Add explanations

CLAUDE.md

@@ -43,7 +43,9 @@ BNetzA ZIP ──[Stage 1: amateurfunk_fetch.py]──► data/<slug>/
                                                   └── manifest.json
 
 data/ ──[Stage 2: amateurfunk_anki.py]──► anki/
-                                            ├── amateurfunk-technische-kenntnisse.apkg
+                                            ├── amateurfunk-technische-kenntnisse-n.apkg
+                                            ├── amateurfunk-technische-kenntnisse-e.apkg
+                                            ├── amateurfunk-technische-kenntnisse-a.apkg
                                             ├── amateurfunk-betriebliche-kenntnisse.apkg
                                             └── amateurfunk-kenntnisse-von-vorschriften.apkg
 ```
@@ -66,14 +68,20 @@ data/ ──[Stage 2: amateurfunk_anki.py]──► anki/
 
 1. Read the latest edition from `data/` (following
    `manifest-latest.json` to a per-edition directory).
-2. Split the catalog into three categories — one per top-level
+2. Split the catalog into five categories. Betriebliche and
-   Prüfungsteil (Technische / Betriebliche / Vorschriften). The
+   Vorschriften get one deck each (shared across every candidate).
-   license-class axis is mapped into tags, not separate packages.
+   Technische is additionally fanned out per license class into three
+   decks (N / E / A) using a strict equality split on the question's
+   `class` field. The `klasse-N|E|A` tag is still emitted on every
+   note for inside-Anki filtering.
 3. Render every question as an Anki note: shuffled A/B/C/D choices on
    the front, the displayed position of the correct answer on the
    back. Inline `$...$` LaTeX is converted to MathJax `\(...\)`
    delimiters; the catalog's safe inline markup (`<u>...</u>`) is
-   preserved.
+   preserved. If the question's number has an entry in
+   `explanations.json` (see EXPLANATIONS.md), an English explanation
+   block is appended to the back; a "low confidence" badge shows for
+   entries with `confidence < 7`.
 4. Hand-roll the v11 Anki collection (SQLite + JSON config) and
    package it as a `.apkg` ZIP with deterministic timestamps. Same
    input → byte-identical output across runs.
@@ -98,10 +106,16 @@ SVG dark-mode handling, schema choices) live in `DESIGN.md` §7.
 
 ## Working on this repo
 
+- `EXPLANATIONS.md` — the editorial contract for agents asked to add
+  or improve per-question explanations. The schema, the workflows
+  ("explain everything unexplained", "improve everything below
+  confidence 7"), and the source/confidence guidance live there.
+  `explanations.json` is an empty `{}` until agents populate it.
 - Start from `DESIGN.md` — it has the JSON schema, the question/answer
   conventions (answer A is always correct upstream → consumers shuffle
   before display), the LaTeX-in-questions caveat, the exam-structure
-  rationale for the three Anki packages, and per-stage design notes.
+  rationale for the Anki package layout (Betriebliche, Vorschriften,
+  and Technische split per license class), and per-stage design notes.
 - Do not invent new download URLs; the ones in `DESIGN.md` were
   verified against the live BNetzA site.
 - When BNetzA publishes a new edition, expect a new

DESIGN.md

@@ -491,28 +491,43 @@ output artifacts.
 
 ```
 anki/
-  amateurfunk-technische-kenntnisse.apkg          (~2.5 MB, 1374 cards)
+  amateurfunk-technische-kenntnisse-n.apkg        (195 cards)
-  amateurfunk-betriebliche-kenntnisse.apkg        (~130 KB,  172 cards)
+  amateurfunk-technische-kenntnisse-e.apkg        (463 cards)
-  amateurfunk-kenntnisse-von-vorschriften.apkg    (~75 KB,   204 cards)
+  amateurfunk-technische-kenntnisse-a.apkg        (716 cards)
+  amateurfunk-betriebliche-kenntnisse.apkg        (172 cards)
+  amateurfunk-kenntnisse-von-vorschriften.apkg    (204 cards)
 ```
 
-One `.apkg` per top-level Prüfungsteil. The card counts match the
+Five `.apkg` files. Betriebliche and Vorschriften are shared across
-study pools from §3 axis 1: Technische = 1374 (cumulative N+E+A);
+every candidate (class-1-only in the data per §3 axis 2) and stay as
-Betriebliche = 172 (shared across all candidates); Vorschriften = 204
+one deck each. Technische is fanned out per license class using a
-(also shared). License class is conveyed via tags
+**strict equality split** on the question's `class` field — class-1
-(`klasse-N` / `klasse-E` / `klasse-A`), not separate packages —
+questions land in the N deck only, class-2 in E only, class-3 in A
-Anki users browse decks by topic and filter by class.
+only. The card counts therefore equal the new-at-this-class slices
+from §3, not the cumulative study pools: a candidate studying for
+class E imports Technische-N + Technische-E + Betriebliche +
+Vorschriften.
+
+The Technische deck names use Anki's `::` hierarchy separator
+(`Amateurfunk::Technische Kenntnisse::N`) so the three decks render
+as children of a shared parent in Anki's deck browser. The
+`klasse-N` / `klasse-E` / `klasse-A` tag is still emitted on every
+note — redundant within each Technische deck but useful in Betr/Vor
+for in-Anki filtering, and harmless besides.
 
 ### Steps
 
 1. **Load** the latest catalog: follow `manifest-latest.json` to a
    per-edition directory, read the catalog JSON, the per-edition
    `manifest.json`, and index the `svgs/` folder.
-2. **Categorize** the question tree into three `Category` objects,
+2. **Categorize** the question tree into five `Category` objects.
-   one per top-level Prüfungsteil. Each category carries the flat
+   Betriebliche and Vorschriften get one each. Technische is
-   list of every question that lives anywhere under it, along with
+   additionally split into three sub-categories (one per license
-   each question's path through the section tree (used for the card
+   class) via a strict equality match on the question's `class`
-   breadcrumb and the path tags).
+   field. Each category carries the flat list of every question that
+   lives anywhere under it, along with each question's path through
+   the section tree (used for the card breadcrumb and the path
+   tags).
 3. **For each category, render every question** as an Anki note:
    - Compute a stable per-question seed from `--seed` + question
      number; shuffle the A/B/C/D choices using it. The displayed
@@ -594,13 +609,19 @@ byte-identical sha256 on each `.apkg`. Verified during review.
 - **Slug for filenames.** `slugify()` maps German umlauts and ß
   to ASCII digraphs before NFKD-normalizing the rest, then keeps
   only `[a-z0-9-]`. Yields readable, sortable filenames
-  (`amateurfunk-technische-kenntnisse.apkg`).
+  (`amateurfunk-technische-kenntnisse-n.apkg`). The three
+  Technische decks append the class letter (`-n`, `-e`, `-a`)
+  to the shared base slug; Betriebliche and Vorschriften use the
+  base slug as-is.
 - **Tags.** Each note carries `klasse-N|E|A` plus
   `pfad-<slugified-section>` for every section level below the
-  top-level Prüfungsteil. Per-question number tags are
+  top-level Prüfungsteil. The `klasse-*` tag is redundant within
-  deliberately *not* emitted — they would create ~1750 singletons
+  the three Technische decks (every card already shares one class)
-  in Anki's tag tree, and the number is already in the dedicated
+  but is still emitted there for uniformity, and remains the
-  `Number` field for search.
+  primary filter axis inside the Betriebliche/Vorschriften decks.
+  Per-question number tags are deliberately *not* emitted — they
+  would create ~1750 singletons in Anki's tag tree, and the number
+  is already in the dedicated `Number` field for search.
 - **Breadcrumb consistency.** The visible card breadcrumb and the
   stored `Path` field both go through one `display_path()`
   helper, so they never drift. The boilerplate
@@ -611,6 +632,22 @@ byte-identical sha256 on each `.apkg`. Verified during review.
   the card background. We inject a white `<rect>` as the first
   painted element when packaging the SVG into the `.apkg`. The
   on-disk extracted files are left untouched.
+- **Explanations layer (optional).** `amateurfunk_anki.py` loads
+  `explanations.json` (CLI: `--explanations`) and, for any
+  question number found there, appends an English explanation
+  block to the back of the card. The block is styled distinctly
+  (serif italic body, sans-serif metadata, top border) and shows
+  a small "low confidence" badge when the entry's `confidence`
+  field is below `LOW_CONFIDENCE_THRESHOLD` (= 7). `revision` and
+  the raw `confidence` number are editorial-only and never
+  displayed. A missing file is treated as an empty database; a
+  malformed entry is a hard `AnkiBuildError`. The editorial
+  contract (schema, sourcing, agent workflows) lives in
+  `EXPLANATIONS.md`; this file documents only the wiring on the
+  build side. Note-GUID stability across explanation edits: the
+  GUID is keyed on `category.slug:number`, not on field content,
+  so adding/editing an entry updates the existing Anki note on
+  re-import rather than producing a duplicate.
 
 ### Testing focus
 
@@ -618,8 +655,11 @@ Network-free, fixture-driven. Real Anki is NOT required to
 verify the output — tests open the `.apkg` ZIP directly, parse
 the embedded SQLite, and check structural invariants.
 
-- Top-level structure: one `.apkg` per Prüfungsteil; expected
+- Top-level structure: five `.apkg` files (Betriebliche +
-  filename slugs.
+  Vorschriften + Technische×{N,E,A}); expected filename slugs.
+- Technische class partition: each Technische deck contains
+  exactly the questions whose `class` field matches that letter,
+  and no others.
 - Note count and field shape per category.
 - Tag mapping: a class-1 question carries `klasse-N`, no
   `nummer-*` tags.
@@ -643,6 +683,13 @@ the embedded SQLite, and check structural invariants.
 - Deterministic build: two consecutive builds against the same
   catalog with a fixed `--epoch` produce byte-identical `.apkg`
   files.
+- Explanations layer: a present entry produces a styled
+  explanation block on the back card; a missing entry leaves the
+  card unchanged; `confidence < 7` surfaces the "low confidence"
+  badge; URL sources render as `<a>`, citation sources render as
+  escaped text; a missing `explanations.json` is silently empty;
+  a malformed entry (bad type, out-of-range confidence, missing
+  field) raises `AnkiBuildError`.
 
 ### What we deliberately do NOT do in Stage 2
 

EXPLANATIONS.md

@@ -0,0 +1,425 @@
+# Explanations — agent-authored answer rationales
+
+This document is the working contract for any AI agent (or human)
+asked to **add or improve explanations** for questions in the BNetzA
+amateur-radio catalog. The decks built by `amateurfunk_anki.py`
+optionally append an English explanation block to the back of each
+card; the text of those explanations lives in `explanations.json` at
+the repo root and is edited by hand (or by an agent following this
+file).
+
+The build is non-blocking on this data: a missing file or missing
+entry just produces a card without an explanation block. Adding an
+entry is purely additive — no regenerate ceremony beyond
+`make anki`.
+
+---
+
+## 1. File location and shape
+
+- **Path:** `explanations.json` at the repo root. Tracked in git.
+- **Encoding:** UTF-8, no BOM, two-space indent for readability.
+- **Top level:** a JSON object keyed by **question number** exactly
+  as it appears in the catalog (e.g. `"NA101"`, `"BA205"`,
+  `"EH410"`). Keys are case-sensitive.
+- **Sort order:** keep keys in alphabetical order. Diffs stay clean
+  and merge conflicts get easier; the build does not care.
+
+### Per-entry schema
+
+Every entry MUST have exactly these four fields:
+
+| Field         | Type    | Constraint                                 |
+|---------------|---------|--------------------------------------------|
+| `revision`    | integer | `>= 1`. Starts at `1`, bumps on improvement |
+| `explanation` | string  | Non-empty. **English.** Terse. WHY-focused |
+| `source`      | string  | Non-empty. URL or citation like `AFuV §16(2)` |
+| `confidence`  | integer | `1..10` inclusive. See scale in §5         |
+
+Extra keys are rejected by `load_explanations()` — the build fails
+with `unknown fields [...]` listing them. The loader is similarly
+strict about types: a JSON `true` will not satisfy the integer
+contract for `revision` or `confidence`. If you need to track
+editorial metadata that isn't shown on the card, propose a schema
+change rather than smuggling fields in.
+
+Top-level keys (the question numbers) must also match the catalog
+exactly. An entry keyed on a number that no live question carries
+(typo, stale ID after a catalog revision) fails the build with
+`explanation keys not present in the catalog: ...`. Fix the key, or
+remove the entry.
+
+### Minimal example
+
+```json
+{
+  "NA101": {
+    "revision": 1,
+    "explanation": "100 m weighs 210 g, so 55 g is 55/210 of 100 m ≈ 26.2 m. Mass scales linearly with length for the same wire.",
+    "source": "https://50ohm.de/lernen/wissen/elektrotechnik-mathematik",
+    "confidence": 8
+  }
+}
+```
+
+The build appends this on the back card as a styled block headed
+**Explanation**, with the source rendered as a clickable link when
+it looks like an HTTP(S) URL.
+
+---
+
+## 2. How the explanation reaches the card
+
+`amateurfunk_anki.py` does the wiring; the agent does not need to
+modify Python code to ship a new explanation. End-to-end:
+
+1. `load_explanations(./explanations.json)` parses the file and
+   validates each entry against the schema above. A malformed entry
+   is a hard build error.
+2. While rendering a card, `render_question()` looks up the
+   question's `number` in the dict. On a hit, `render_explanation()`
+   produces an HTML block: header "Explanation", the body (with
+   inline `$...$` LaTeX rewritten to MathJax `\(...\)`, same as the
+   question text), and a "Source: ..." line.
+3. The block lands inside `.af-back` at the very end, styled by the
+   `.af-explanation*` CSS rules — serif italic body, sans-serif
+   metadata, separated from the answer by a top border. When the
+   entry's `confidence` is **below 7**, a small "low confidence"
+   badge appears next to the **Explanation** header — a hint to the
+   learner that the reasoning may be incomplete or weakly sourced.
+   `revision` is never shown on the card.
+4. The note GUID is keyed on `category.slug:number`, NOT on field
+   content, so adding/changing an explanation does **not** create
+   duplicate cards on re-import — Anki updates the existing note in
+   place.
+
+Run `make anki` (or `python3 amateurfunk_anki.py`) after editing the
+file. The CLI summary line will show `... N with explanations` per
+deck.
+
+---
+
+## 3. Locating a question to explain
+
+Questions live in the JSON catalog under the per-edition data
+directory:
+
+```
+data/<slug>/fragenkatalog<edition>.json
+```
+
+To find one by number, e.g. `EH410`:
+
+```sh
+python3 -c '
+import json, glob
+for path in glob.glob("data/*/fragenkatalog*.json"):
+    cat = json.load(open(path))
+    def walk(node, prefix=()):
+        for q in node.get("questions", []) or []:
+            if q["number"] == "EH410":
+                print(json.dumps(q, ensure_ascii=False, indent=2))
+                print(" / ".join(prefix))
+        for s in node.get("sections", []) or []:
+            walk(s, prefix + (s.get("title", "?"),))
+    walk(cat)
+'
+```
+
+You get the question stem, the four answers (`answer_a` is the
+**correct** one upstream; the deck shuffles before display), the
+class digit (`"1"` = N, `"2"` = E, `"3"` = A), and the section path.
+That path tells you the topic context — useful when choosing a
+source citation.
+
+---
+
+## 4. Writing the explanation
+
+### Style
+
+- **Language: English.** The questions and answers stay in German
+  on the card; the explanation is the one English element. Don't
+  switch.
+- **Terse.** Aim for 1–3 sentences. The card already shows the
+  question, the correct answer text, and the breadcrumb — the
+  explanation should add the missing *why*, not restate any of
+  those.
+- **Lead with the reasoning.** Don't open with "The correct answer
+  is X because..." — the card already declares the correct answer
+  one line above. Go straight to the principle, the formula, the
+  rule of thumb.
+- **Math.** Use inline `$...$`; it's rewritten to MathJax on the
+  card, same as for question text. Don't use display math `$$...$$`
+  (the deck has no MathJax support for those).
+- **Inline emphasis.** `<u>...</u>` is the only HTML you should
+  type by hand. Everything else gets HTML-escaped.
+- **No spoilers about other answers.** If a distractor is a common
+  trap, naming the trap is fine; quoting the distractor's text is
+  noise.
+
+### What "WHY-focused" means in practice
+
+| Less useful                                  | More useful                                   |
+|----------------------------------------------|-----------------------------------------------|
+| "The correct answer is 26.2 m."              | "55 g is 55/210 of 100 m by mass scaling."    |
+| "Because regulations require it."            | "AFuV §16(2): only class A may operate ≤10 m." |
+| "It's how the formula works."                | "Q is reactance over resistance, so Q rises as R falls." |
+
+If the *only* honest explanation is "memorize the table" (e.g. a
+band-plan lookup), say that plainly and cite the band plan in
+`source` — confidence stays low (3–4) until someone finds a deeper
+hook.
+
+---
+
+## 5. Confidence scale
+
+| Score | Meaning                                                                |
+|------:|------------------------------------------------------------------------|
+| 10    | Direct quote / arithmetic restatement from a primary legal source.     |
+|  8–9  | Derivation from a well-known formula or law text; no ambiguity.        |
+|  6–7  | Reasoning sound but condensed; a reviewer might want one more line.    |
+|  4–5  | Educated guess based on context; corroborating source is weak.         |
+|  1–3  | Best-effort placeholder. Flag for re-explanation.                      |
+
+When in doubt, score lower. The query "explain everything below
+confidence 7" is a real workflow — a too-generous score hides work
+that should be redone.
+
+The score is also capped by the source tier — see §6 "Source ↔
+confidence linkage". A tier-8 (general web) source caps confidence
+at 5 regardless of how persuasive the prose is.
+
+**Card surface.** Confidence below 7 also surfaces a "low
+confidence" badge on the card itself (see §2). This is deliberate:
+the learner sees that the explanation is provisional rather than
+trusting it as definitive. Scoring 7 turns the badge off, so do
+not nudge a 6 to a 7 just to clear the badge — fix the explanation
+or the source first.
+
+---
+
+## 6. Sources
+
+Use the **strictest tier that actually answers the question**. Drop
+down a tier only when the higher one doesn't cover the topic — never
+because a lower-tier source is easier to find. The exam tests
+knowledge of German law and BNetzA-curated material; a primary-law
+question backed by a random tutorial site is a worse explanation
+than no explanation at all.
+
+The full priority order, top (most preferred) to bottom:
+
+1. **Primary German law.**
+   - AFuG (Amateurfunkgesetz), AFuV (Amateurfunkverordnung), BEMFV,
+     TKG when applicable, plus frequency-allocation ordinances
+     (vfg / VVnömL).
+   - Canonical URL pattern: `https://www.gesetze-im-internet.de/<gesetz>/__<§>.html`.
+     Cite as `AFuV §16(2)` style. Use the URL form when possible so
+     the card link is clickable.
+   - If a question's answer is fixed by German law, **no lower tier
+     is acceptable** as the sole source.
+
+2. **EU / international regulation binding on Germany.**
+   - CEPT recommendations (T/R 61-01, T/R 61-02), ITU Radio
+     Regulations, ECC decisions, EU EMC / RED directives.
+   - Use `docdb.cept.org/...` or the ITU document portal as the
+     canonical URL.
+   - Reach for this tier when the question references CEPT
+     licensing, foreign operation, or harmonised band plans that
+     AFuV doesn't restate.
+
+3. **BNetzA official publications.**
+   - The question catalog's own `README.txt`, BNetzA "Vfg"
+     announcements, the German band plan as published by BNetzA,
+     official explanatory notes.
+   - Useful when primary law is too terse to convey the *why* on
+     its own — these are the regulator's own gloss on the law.
+
+4. **German amateur-radio organisations.**
+   - DARC publications and band plan (`darc.de`), Runder Tisch
+     Amateurfunk (RTA) papers, **50ohm.de** (the community study
+     site linked from `README.md`).
+   - Exam-aligned by community convention; well-suited to the WHY
+     when law text alone isn't pedagogically clear.
+
+5. **International amateur-radio organisations.**
+   - IARU Region 1 documents (Germany is in R1) and DARC's IARU R1
+     liaison material; ARRL publications when they explain the
+     underlying physics or an internationally harmonised band plan.
+   - ARRL is **not** a source for German licensing rules — don't
+     use it that way.
+
+6. **Standards bodies and primary technical references.**
+   - NIST, IEEE, IEC, ITU-R recommendations, ETSI standards (when
+     not already binding via tier 2).
+   - Appropriate for purely technical questions — units, defining
+     equations, definitions — where no amateur-radio-specific
+     source applies.
+
+7. **Established engineering references.**
+   - University lecture notes hosted by the university, recognised
+     EE textbook publishers, well-known authors' personal sites.
+   - Use only when tiers 1–6 don't cover the topic.
+
+8. **General web technical sites — last resort.**
+   - Wikipedia, tutorial sites (`allaboutcircuits.com`,
+     `amateur-radio-wiki.net`, `biopac.com`, etc.), random
+     manufacturer pages, blog posts.
+   - **Treat citations at this tier as temporary.** Score
+     `confidence` ≤ 5 and consider the entry a §8.3 candidate for
+     re-sourcing the moment a higher-tier source is identified.
+     Never use a tier-8 source as the sole citation for a question
+     that has a clear answer in tiers 1–4.
+
+### Source ↔ confidence linkage
+
+The §5 confidence scale isn't independent of the tier:
+
+- Tier 1–2 + sound derivation → 9–10 is appropriate.
+- Tier 3–4 → cap at 8 unless the source quotes a primary citation.
+- Tier 5–6 → cap at 7.
+- Tier 7 → cap at 6.
+- Tier 8 (sole source) → cap at 5. Flag for re-sourcing.
+
+This isn't a hard limit the validator enforces — it's editorial
+discipline. If you find yourself wanting confidence 9 with a
+tier-7 source, you almost always either (a) need to find the
+tier-1–4 source that backs the same claim, or (b) are about to
+overstate certainty.
+
+### Form
+
+If the source is a URL, store it as one — `_source_html()` makes
+plain `http://` / `https://` strings clickable. Mixed text+URL
+sources are stored verbatim and rendered as escaped text (no auto
+link), so prefer one or the other. A short citation like
+`AFuV §16(2), gesetze-im-internet.de/afuv_2005/__16.html` is
+acceptable but loses click-through. When possible, narrow the
+source to a single canonical URL.
+
+---
+
+## 7. Revisions
+
+`revision` is editorial bookkeeping; it never appears on the card.
+
+- A brand-new entry starts at `revision: 1`.
+- **Bump** when you materially improve the explanation, broaden the
+  source, or fix an error. A typo fix is not a revision bump.
+- Don't lower the revision. If a previous revision was wrong, fix
+  it forward and explain the fix in the commit message.
+- A revision bump usually pairs with a confidence bump (or, more
+  rarely, with a deliberate confidence reset to indicate the
+  reviewer is less sure than the prior author was).
+
+---
+
+## 8. Workflows
+
+These are the common ways the file gets edited. The agent should
+read the request and pick one.
+
+### 8.1 "Explain question X"
+
+1. Look up X in the catalog (§3) — read the stem, the correct
+   answer (`answer_a` upstream), and the section path.
+2. Identify *why* the correct answer is correct. Reach for a
+   primary source before paraphrasing 50ohm.de.
+3. Draft the body per §4, pick a source per §6, score honestly
+   per §5.
+4. Open `explanations.json`, insert the new entry in **alphabetical
+   key order**, save.
+5. Run `python3 -m unittest test_amateurfunk_anki` — the schema
+   validator runs as part of the build path used by tests, so a
+   typo'd entry fails fast.
+6. Rebuild decks (`make anki`) and spot-check the new entry on
+   the back of the card.
+
+### 8.2 "Explain every question that doesn't have an entry yet"
+
+This is bulk work — handle it in chunks, not as one giant batch.
+
+1. Compute the unexplained set:
+
+   ```sh
+   python3 -c '
+   import json, glob
+   catalog = next(iter(glob.glob("data/*/fragenkatalog*.json")))
+   exp = json.load(open("explanations.json"))
+   def walk(n):
+       for q in n.get("questions", []) or []: yield q["number"]
+       for s in n.get("sections", []) or []: yield from walk(s)
+   missing = sorted(set(walk(json.load(open(catalog)))) - exp.keys())
+   print(len(missing), "questions unexplained")
+   print("\n".join(missing[:20]))
+   '
+   ```
+
+2. Process in groups of ~10 within one topic. Same section path
+   often shares one source citation, so context stays warm and
+   confidence scores stay consistent.
+3. After each group, save and run the tests. Don't accumulate
+   hundreds of unsaved entries.
+
+### 8.3 "Improve all entries with confidence below N"
+
+1. List low-confidence entries:
+
+   ```sh
+   python3 -c '
+   import json
+   exp = json.load(open("explanations.json"))
+   for k, v in sorted(exp.items()):
+       if v["confidence"] < 7: print(k, v["confidence"], "—", v["source"])
+   '
+   ```
+
+2. For each, look up the question (§3), seek a stronger source
+   (§6), rewrite if the new source changes the framing.
+3. **Bump `revision`** and update `confidence` to reflect the new
+   evidence. Do not just bump confidence without changing anything
+   else — that's how bad explanations entrench themselves.
+
+### 8.4 "This explanation is wrong"
+
+1. Read the current entry. Confirm the error against the catalog
+   stem and the cited source.
+2. Rewrite the body, replace the source if needed, bump
+   `revision`, and re-score `confidence` from scratch (don't carry
+   over the prior score).
+3. Commit message should name the question number and summarize
+   what was wrong — future agents need to know what kind of
+   mistake to look for elsewhere.
+
+---
+
+## 9. Validation checklist
+
+Before saving, confirm:
+
+- [ ] Key is the exact catalog `number` (case-sensitive).
+- [ ] All four fields present, correct types, in range.
+- [ ] `explanation` is English, terse, WHY-focused.
+- [ ] `source` is non-empty and as primary as possible.
+- [ ] `revision` is correct (1 for new, bumped for improvement).
+- [ ] `confidence` honestly reflects how well the source supports
+      the explanation.
+- [ ] Keys remain in alphabetical order in the file.
+- [ ] `python3 -m unittest test_amateurfunk_anki` passes.
+
+---
+
+## 10. What's out of scope here
+
+- **Translating questions into English.** The cards are bilingual
+  by design: German question + English explanation. Don't
+  paraphrase the German.
+- **Editing the catalog itself.** `data/` is a build artifact —
+  refreshed by `amateurfunk_fetch.py` from BNetzA. Mistakes in the
+  upstream questions are upstream's to fix.
+- **Removing explanations.** If an entry is genuinely useless, fix
+  it forward (§8.4). Don't delete keys — that loses the audit
+  trail. (The schema doesn't currently encode "retracted", so if
+  you ever need it, propose a schema change.)

README.md

@@ -13,9 +13,12 @@ make test      # run both test suites
 make clean     # remove data/ and anki/
 ```
 
-Output: three `.apkg` files under `anki/`, one per Prüfungsteil
+Output: five `.apkg` files under `anki/`. Betriebliche and
-(Technische / Betriebliche / Vorschriften). Import them into Anki —
+Vorschriften get one deck each (shared across all license classes);
-re-importing a newer build preserves your review history.
+Technische is split per class into three decks (N / E / A) following
+the catalog's `class` field. A class-A candidate who wants every
+Technische question imports all three Technische decks. Re-importing
+a newer build preserves your review history.
 
 ## Exam question source
 
@@ -47,11 +50,27 @@ question catalog the decks built here are based on.
 
 Python 3.11+, standard library only. No third-party dependencies.
 
+## Explanations
+
+The back of each card optionally carries a terse English explanation
+of *why* the right answer is right. Explanations are not part of the
+BNetzA catalog — they're authored separately (by humans or AI agents)
+into `explanations.json` at the repo root. The build is non-blocking
+on this: questions without an entry just show no explanation block.
+Entries with `confidence < 7` render a small "low confidence" badge
+so learners know the reasoning is provisional.
+
+`EXPLANATIONS.md` is the editorial contract: schema, sourcing
+guidance, confidence scale, and the workflows an AI agent should
+follow when asked to add or improve entries.
+
 ## More
 
 - `CLAUDE.md` — project orientation, pipeline overview.
 - `DESIGN.md` — source-discovery notes, JSON schema, per-stage
   design contracts.
+- `EXPLANATIONS.md` — schema + workflows for the explanations
+  database.
 
 ## License
 

amateurfunk_anki.py

@@ -7,10 +7,10 @@ script does:
   1. Read the per-edition directory written by `amateurfunk_fetch.py`
      (the JSON catalog, the `svgs/` folder, and the per-edition
      `manifest.json`).
-  2. Split the catalog into three categories — one per top-level
+  2. Split the catalog into five categories. Betriebliche and
-     Prüfungsteil: Technische / Betriebliche / Vorschriften. (The
+     Vorschriften get one each. Technische is additionally fanned
-     license-class axis is mapped into tags, not into separate
+     out per license class into three decks (N / E / A) via a
-     packages.)
+     strict equality split on the question's `class` field.
   3. For each category, render every question as an Anki note: front
      shows the (shuffled) A/B/C/D choices, back names the displayed
      position of the correct answer.
@@ -20,8 +20,8 @@ script does:
 
 The script is intentionally a single file with stdlib only. Readability
 beats cleverness here — most of the bytes below are docstrings and
-comments. Performance is a non-goal (we build three small decks once
+comments. Performance is a non-goal (we build a handful of small
-per upstream edition).
+decks once per upstream edition).
 """
 
 import argparse
@@ -52,6 +52,23 @@ DEFAULT_DATA_DIR = Path("data")
 # Default destination for the generated `.apkg` files.
 DEFAULT_OUT_DIR = Path("anki")
 
+# Default location of the per-question explanations database. The
+# file is editorial content (tracked in git), not a build artifact.
+# A missing file is treated as "no explanations available" — the
+# decks still build cleanly.
+DEFAULT_EXPLANATIONS_PATH = Path("explanations.json")
+
+# Required fields on every entry in the explanations database. See
+# EXPLANATIONS.md for the full editorial contract.
+EXPLANATION_FIELDS = ("revision", "explanation", "source", "confidence")
+
+# Confidence values strictly below this threshold are surfaced on
+# the card as a "low confidence" badge — a learner-facing warning
+# that the explanation may be incomplete or weakly sourced. Matches
+# the editorial cutoff used in EXPLANATIONS.md §5/§8.3 ("everything
+# below 7 is review-worthy").
+LOW_CONFIDENCE_THRESHOLD = 7
+
 # Exit codes. The builder is much simpler than the fetcher — there is
 # no "operator must resolve local state" case here, so two codes are
 # enough.
@@ -75,6 +92,16 @@ TOP_LEVEL_PREFIX = "Prüfungsfragen im Prüfungsteil: "
 # search for.
 CLASS_TAGS = {"1": "N", "2": "E", "3": "A"}
 
+# Ordered class digits used when fanning Technische Kenntnisse out
+# into one deck per license class. Order controls the on-disk file
+# ordering and the Anki deck-tree ordering.
+CLASS_ORDER = [("1", "N"), ("2", "E"), ("3", "A")]
+
+# Short title of the top-level Prüfungsteil that gets fanned out
+# per license class. The other two (Betriebliche, Vorschriften) are
+# shared across all candidates and stay as a single deck each.
+TECHNISCHE_SHORT_TITLE = "Technische Kenntnisse"
+
 # Fallback build epoch if neither the manifest nor `--epoch` supplies
 # one. Picked as 0 so missing-metadata builds are still deterministic
 # and obviously wrong (timestamps would all show 1970).
@@ -158,18 +185,139 @@ def load_latest_catalog(data_dir):
     return edition_dir, manifest, catalog
 
 
+# ============================================================================
+# Explanations database
+# ============================================================================
+
+
+def load_explanations(path):
+    """Return the per-question explanations dict from a JSON file.
+
+    A missing file is treated as an empty database — the build still
+    runs and simply doesn't append an explanation block to any card.
+    A malformed file (bad JSON, wrong shape, missing fields, out of
+    range confidence/revision) is a hard error so editorial mistakes
+    don't silently ship.
+
+    Expected on-disk shape (full schema lives in `EXPLANATIONS.md`):
+
+        {
+            "NA101": {
+                "revision": 1,
+                "explanation": "Terse English text explaining why ...",
+                "source": "https://... or AFuV §16(2)",
+                "confidence": 7
+            },
+            ...
+        }
+    """
+    if path is None or not path.exists():
+        return {}
+    try:
+        raw = json.loads(path.read_text("utf-8"))
+    except (OSError, json.JSONDecodeError) as e:
+        raise AnkiBuildError(
+            f"could not read explanations file {path}: {e}"
+        ) from e
+    if not isinstance(raw, dict):
+        raise AnkiBuildError(
+            f"explanations file {path} must contain a JSON object at the top level"
+        )
+    cleaned = {}
+    for key, value in raw.items():
+        _validate_explanation(key, value)
+        cleaned[str(key)] = value
+    return cleaned
+
+
+def _validate_explanation(key, value):
+    """Raise `AnkiBuildError` if one explanation entry is malformed.
+
+    Strict shape check: the four documented fields are required, no
+    extras allowed, and integer fields reject `bool` (which would
+    pass `isinstance(..., int)` because `bool` subclasses `int` in
+    Python). Extras are rejected so a stray `"note"` or `"author"`
+    field can't accumulate silently and drift from the documented
+    schema in EXPLANATIONS.md.
+    """
+    if not isinstance(value, dict):
+        raise AnkiBuildError(
+            f"explanation {key!r} must be a JSON object"
+        )
+    missing = [f for f in EXPLANATION_FIELDS if f not in value]
+    if missing:
+        raise AnkiBuildError(
+            f"explanation {key!r} missing required fields: {missing}"
+        )
+    extra = sorted(set(value) - set(EXPLANATION_FIELDS))
+    if extra:
+        raise AnkiBuildError(
+            f"explanation {key!r}: unknown fields {extra}"
+        )
+    if type(value["revision"]) is not int or value["revision"] < 1:
+        raise AnkiBuildError(
+            f"explanation {key!r}: revision must be a positive integer"
+        )
+    if (
+        type(value["confidence"]) is not int
+        or not 1 <= value["confidence"] <= 10
+    ):
+        raise AnkiBuildError(
+            f"explanation {key!r}: confidence must be an integer in 1..10"
+        )
+    if not isinstance(value["explanation"], str) or not value["explanation"].strip():
+        raise AnkiBuildError(
+            f"explanation {key!r}: explanation must be a non-empty string"
+        )
+    if not isinstance(value["source"], str) or not value["source"].strip():
+        raise AnkiBuildError(
+            f"explanation {key!r}: source must be a non-empty string"
+        )
+
+
+def _check_explanation_keys_against_catalog(explanations, categories):
+    """Raise `AnkiBuildError` if any explanation key has no matching question.
+
+    EXPLANATIONS.md makes the catalog `number` part of the editorial
+    contract. A typo like `"NA10I"` for `"NA101"` would otherwise
+    pass schema validation and silently never appear on any card.
+    We fail the build instead, listing the unknown keys so the
+    editor can fix them.
+    """
+    if not explanations:
+        return
+    known = set()
+    for category in categories:
+        for item in category.questions:
+            number = item.question.get("number")
+            if number is not None:
+                known.add(str(number))
+    unknown = sorted(set(explanations) - known)
+    if unknown:
+        raise AnkiBuildError(
+            "explanation keys not present in the catalog: "
+            + ", ".join(unknown)
+        )
+
+
 # ============================================================================
 # Category collection
 # ============================================================================
 
 
 def collect_categories(catalog):
-    """Split the catalog into one `Category` per top-level section.
+    """Split the catalog into one `Category` per output `.apkg`.
 
-    Per DESIGN.md §3 axis 1, the catalog's three top-level sections
+    Per DESIGN.md §3 axis 1 the catalog has three top-level sections
-    are exactly the three Prüfungsteile. We walk each subtree and
+    (the three Prüfungsteile). Technische Kenntnisse is additionally
-    flatten every question it contains, recording the section path so
+    fanned out per license class (N/E/A) — see DESIGN.md §7 "Output
-    cards can show a breadcrumb and we can emit path-shaped tags.
+    layout". Betriebliche and Vorschriften are class-1-only in the
+    data but apply to every candidate, so they stay as one category
+    each.
+
+    For every category we flatten the subtree into a list of
+    `QuestionItem`s carrying the section path, used downstream for
+    the breadcrumb and the `pfad-*` tags.
     """
     sections = catalog.get("sections")
     if not isinstance(sections, list):
@@ -181,6 +329,9 @@ def collect_categories(catalog):
         questions = []
         _collect_questions(section, (title,), questions)
         short_title = _short_category_title(title)
+        if short_title == TECHNISCHE_SHORT_TITLE:
+            categories.extend(_split_by_class(title, short_title, questions))
+        else:
             categories.append(
                 Category(
                     title=title,
@@ -192,6 +343,33 @@ def collect_categories(catalog):
     return categories
 
 
+def _split_by_class(title, short_title, questions):
+    """Fan one Prüfungsteil out into one `Category` per license class.
+
+    Strict equality split on the question's `class` field: a class-1
+    question lands in the N deck only, class-2 in E only, class-3 in
+    A only. This matches the catalog's `class` field as written, not
+    the cumulative study pool (a class-A candidate who wants every
+    Technische question imports all three decks).
+
+    The short title uses Anki's `::` deck-hierarchy separator so the
+    three decks render as children of a shared parent in Anki's deck
+    browser (e.g. `Amateurfunk::Technische Kenntnisse::N`).
+    """
+    base_slug = slugify(short_title)
+    for digit, letter in CLASS_ORDER:
+        subset = [
+            item for item in questions
+            if str(item.question.get("class")) == digit
+        ]
+        yield Category(
+            title=f"{title} — Klasse {letter}",
+            short_title=f"{short_title}::{letter}",
+            slug=f"{base_slug}-{letter.lower()}",
+            questions=subset,
+        )
+
+
 def _collect_questions(section, path, out):
     """Recurse into a section node and append every question to `out`.
 
@@ -337,18 +515,22 @@ def randomized_answers(question, seed):
 # ============================================================================
 
 
-def render_question(item, media, seed):
+def render_question(item, media, seed, explanations=None):
     """Render one question as `(front_html, back_html, correct_label)`.
 
     The front shows the question stem, an optional figure, and the
     four shuffled answer choices as an ordered list with `type="A"`.
     The back names the displayed position of the correct answer and
-    repeats its text/figure.
+    repeats its text/figure, and — when the question's number has an
+    entry in `explanations` — appends a styled English explanation
+    block at the very end of the back card.
 
     `media` is the `MediaRegistry` for this category; it records
     which figure files are actually used so the packager can include
-    only those.
+    only those. `explanations` is the dict returned by
+    `load_explanations()`; `None` is treated as "no explanations".
     """
+    explanations = explanations or {}
     question = item.question
     choices, correct_label, correct = randomized_answers(question, seed)
     number = html.escape(str(question.get("number", "")))
@@ -390,10 +572,54 @@ def render_question(item, media, seed):
         back_parts.append(
             f'<div class="af-image af-correct-image">{correct_image}</div>'
         )
+    explanation = explanations.get(str(question.get("number", "")))
+    if explanation:
+        back_parts.append(render_explanation(explanation))
     back_parts.append("</div>")
     return "".join(front_parts), "".join(back_parts), correct_label
 
 
+def render_explanation(explanation):
+    """Render one explanation entry as an HTML block for the card back.
+
+    Inline `$...$` LaTeX in the body is rewritten to MathJax via
+    `text_html()`, same as the rest of the card. Sources that look
+    like an HTTP(S) URL become a clickable link; anything else (a
+    citation like "AFuV §16(2)") is rendered as plain text. The
+    `revision` number stays editorial-only and is never displayed;
+    `confidence` is also normally hidden, but surfaces as a small
+    "low confidence" badge in the header when it falls below
+    `LOW_CONFIDENCE_THRESHOLD` — a hint to the learner that this
+    explanation needs more work and to a reviewer that it's a
+    candidate for §8.3 in EXPLANATIONS.md.
+    """
+    body = text_html(explanation["explanation"])
+    source_html = _source_html(explanation["source"])
+    badge = ""
+    if explanation["confidence"] < LOW_CONFIDENCE_THRESHOLD:
+        badge = (
+            '<span class="af-explanation-low-confidence" '
+            'title="This explanation is weakly sourced or incomplete">'
+            'low confidence</span>'
+        )
+    return (
+        '<div class="af-explanation">'
+        f'<div class="af-explanation-header">Explanation{badge}</div>'
+        f'<div class="af-explanation-body">{body}</div>'
+        f'<div class="af-explanation-source">Source: {source_html}</div>'
+        '</div>'
+    )
+
+
+def _source_html(source):
+    """Return HTML for a `source` field: clickable link iff plain URL."""
+    text = source.strip()
+    if re.match(r"^https?://\S+$", text):
+        escaped = html.escape(text, quote=True)
+        return f'<a href="{escaped}">{escaped}</a>'
+    return html.escape(text)
+
+
 def display_path(path):
     """Return the user-facing section path with the top prefix stripped.
 
@@ -589,14 +815,20 @@ def tags_for_item(item):
 # ============================================================================
 
 
-def build_apkg_for_category(category, edition_dir, out_path, seed, build_epoch):
+def build_apkg_for_category(
+    category, edition_dir, out_path, seed, build_epoch, explanations=None,
+):
     """Write one category as an Anki `.apkg` file.
 
     Renders every question to a note, hand-rolls the v11 SQLite
     collection, and writes the final ZIP with deterministic
-    timestamps. Returns a small result dict describing what was
+    timestamps. `explanations` is the dict returned by
+    `load_explanations()` (or `None`); when a question's number is
+    present there, an English explanation block is appended to the
+    card back. Returns a small result dict describing what was
     written (for the CLI summary).
     """
+    explanations = explanations or {}
     tmp_dir = out_path.parent / f".{out_path.name}.tmp"
     db_path = tmp_dir / "collection.anki2"
 
@@ -616,10 +848,15 @@ def build_apkg_for_category(category, edition_dir, out_path, seed, build_epoch):
 
     try:
         notes = []
+        applied_explanations = 0
         for ordinal, item in enumerate(category.questions):
-            front, back, correct_label = render_question(item, media, seed)
+            front, back, correct_label = render_question(
+                item, media, seed, explanations,
+            )
             question = item.question
             number = str(question.get("number", f"q{ordinal}"))
+            if number in explanations:
+                applied_explanations += 1
             note_id = stable_id("note", f"{category.slug}:{number}")
             card_id = stable_id("card", f"{category.slug}:{number}")
             fields = [
@@ -662,6 +899,7 @@ def build_apkg_for_category(category, edition_dir, out_path, seed, build_epoch):
         "questions": len(category.questions),
         "media": len(media_paths),
         "missing_media": sorted(set(media.missing)),
+        "explanations": applied_explanations,
     }
 
 
@@ -1169,6 +1407,50 @@ CARD_CSS = """
   color: #0b6b3a;
   margin-bottom: 0.5rem;
 }
+.af-explanation {
+  margin-top: 1.5rem;
+  padding-top: 0.75rem;
+  border-top: 1px solid #ccc;
+  font-family: Georgia, "Times New Roman", serif;
+  font-size: 15px;
+  line-height: 1.5;
+  color: #333;
+  font-style: italic;
+}
+.af-explanation-header {
+  font-family: Arial, sans-serif;
+  font-style: normal;
+  font-weight: bold;
+  text-transform: uppercase;
+  letter-spacing: 0.05em;
+  font-size: 11px;
+  color: #777;
+  margin-bottom: 0.4rem;
+}
+.af-explanation-low-confidence {
+  display: inline-block;
+  margin-left: 0.5rem;
+  padding: 1px 6px;
+  border-radius: 3px;
+  background: #fff4d6;
+  color: #8a5a00;
+  font-weight: bold;
+  letter-spacing: 0.04em;
+  font-size: 10px;
+}
+.af-explanation-body {
+  margin-bottom: 0.6rem;
+}
+.af-explanation-source {
+  font-family: Arial, sans-serif;
+  font-style: normal;
+  font-size: 12px;
+  color: #666;
+}
+.af-explanation-source a {
+  color: #1a73e8;
+  text-decoration: none;
+}
 """
 
 
@@ -1199,18 +1481,24 @@ def build_epoch_from_manifest(manifest, override_epoch=None):
         ) from e
 
 
-def build_all(data_dir, out_dir, seed, override_epoch=None):
+def build_all(data_dir, out_dir, seed, override_epoch=None, explanations_path=None):
     """Build every category's `.apkg` and return their result dicts.
 
-    Loads the latest fetched catalog, picks a build epoch, then walks
+    Loads the latest fetched catalog and the explanations database,
-    the three top-level categories writing one `.apkg` each. Raises
+    picks a build epoch, then walks every category writing one
-    `AnkiBuildError` on configuration / catalog problems.
+    `.apkg` each. Raises `AnkiBuildError` on configuration / catalog
+    / explanation-schema problems.
     """
     edition_dir, manifest, catalog = load_latest_catalog(data_dir)
     build_epoch = build_epoch_from_manifest(manifest, override_epoch)
+    explanations = load_explanations(
+        explanations_path if explanations_path is not None
+        else DEFAULT_EXPLANATIONS_PATH
+    )
     categories = collect_categories(catalog)
     if not categories:
         raise AnkiBuildError("catalog has no categories")
+    _check_explanation_keys_against_catalog(explanations, categories)
 
     results = []
     for category in categories:
@@ -1222,6 +1510,7 @@ def build_all(data_dir, out_dir, seed, override_epoch=None):
                 out_path,
                 seed=seed,
                 build_epoch=build_epoch,
+                explanations=explanations,
             )
         )
     return results
@@ -1269,6 +1558,15 @@ def _parse_args(argv):
             "derived from manifest.json's fetched_at"
         ),
     )
+    parser.add_argument(
+        "--explanations",
+        type=Path,
+        default=DEFAULT_EXPLANATIONS_PATH,
+        help=(
+            "JSON file with per-question explanations (default: "
+            "./explanations.json; missing file is treated as empty)"
+        ),
+    )
     return parser.parse_args(argv)
 
 
@@ -1281,6 +1579,7 @@ def main(argv=None):
             args.out,
             seed=args.seed,
             override_epoch=args.epoch,
+            explanations_path=args.explanations,
         )
     except AnkiBuildError as e:
         print(f"error: {e}", file=sys.stderr)
@@ -1294,7 +1593,8 @@ def main(argv=None):
     for result in results:
         print(
             f"wrote {result['path']} "
-            f"({result['questions']} cards, {result['media']} media files)"
+            f"({result['questions']} cards, {result['media']} media files, "
+            f"{result['explanations']} with explanations)"
         )
         if result["missing_media"]:
             print(

explanations.json

@@ -0,0 +1 @@
+{}

test_amateurfunk_anki.py

@@ -32,6 +32,7 @@ def make_catalog():
                         "questions": [
                             question("NA101", "1", "Was ist A?", "NA101_q"),
                             question("EA202", "2", "Was ist E?", None),
+                            question("AA303", "3", "Was ist A-Klasse?", None),
                         ],
                     }
                 ],
@@ -114,23 +115,68 @@ class TestAnkiBuild(unittest.TestCase):
         self.root = Path(self.tmp.name)
         self.data_dir = make_fetched_data(self.root)
         self.out_dir = self.root / "anki"
+        # Hermetic explanations file: empty by default so tests don't
+        # pick up the real repo's explanations.json via the CLI
+        # default. Individual tests overwrite this file as needed.
+        self.explanations_path = self.root / "explanations.json"
+        self.explanations_path.write_text("{}", encoding="utf-8")
 
-    def test_builds_one_apkg_per_top_level_category(self):
+    def _build_all(self):
-        results = aa.build_all(self.data_dir, self.out_dir, seed="test-seed")
+        return aa.build_all(
+            self.data_dir,
+            self.out_dir,
+            seed="test-seed",
+            explanations_path=self.explanations_path,
+        )
+
+    def test_builds_one_apkg_per_category(self):
+        results = self._build_all()
         paths = sorted(path.name for path in self.out_dir.glob("*.apkg"))
         self.assertEqual(
             paths,
             [
                 "amateurfunk-betriebliche-kenntnisse.apkg",
                 "amateurfunk-kenntnisse-von-vorschriften.apkg",
-                "amateurfunk-technische-kenntnisse.apkg",
+                "amateurfunk-technische-kenntnisse-a.apkg",
+                "amateurfunk-technische-kenntnisse-e.apkg",
+                "amateurfunk-technische-kenntnisse-n.apkg",
             ],
         )
-        self.assertEqual([r["questions"] for r in results], [2, 1, 1])
+        by_deck = {r["deck"]: r["questions"] for r in results}
+        self.assertEqual(
+            by_deck,
+            {
+                "Amateurfunk::Technische Kenntnisse::N": 1,
+                "Amateurfunk::Technische Kenntnisse::E": 1,
+                "Amateurfunk::Technische Kenntnisse::A": 1,
+                "Amateurfunk::Betriebliche Kenntnisse": 1,
+                "Amateurfunk::Kenntnisse von Vorschriften": 1,
+            },
+        )
+
+    def test_technische_decks_partition_strictly_by_class_field(self):
+        self._build_all()
+        per_class_numbers = {}
+        for letter in ("n", "e", "a"):
+            apkg = self.out_dir / f"amateurfunk-technische-kenntnisse-{letter}.apkg"
+            db_path, _media, _names = extract_collection(apkg, self.root / letter)
+            conn = sqlite3.connect(db_path)
+            try:
+                numbers = [
+                    row[0].split(aa.FIELD_SEP, 1)[0]
+                    for row in conn.execute("select flds from notes")
+                ]
+            finally:
+                conn.close()
+            per_class_numbers[letter.upper()] = sorted(numbers)
+        self.assertEqual(
+            per_class_numbers,
+            {"N": ["NA101"], "E": ["EA202"], "A": ["AA303"]},
+        )
 
     def test_apkg_contains_notes_cards_and_media(self):
-        aa.build_all(self.data_dir, self.out_dir, seed="test-seed")
+        self._build_all()
-        apkg = self.out_dir / "amateurfunk-technische-kenntnisse.apkg"
+        apkg = self.out_dir / "amateurfunk-technische-kenntnisse-n.apkg"
         db_path, media, names = extract_collection(apkg, self.root)
 
         self.assertIn("collection.anki2", names)
@@ -150,8 +196,8 @@ class TestAnkiBuild(unittest.TestCase):
         try:
             note_count = conn.execute("select count(*) from notes").fetchone()[0]
             card_count = conn.execute("select count(*) from cards").fetchone()[0]
-            self.assertEqual(note_count, 2)
+            self.assertEqual(note_count, 1)
-            self.assertEqual(card_count, 2)
+            self.assertEqual(card_count, 1)
             fields = [row[0] for row in conn.execute("select flds from notes")]
         finally:
             conn.close()
@@ -253,15 +299,190 @@ class TestAnkiBuild(unittest.TestCase):
         self.assertIn('<rect data-af-white-background="1"', once)
         self.assertLess(once.index("<rect"), once.index("<path"))
 
+    def test_explanation_is_appended_to_back_when_present(self):
+        self.explanations_path.write_text(
+            json.dumps({
+                "NA101": {
+                    "revision": 1,
+                    "explanation": "Doubling the voltage halves the current for fixed power.",
+                    "source": "https://example.invalid/ohms-law",
+                    "confidence": 8,
+                }
+            }),
+            encoding="utf-8",
+        )
+        results = self._build_all()
+        n_result = next(
+            r for r in results
+            if r["deck"] == "Amateurfunk::Technische Kenntnisse::N"
+        )
+        self.assertEqual(n_result["explanations"], 1)
+
+        apkg = self.out_dir / "amateurfunk-technische-kenntnisse-n.apkg"
+        db_path, _media, _names = extract_collection(apkg, self.root / "n")
+        conn = sqlite3.connect(db_path)
+        try:
+            fields = [row[0] for row in conn.execute("select flds from notes")]
+        finally:
+            conn.close()
+        joined = "\n".join(fields)
+        self.assertIn("af-explanation", joined)
+        self.assertIn("Doubling the voltage halves", joined)
+        self.assertIn(
+            '<a href="https://example.invalid/ohms-law">'
+            'https://example.invalid/ohms-law</a>',
+            joined,
+        )
+
+    def test_low_confidence_explanation_shows_badge(self):
+        item = aa.QuestionItem(
+            question=question("TEST123", "1", "Prompt?", None),
+            path=("Prüfungsfragen im Prüfungsteil: Technische Kenntnisse", "Leaf"),
+        )
+        low = {"TEST123": {
+            "revision": 1,
+            "explanation": "Weak guess.",
+            "source": "TBD",
+            "confidence": aa.LOW_CONFIDENCE_THRESHOLD - 1,
+        }}
+        high = {"TEST123": {
+            "revision": 1,
+            "explanation": "Solid reasoning.",
+            "source": "AFuV §1",
+            "confidence": aa.LOW_CONFIDENCE_THRESHOLD,
+        }}
+        media = aa.MediaRegistry(self.root / "missing")
+        _f1, back_low, _l1 = aa.render_question(item, media, "seed", low)
+        _f2, back_high, _l2 = aa.render_question(item, media, "seed", high)
+        self.assertIn("af-explanation-low-confidence", back_low)
+        self.assertIn("low confidence", back_low)
+        self.assertNotIn("af-explanation-low-confidence", back_high)
+        self.assertNotIn("low confidence", back_high)
+
+    def test_missing_explanation_leaves_card_unchanged(self):
+        self._build_all()
+        apkg = self.out_dir / "amateurfunk-technische-kenntnisse-e.apkg"
+        db_path, _media, _names = extract_collection(apkg, self.root / "e")
+        conn = sqlite3.connect(db_path)
+        try:
+            fields = [row[0] for row in conn.execute("select flds from notes")]
+        finally:
+            conn.close()
+        self.assertNotIn("af-explanation", "\n".join(fields))
+
+    def test_non_url_source_is_rendered_as_plain_text(self):
+        item = aa.QuestionItem(
+            question=question("TEST123", "1", "Prompt?", None),
+            path=("Prüfungsfragen im Prüfungsteil: Technische Kenntnisse", "Leaf"),
+        )
+        explanations = {
+            "TEST123": {
+                "revision": 1,
+                "explanation": "Because $P = U \\cdot I$ and the spec fixes P.",
+                "source": "AFuV §16(2)",
+                "confidence": 9,
+            }
+        }
+        _front, back, _label = aa.render_question(
+            item, aa.MediaRegistry(self.root / "missing"), "seed", explanations,
+        )
+        self.assertIn("AFuV §16(2)", back)
+        self.assertNotIn("<a href=", back)
+        # Inline LaTeX in the explanation body goes through text_html().
+        self.assertIn(r"\(P = U \cdot I\)", back)
+
+    def test_unknown_explanation_key_fails_the_build(self):
+        self.explanations_path.write_text(
+            json.dumps({
+                "NA10I": {  # typo: 'I' instead of '1' — not in the catalog
+                    "revision": 1,
+                    "explanation": "ok",
+                    "source": "ok",
+                    "confidence": 5,
+                }
+            }),
+            encoding="utf-8",
+        )
+        with self.assertRaises(aa.AnkiBuildError) as ctx:
+            self._build_all()
+        self.assertIn("NA10I", str(ctx.exception))
+
+    def test_schema_rejects_bool_in_integer_fields(self):
+        for bad_field in ("revision", "confidence"):
+            entry = {
+                "revision": 1,
+                "explanation": "ok",
+                "source": "ok",
+                "confidence": 5,
+            }
+            entry[bad_field] = True
+            self.explanations_path.write_text(
+                json.dumps({"NA101": entry}),
+                encoding="utf-8",
+            )
+            with self.assertRaises(aa.AnkiBuildError):
+                self._build_all()
+
+    def test_schema_rejects_extra_fields(self):
+        self.explanations_path.write_text(
+            json.dumps({
+                "NA101": {
+                    "revision": 1,
+                    "explanation": "ok",
+                    "source": "ok",
+                    "confidence": 5,
+                    "author": "claude",
+                }
+            }),
+            encoding="utf-8",
+        )
+        with self.assertRaises(aa.AnkiBuildError) as ctx:
+            self._build_all()
+        self.assertIn("author", str(ctx.exception))
+
+    def test_invalid_explanation_schema_raises_build_error(self):
+        self.explanations_path.write_text(
+            json.dumps({
+                "NA101": {
+                    "revision": 1,
+                    "explanation": "ok",
+                    "source": "ok",
+                    "confidence": 11,  # out of range
+                }
+            }),
+            encoding="utf-8",
+        )
+        with self.assertRaises(aa.AnkiBuildError):
+            self._build_all()
+
+    def test_missing_explanations_file_is_treated_as_empty(self):
+        missing = self.root / "does-not-exist.json"
+        results = aa.build_all(
+            self.data_dir,
+            self.out_dir,
+            seed="test-seed",
+            explanations_path=missing,
+        )
+        self.assertTrue(all(r["explanations"] == 0 for r in results))
+
     def test_main_returns_success(self):
-        rc = aa.main(["--data", str(self.data_dir), "--out", str(self.out_dir)])
+        rc = aa.main([
+            "--data", str(self.data_dir),
+            "--out", str(self.out_dir),
+            "--explanations", str(self.explanations_path),
+        ])
         self.assertEqual(rc, aa.EXIT_OK)
 
     def test_build_is_byte_deterministic_for_same_input(self):
         out_a = self.root / "anki-a"
         out_b = self.root / "anki-b"
-        aa.main(["--data", str(self.data_dir), "--out", str(out_a), "--epoch", "1234567890"])
+        common = [
-        aa.main(["--data", str(self.data_dir), "--out", str(out_b), "--epoch", "1234567890"])
+            "--data", str(self.data_dir),
+            "--explanations", str(self.explanations_path),
+            "--epoch", "1234567890",
+        ]
+        aa.main([*common, "--out", str(out_a)])
+        aa.main([*common, "--out", str(out_b)])
 
         for first in sorted(out_a.glob("*.apkg")):
             second = out_b / first.name