feat(iter-8): anti-join levels 66–67 — LEFT JOIN…IS NULL and NOT EXISTS

IMPROVEMENT_PLAN.md

@@ -38,9 +38,18 @@ Implemented in the follow-up PR:
   aria-labels on icon buttons, and a stacking responsive layout below the `lg` breakpoint.
 - **Pillar 5:** MIT LICENSE, CONTRIBUTING.md with a level-authoring guide.
 
-Still open: build-time expected results (3.4), the expected-vs-actual value diff (4.1 — the
-shape reveal shipped; a value-level diff is a deliberate product decision), and a deeper
-mobile/touch pass (4.4).
+Implemented since (iteration 8):
+
+- **4.5 (partial):** the anti-join syllabus gap is closed — level 66 teaches the
+  `LEFT JOIN … IS NULL` anti-join, level 67 teaches correlated `NOT EXISTS` (and the
+  `NOT IN` NULL trap). Remaining 4.5 gaps: a dedicated date-functions drill and earlier
+  `CASE` exposure are partially covered by existing levels (28, 38, 62, 65) — revisit
+  only if player feedback shows a hole.
+
+Still open: build-time expected results (3.4 — deprioritized: the level-up modal now
+deliberately reveals the model answer after a win, so hiding solutions from the bundle
+buys nothing), the expected-vs-actual value diff (4.1 — the shape reveal shipped; a
+value-level diff is a deliberate product decision), and a deeper mobile/touch pass (4.4).
 
 ---
 

evolution_log.md

@@ -4,6 +4,32 @@ Each entry records one autonomous improvement iteration.
 
 ---
 
+## Iteration 8 — 2026-06-12
+
+### [Game Design Tweaks]
+
+- **Levels 66–67: the anti-join gap is closed.** The 65-level catalog contained exactly one
+  `EXISTS` and no dedicated `LEFT JOIN … IS NULL` drill — yet "find the rows in A with no
+  match in B" is among the most common SQL interview questions (IMPROVEMENT_PLAN.md §4.5).
+  - **Level 66** *(Expert, difficulty 3)* — "Unfinanced Fleet Exposure (Anti-Join)": vessels
+    with no trade finance facility on file, via `LEFT JOIN trade_finance_facilities … WHERE
+    f.facility_id IS NULL`, ordered by `dwt_tonnes` DESC. The description teaches why the
+    NULL test belongs on the right table's primary key. The four matching vessels are exactly
+    the ones not owned by LCB customers — uncollateralised third-party carriers, which makes
+    the compliance narrative true in the data, not just flavour text.
+  - **Level 67** *(Expert, difficulty 4)* — "Lending Whitespace: Depositors Without Loans
+    (NOT EXISTS)": customers holding accounts but no loans, with `total_deposits` aggregated,
+    via correlated `NOT EXISTS`. The description teaches the `NOT IN` NULL trap (one NULL in
+    the subquery silently yields zero rows) — the reason production code prefers `NOT EXISTS`.
+  - Both levels' results were verified non-empty and deterministic against the live seed
+    (distinct ORDER BY keys, 4 and 5 rows respectively) before authoring; the generic
+    level-integrity and determinism suites picked them up automatically (375 tests green).
+- **`LevelUpModal` next-hint map** extended with entries for 66 (LEFT JOIN anti-join) and 67
+  (NOT EXISTS); sentinel advanced 65 → 67. Level counts everywhere else derive from
+  `levels.length` (Iteration 5's progression module), so no other touch points existed.
+
+---
+
 ## Iteration 7 — 2026-06-12
 
 ### [UI/UX Improvements]

src/components/LevelUpModal.tsx

@@ -21,6 +21,8 @@ const EPOCH_NEXT_HINT: Record<number, string> = {
   63: 'Rolling volatility: SQRT(AVG(x²) − AVG(x)²) — population std dev via window functions',
   64: 'NTILE bucketing: equal-count portfolio tranching for Basel III capital adequacy reporting',
   65: 'Running balance: SUM() OVER (ROWS UNBOUNDED PRECEDING) — the universal treasury ledger pattern',
+  66: 'Anti-joins: LEFT JOIN … IS NULL — finding the rows that have no match',
+  67: 'NOT EXISTS: the correlated anti-join without the NOT IN NULL trap',
 };
 
 export default function LevelUpModal() {
@@ -57,7 +59,7 @@ export default function LevelUpModal() {
     return () => window.removeEventListener('keydown', onKey);
   }, [showLevelUp, handleClose]);
 
-  const nextHintKey = [10, 20, 30, 40, 44, 54, 57, 59, 61, 63, 64, 65].find((k) => currentLevel < k) ?? 65;
+  const nextHintKey = [10, 20, 30, 40, 44, 54, 57, 59, 61, 63, 64, 65, 66, 67].find((k) => currentLevel < k) ?? 67;
   const xpEarned    = xpFor(currentLevel);
 
   return (

src/data/levels/expert.ts

@@ -955,4 +955,81 @@ SELECT month,
     epoch: 'Expert',
     difficulty: 4,
   },
+
+  // ============================================================
+  // ANTI-JOIN PATTERNS (Levels 66–67)
+  // LEFT JOIN … IS NULL; correlated NOT EXISTS
+  // ============================================================
+
+  {
+    id: 66,
+    title: 'Unfinanced Fleet Exposure (Anti-Join)',
+    description: `Trade Finance Compliance has flagged a gap: some vessels moving cargo through our network have **no trade finance facility on file** — no Letter of Credit, no Shipping Guarantee, nothing. Every one of them is uncollateralised operational risk.
+
+Find every vessel with no row in \`trade_finance_facilities\`. Return \`vessel_name\`, \`vessel_type\`, \`flag_state\`, and \`dwt_tonnes\`, ordered by \`dwt_tonnes\` descending (largest exposure first).
+
+Use the **anti-join** pattern: \`LEFT JOIN\` the facilities table, then keep only the rows where the join found no match — \`WHERE f.facility_id IS NULL\`. Test a column that can never be NULL in a real match (the primary key is the safe choice): if it's NULL after a LEFT JOIN, the match doesn't exist.
+
+"Find the rows in A with no match in B" is one of the most common interview questions there is — customers with no orders, users with no logins, products never sold. This LEFT JOIN form is the classic answer.`,
+    hint: "LEFT JOIN trade_finance_facilities f ON f.vessel_id = v.vessel_id, then WHERE f.facility_id IS NULL. The NULL test must be on the right-hand (facilities) table — that's what marks an unmatched row.",
+    seedQuery: `SELECT v.vessel_name,
+       v.vessel_type,
+       v.flag_state,
+       v.dwt_tonnes
+  FROM vessels v
+  LEFT JOIN trade_finance_facilities f ON
+ WHERE f.facility_id IS
+ ORDER BY v.dwt_tonnes DESC`,
+    solutionQuery: `SELECT v.vessel_name,
+       v.vessel_type,
+       v.flag_state,
+       v.dwt_tonnes
+  FROM vessels v
+  LEFT JOIN trade_finance_facilities f ON f.vessel_id = v.vessel_id
+ WHERE f.facility_id IS NULL
+ ORDER BY v.dwt_tonnes DESC`,
+    epoch: 'Expert',
+    difficulty: 3,
+  },
+
+  {
+    id: 67,
+    title: 'Lending Whitespace: Depositors Without Loans (NOT EXISTS)',
+    description: `The Retail Lending desk wants a cross-sell target list: customers who keep deposits with LCB but have **never taken a loan** from us. Their balances tell us they trust the bank — the lending relationship is pure whitespace.
+
+Find every customer who holds at least one account but has no row in \`loans\`. Return \`customer_name\`, \`segment\`, \`credit_score\`, and \`total_deposits\` — the sum of their account balances, rounded to 2 dp — ordered by \`total_deposits\` descending.
+
+Use a correlated \`NOT EXISTS\` subquery: \`WHERE NOT EXISTS (SELECT 1 FROM loans l WHERE l.customer_id = c.customer_id)\`. The inner JOIN to \`accounts\` already restricts the list to account holders, so the aggregation and the anti-join compose in one pass.
+
+Prefer \`NOT EXISTS\` over \`NOT IN\` for anti-joins: if the \`NOT IN\` subquery ever returns a NULL, the whole predicate goes unknown and you silently get **zero rows** — a classic production bug. \`NOT EXISTS\` has no such trap, and optimisers handle it well.`,
+    hint: "JOIN accounts for the deposit sum, GROUP BY the customer, and add WHERE NOT EXISTS (SELECT 1 FROM loans l WHERE l.customer_id = c.customer_id). The subquery is correlated — it references the outer customer row.",
+    seedQuery: `SELECT c.customer_name,
+       c.segment,
+       c.credit_score,
+       ROUND(SUM(a.balance), 2) AS total_deposits
+  FROM customers c
+  JOIN accounts a ON a.customer_id = c.customer_id
+ WHERE NOT EXISTS (
+         SELECT 1
+           FROM loans l
+          WHERE
+       )
+ GROUP BY
+ ORDER BY total_deposits DESC`,
+    solutionQuery: `SELECT c.customer_name,
+       c.segment,
+       c.credit_score,
+       ROUND(SUM(a.balance), 2) AS total_deposits
+  FROM customers c
+  JOIN accounts a ON a.customer_id = c.customer_id
+ WHERE NOT EXISTS (
+         SELECT 1
+           FROM loans l
+          WHERE l.customer_id = c.customer_id
+       )
+ GROUP BY c.customer_id, c.customer_name, c.segment, c.credit_score
+ ORDER BY total_deposits DESC`,
+    epoch: 'Expert',
+    difficulty: 4,
+  },
 ];