /tmp/c.xml              # e.g. 'Version:</strong> v0.0.6 (...)'
 grep -c '{{BUILD_VERSION}}' /tmp/c.xml                     # 0 in a release export (1 in committed source)