docs: implementation plan for GIAS code dictionaries

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-09 10:10:26 +01:00
parent 1ae5762a0a
commit 08bd86db05
1 changed files with 799 additions and 0 deletions
@@ -0,0 +1,799 @@
+# GIAS Code Dictionaries Implementation Plan
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** Store the six GIAS classification fields as official DfE integer codes in the marts and translate code → name in application code, leaving the API contract (name strings) unchanged.
+
+**Architecture:** A generation script downloads the public GIAS bulk CSV and emits the dictionaries (Python dicts + a dbt seed) from real data. The tap ingests the `(code)` columns, staging casts them, `dim_school`/`dim_location` keep only codes, and translation happens in exactly two places: `backend/data_loader.py` right after `pd.read_sql`, and `pipeline/scripts/sync_typesense.py` before indexing. A dbt seed test warns when DfE adds/renames a value; a parity test keeps the backend and pipeline dictionary copies identical.
+
+**Tech Stack:** Singer SDK tap, dbt (Postgres), FastAPI + pandas, Typesense sync script, pytest.
+
+**Spec:** `docs/superpowers/specs/2026-07-09-gias-code-dictionaries-design.md`
+
+## Global Constraints
+
+- **Numeric code values are never assumed.** Every literal code used in SQL or yml (status filter, sixth-form derivation, phase cascade) must be verified against `pipeline/transform/seeds/gias_code_names.csv` generated in Task 1 from the live CSV. The literals written in this plan are best-current-knowledge and each carries a verification step.
+- **Names served by the API must stay byte-identical** to today's strings (e.g. `Does not apply`, `Open, but proposed to close`) — UI heuristics compare exact strings.
+- The `(name)` columns stay declared in the tap and present in raw; staging stops exposing them.
+- `dim_school` and `dim_location` status filters must stay identical (API inner-joins them).
+- Backend tests run via: `uv run --with-requirements requirements.txt --with pytest --with "httpx==0.27.0" python -m pytest backend/tests -v` (no local pytest exists).
+- dbt cannot run locally — dbt changes are verified statically (grep / yaml parse) + CI.
+- Never push to `main`. Work on branch `feat/gias-code-dictionaries` (branch off `docs/gias-code-dictionaries` so the spec is included, or off `main` if that has merged).
+- Commits end with: `Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>`
+- Deploy runbook (accepted window, spec §7): merge → deploy → trigger `school_data_daily` immediately. No code-level fallback for the old-schema window.
+
+---
+
+### Task 1: Dictionary generation script, canonical module, pipeline copy, seed
+
+**Files:**
+- Create: `pipeline/scripts/generate_gias_codes.py`
+- Create: `backend/gias_codes.py` (content generated by the script)
+- Create: `pipeline/scripts/gias_codes.py` (byte-identical copy)
+- Create: `pipeline/transform/seeds/gias_code_names.csv` (generated)
+- Test: `backend/tests/test_gias_codes.py`
+
+**Interfaces:**
+- Produces: `backend/gias_codes.py` exporting `SCHOOL_TYPE`, `ESTABLISHMENT_STATUS`, `PHASE_OF_EDUCATION`, `OFFICIAL_SIXTH_FORM`, `RELIGIOUS_CHARACTER`, `ADMISSIONS_POLICY` (each `dict[int, str]`) and `translate(code, mapping) -> str | None`. Task 4 imports these; Task 5 imports the pipeline copy; Task 3 reads code literals from the seed CSV.
+
+- [ ] **Step 1: Write the failing tests**
+
+Create `backend/tests/test_gias_codes.py`:
+
+```python
+"""Tests for the GIAS code->name dictionaries (spec 2026-07-09).
+
+The dictionaries are generated from the live GIAS bulk CSV by
+pipeline/scripts/generate_gias_codes.py — these tests assert the module's
+contract, key sentinel values the marts/UI depend on, and that the pipeline
+copy has not drifted from the canonical backend module.
+"""
+
+import math
+from pathlib import Path
+
+from backend.gias_codes import (
+    ADMISSIONS_POLICY,
+    ESTABLISHMENT_STATUS,
+    OFFICIAL_SIXTH_FORM,
+    PHASE_OF_EDUCATION,
+    RELIGIOUS_CHARACTER,
+    SCHOOL_TYPE,
+    translate,
+)
+
+REPO = Path(__file__).resolve().parents[2]
+
+
+def test_translate_known_code():
+    open_code = next(c for c, n in ESTABLISHMENT_STATUS.items() if n == "Open")
+    assert translate(open_code, ESTABLISHMENT_STATUS) == "Open"
+
+
+def test_translate_unknown_code_degrades_gracefully():
+    assert translate(9999, ESTABLISHMENT_STATUS) == "Unknown (9999)"
+
+
+def test_translate_none_and_nan_return_none():
+    assert translate(None, ESTABLISHMENT_STATUS) is None
+    assert translate(float("nan"), ESTABLISHMENT_STATUS) is None
+
+
+def test_translate_accepts_float_codes():
+    # pd.read_sql yields float columns when NULLs are present
+    open_code = next(c for c, n in ESTABLISHMENT_STATUS.items() if n == "Open")
+    assert translate(float(open_code), ESTABLISHMENT_STATUS) == "Open"
+
+
+def test_sentinel_names_present():
+    """Names the marts/UI compare against must exist verbatim."""
+    assert "Open" in ESTABLISHMENT_STATUS.values()
+    assert "Open, but proposed to close" in ESTABLISHMENT_STATUS.values()
+    assert "Has a sixth form" in OFFICIAL_SIXTH_FORM.values()
+    assert "Primary" in PHASE_OF_EDUCATION.values()
+    assert "Secondary" in PHASE_OF_EDUCATION.values()
+    assert "Does not apply" in RELIGIOUS_CHARACTER.values()
+    assert all(len(d) > 0 for d in (
+        SCHOOL_TYPE, ESTABLISHMENT_STATUS, PHASE_OF_EDUCATION,
+        OFFICIAL_SIXTH_FORM, RELIGIOUS_CHARACTER, ADMISSIONS_POLICY,
+    ))
+
+
+def test_pipeline_copy_is_identical():
+    canonical = (REPO / "backend" / "gias_codes.py").read_text()
+    copy = (REPO / "pipeline" / "scripts" / "gias_codes.py").read_text()
+    assert canonical == copy, (
+        "pipeline/scripts/gias_codes.py has drifted from backend/gias_codes.py — "
+        "regenerate with pipeline/scripts/generate_gias_codes.py and copy the file"
+    )
+
+
+def test_seed_matches_dictionaries():
+    import csv
+    fields = {
+        "school_type": SCHOOL_TYPE,
+        "establishment_status": ESTABLISHMENT_STATUS,
+        "phase_of_education": PHASE_OF_EDUCATION,
+        "official_sixth_form": OFFICIAL_SIXTH_FORM,
+        "religious_character": RELIGIOUS_CHARACTER,
+        "admissions_policy": ADMISSIONS_POLICY,
+    }
+    seed_path = REPO / "pipeline" / "transform" / "seeds" / "gias_code_names.csv"
+    seed: dict[str, dict[int, str]] = {k: {} for k in fields}
+    with open(seed_path, newline="") as fh:
+        for row in csv.DictReader(fh):
+            seed[row["field"]][int(row["code"])] = row["name"]
+    assert seed == fields
+```
+
+- [ ] **Step 2: Run tests to verify they fail**
+
+Run: `cd /Users/tudor/projects/school_compare && uv run --with-requirements requirements.txt --with pytest --with "httpx==0.27.0" python -m pytest backend/tests/test_gias_codes.py -v`
+Expected: FAIL at import — `ModuleNotFoundError: No module named 'backend.gias_codes'`.
+
+- [ ] **Step 3: Write the generation script**
+
+Create `pipeline/scripts/generate_gias_codes.py`:
+
+```python
+"""Generate GIAS code->name dictionaries from the live bulk CSV.
+
+Writes:
+  - backend/gias_codes.py                        (canonical Python module)
+  - pipeline/scripts/gias_codes.py               (byte-identical copy)
+  - pipeline/transform/seeds/gias_code_names.csv (dbt seed for drift test)
+
+Run from the repo root whenever the dbt drift test warns that DfE
+added/renamed a value:  python pipeline/scripts/generate_gias_codes.py
+"""
+
+from __future__ import annotations
+
+import io
+import sys
+from datetime import date, timedelta
+from pathlib import Path
+
+import pandas as pd
+import requests
+
+GIAS_URL = (
+    "https://ea-edubase-api-prod.azurewebsites.net"
+    "/edubase/downloads/public/edubasealldata{date}.csv"
+)
+
+# (CSV code column, CSV name column, python dict name, seed field key)
+FIELDS = [
+    ("TypeOfEstablishment (code)", "TypeOfEstablishment (name)", "SCHOOL_TYPE", "school_type"),
+    ("EstablishmentStatus (code)", "EstablishmentStatus (name)", "ESTABLISHMENT_STATUS", "establishment_status"),
+    ("PhaseOfEducation (code)", "PhaseOfEducation (name)", "PHASE_OF_EDUCATION", "phase_of_education"),
+    ("OfficialSixthForm (code)", "OfficialSixthForm (name)", "OFFICIAL_SIXTH_FORM", "official_sixth_form"),
+    ("ReligiousCharacter (code)", "ReligiousCharacter (name)", "RELIGIOUS_CHARACTER", "religious_character"),
+    ("AdmissionsPolicy (code)", "AdmissionsPolicy (name)", "ADMISSIONS_POLICY", "admissions_policy"),
+]
+
+MODULE_HEADER = '''"""GIAS code -> name dictionaries.
+
+GENERATED by pipeline/scripts/generate_gias_codes.py from the GIAS bulk CSV
+— do not edit by hand; rerun the script when the dbt drift test warns.
+The canonical file is backend/gias_codes.py; pipeline/scripts/gias_codes.py
+must be byte-identical (enforced by backend/tests/test_gias_codes.py).
+"""
+
+from __future__ import annotations
+
+import logging
+import math
+
+logger = logging.getLogger(__name__)
+
+'''
+
+MODULE_FOOTER = '''
+
+def translate(code, mapping: dict[int, str]) -> str | None:
+    """Translate a GIAS code to its display name.
+
+    None/NaN -> None (column absent or suppressed). Unknown codes degrade to
+    "Unknown (<code>)" with a warning so a new DfE value never blanks the UI.
+    """
+    if code is None or (isinstance(code, float) and math.isnan(code)):
+        return None
+    code = int(code)
+    if code not in mapping:
+        logger.warning("Unknown GIAS code %s (not in dictionary)", code)
+        return f"Unknown ({code})"
+    return mapping[code]
+'''
+
+
+def download_csv() -> pd.DataFrame:
+    for day in (date.today(), date.today() - timedelta(days=1)):
+        url = GIAS_URL.format(date=day.strftime("%Y%m%d"))
+        print(f"Downloading {url}")
+        resp = requests.get(url, timeout=300)
+        if resp.status_code == 404:
+            continue
+        resp.raise_for_status()
+        return pd.read_csv(
+            io.StringIO(resp.content.decode("latin-1")),
+            dtype=str, keep_default_na=False,
+        )
+    sys.exit("GIAS CSV not available for today or yesterday")
+
+
+def main() -> None:
+    repo = Path(__file__).resolve().parents[2]
+    df = download_csv()
+
+    module_parts = [MODULE_HEADER]
+    seed_rows: list[tuple[str, int, str]] = []
+
+    for code_col, name_col, dict_name, field_key in FIELDS:
+        pairs = (
+            df[[code_col, name_col]]
+            .loc[lambda d: (d[code_col] != "") & (d[name_col] != "")]
+            .drop_duplicates()
+        )
+        mapping = sorted((int(c), n) for c, n in pairs.itertuples(index=False))
+        dupes = len(mapping) - len({c for c, _ in mapping})
+        if dupes:
+            sys.exit(f"{code_col}: {dupes} codes map to multiple names — investigate before generating")
+        lines = [f"{dict_name}: dict[int, str] = {{"]
+        for code, name in mapping:
+            escaped = name.replace('"', '\\"')
+            lines.append(f'    {code}: "{escaped}",')
+        lines.append("}\n")
+        module_parts.append("\n".join(lines))
+        seed_rows += [(field_key, code, name) for code, name in mapping]
+
+    module = "\n".join(module_parts) + MODULE_FOOTER
+
+    (repo / "backend" / "gias_codes.py").write_text(module)
+    (repo / "pipeline" / "scripts" / "gias_codes.py").write_text(module)
+
+    seed_path = repo / "pipeline" / "transform" / "seeds" / "gias_code_names.csv"
+    with open(seed_path, "w", newline="") as fh:
+        import csv
+        w = csv.writer(fh)
+        w.writerow(["field", "code", "name"])
+        w.writerows(seed_rows)
+
+    print(f"Wrote backend/gias_codes.py, pipeline/scripts/gias_codes.py, {seed_path.name}")
+    print("\nKey codes for the dbt work (Task 3):")
+    for field in ("establishment_status", "phase_of_education", "official_sixth_form"):
+        print(f"  {field}:")
+        for f, code, name in seed_rows:
+            if f == field:
+                print(f"    {code} = {name}")
+
+
+if __name__ == "__main__":
+    main()
+```
+
+- [ ] **Step 4: Run the generator**
+
+Run: `cd /Users/tudor/projects/school_compare && uv run --with pandas --with requests python pipeline/scripts/generate_gias_codes.py`
+Expected: downloads the CSV (~100MB, may take a minute), writes the three files, and prints the status/phase/sixth-form code tables. **Record the printed code tables — Task 3 needs them.** If the download fails twice, report BLOCKED (no network or GIAS outage) rather than inventing dictionary content.
+
+- [ ] **Step 5: Run the tests again**
+
+Run: `cd /Users/tudor/projects/school_compare && uv run --with-requirements requirements.txt --with pytest --with "httpx==0.27.0" python -m pytest backend/tests/test_gias_codes.py -v`
+Expected: 7 passed. If `test_sentinel_names_present` fails, the GIAS vocabulary differs from expectations — inspect the generated module and report DONE_WITH_CONCERNS naming the differing value; do not edit the generated names.
+
+- [ ] **Step 6: Commit**
+
+```bash
+git add pipeline/scripts/generate_gias_codes.py backend/gias_codes.py pipeline/scripts/gias_codes.py pipeline/transform/seeds/gias_code_names.csv backend/tests/test_gias_codes.py
+git commit -m "feat: GIAS code->name dictionaries generated from live bulk CSV
+
+Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>"
+```
+
+---
+
+### Task 2: Tap ingests the (code) columns; staging exposes codes, drops names
+
+**Files:**
+- Modify: `pipeline/plugins/extractors/tap-uk-gias/tap_uk_gias/tap.py` (Singer schema)
+- Modify: `pipeline/transform/models/staging/stg_gias_establishments.sql`
+
+**Interfaces:**
+- Produces: staging columns `school_type_code`, `status_code`, `phase_code`, `official_sixth_form_code`, `religious_character_code`, `admissions_policy_code` (all int) consumed by Task 3. Staging **stops exposing** `school_type`, `status`, `phase`, `official_sixth_form`, `religious_character`, `admissions_policy` (names stay in raw only).
+
+- [ ] **Step 1: Add the six (code) properties to the Singer schema**
+
+In `tap.py`, `GIASEstablishmentsStream.schema`, add each `(code)` property directly above its existing `(name)` sibling:
+
+```python
+        th.Property("TypeOfEstablishment (code)", th.StringType),
+        th.Property("PhaseOfEducation (code)", th.StringType),
+        th.Property("EstablishmentStatus (code)", th.StringType),
+        th.Property("Gender (name)", ...)  # existing line — for placement reference only
+        th.Property("ReligiousCharacter (code)", th.StringType),
+        th.Property("AdmissionsPolicy (code)", th.StringType),
+        th.Property("OfficialSixthForm (code)", th.StringType),
+```
+
+(The exact insertion order doesn't matter — the schema is a dict — but keep each `(code)` adjacent to its `(name)` for readability. Do NOT remove any `(name)` property.)
+
+- [ ] **Step 2: Rewrite the six columns in staging**
+
+In `stg_gias_establishments.sql` `renamed` CTE, replace:
+
+```sql
+        "TypeOfEstablishment (name)"                    as school_type,
+        "PhaseOfEducation (name)"                       as phase,
+        nullif(trim("OfficialSixthForm (name)"), '')    as official_sixth_form,
+        "ReligiousCharacter (name)"                     as religious_character,
+        "AdmissionsPolicy (name)"                       as admissions_policy,
+        "EstablishmentStatus (name)"                    as status,
+```
+
+with:
+
+```sql
+        cast(nullif(trim("TypeOfEstablishment (code)"), '') as integer)  as school_type_code,
+        cast(nullif(trim("PhaseOfEducation (code)"), '') as integer)     as phase_code,
+        cast(nullif(trim("OfficialSixthForm (code)"), '') as integer)    as official_sixth_form_code,
+        cast(nullif(trim("ReligiousCharacter (code)"), '') as integer)   as religious_character_code,
+        cast(nullif(trim("AdmissionsPolicy (code)"), '') as integer)     as admissions_policy_code,
+        cast(nullif(trim("EstablishmentStatus (code)"), '') as integer)  as status_code,
+```
+
+(The name lines are scattered through the CTE — replace each in place; the six name aliases must no longer appear in the model.)
+
+- [ ] **Step 3: Verify statically**
+
+Run:
+```bash
+cd /Users/tudor/projects/school_compare && \
+python3 -c "import ast; ast.parse(open('pipeline/plugins/extractors/tap-uk-gias/tap_uk_gias/tap.py').read()); print('tap OK')" && \
+grep -c "(code)" pipeline/plugins/extractors/tap-uk-gias/tap_uk_gias/tap.py && \
+grep -E "as (school_type|status|phase|official_sixth_form|religious_character|admissions_policy)," pipeline/transform/models/staging/stg_gias_establishments.sql; echo "name-alias grep exit=$? (want 1 = none found)"
+```
+Expected: `tap OK`, code-column count `6`, and the final grep finds nothing (exit 1).
+
+- [ ] **Step 4: Commit**
+
+```bash
+git add pipeline/plugins/extractors/tap-uk-gias/tap_uk_gias/tap.py pipeline/transform/models/staging/stg_gias_establishments.sql
+git commit -m "feat(pipeline): ingest GIAS code columns; staging exposes codes not names
+
+Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>"
+```
+
+---
+
+### Task 3: Marts store codes; dbt tests + drift test
+
+**Files:**
+- Modify: `pipeline/transform/models/marts/dim_school.sql`
+- Modify: `pipeline/transform/models/marts/dim_location.sql`
+- Modify: `pipeline/transform/models/marts/_marts_schema.yml`
+- Create: `pipeline/transform/tests/assert_gias_code_names_match_seed.sql`
+
+**Interfaces:**
+- Consumes: staging code columns from Task 2; code literals from `pipeline/transform/seeds/gias_code_names.csv` (Task 1).
+- Produces: `dim_school` columns `school_type_code`, `status_code`, `phase_code`, `religious_character_code`, `admissions_policy_code` (int) replacing their string columns; `has_sixth_form` unchanged (bool). Task 4's `_MAIN_QUERY` selects these.
+
+**Before writing SQL: open `pipeline/transform/seeds/gias_code_names.csv` and confirm the literals below.** Best-current-knowledge values (VERIFY EACH):
+`establishment_status`: 1 = Open, 3 = "Open, but proposed to close" (2 = Closed, 4 = Proposed to open).
+`phase_of_education`: 0 = Not applicable, 2 = Primary, 4 = Secondary, 7 = All-through.
+`official_sixth_form`: 1 = Has a sixth form, 2 = Does not have a sixth form, 0 = Not applicable.
+If any differ, use the seed's values everywhere below and say so in your report.
+
+- [ ] **Step 1: Rewrite dim_school.sql derivations in code space**
+
+Replace the phase cascade block (`case ... end as phase,`) with:
+
+```sql
+    -- Phase in GIAS code space (see seeds/gias_code_names.csv):
+    -- 2 = Primary, 4 = Secondary, 7 = All-through, 0 = Not applicable.
+    case
+        -- 1. Trust GIAS phase when it's a real value (0 = the catch-all "Not Applicable")
+        when s.phase_code is not null and s.phase_code != 0
+            then s.phase_code
+        -- 2. Infer from statutory age range (independent schools still publish these)
+        when s.statutory_high_age is not null and s.statutory_high_age <= 11 then 2
+        when s.statutory_low_age is not null and s.statutory_low_age >= 11 then 4
+        when s.statutory_low_age is not null and s.statutory_high_age is not null
+             and s.statutory_low_age < 11 and s.statutory_high_age > 11 then 7
+        -- 3. Fallback: infer from school name (covers independents with missing ages)
+        when s.school_name ilike '%primary%'
+             or s.school_name ilike '%infant%'
+             or s.school_name ilike '%junior%'
+             or s.school_name ilike '%preparatory%'
+             or s.school_name ilike '% prep school%'
+             or s.school_name ilike '% prep %'
+            then 2
+        when s.school_name ilike '%secondary%'
+             or s.school_name ilike '%high school%'
+             or s.school_name ilike '%grammar%'
+             or s.school_name ilike '%senior school%'
+             or s.school_name ilike '%upper school%'
+            then 4
+        -- 4. Give up — null renders no phase pill
+        else null
+    end as phase_code,
+```
+
+Replace `s.school_type,` with `s.school_type_code,`; `s.religious_character,` with `s.religious_character_code,`; `s.admissions_policy,` with `s.admissions_policy_code,`; `s.status,` with `s.status_code,`.
+
+Replace the has_sixth_form case with:
+
+```sql
+    -- GIAS OfficialSixthForm in code space: 1 = has, 2 = does not, 0 = N/A.
+    -- Null (rare, new establishments) falls back to the statutory age range.
+    case
+        when s.official_sixth_form_code = 1 then true
+        when s.official_sixth_form_code in (0, 2) then false
+        else coalesce(s.statutory_high_age >= 18, false)
+    end as has_sixth_form,
+```
+
+Replace the status filter with:
+
+```sql
+-- 1 = Open; 3 = Open, but proposed to close (still operating; drops out when
+-- GIAS flips to Closed — marts fully rebuild each run).
+where s.status_code in (1, 3)
+```
+
+- [ ] **Step 2: Same filter in dim_location.sql**
+
+Replace its `where s.status in ('Open', 'Open, but proposed to close')` (and the comment above it) with:
+
+```sql
+-- Must match dim_school's status filter exactly (the API inner-joins the two).
+where s.status_code in (1, 3)
+```
+
+- [ ] **Step 3: Update _marts_schema.yml**
+
+Under `dim_school` columns: rename `phase` → `phase_code` (keep the warn-severity not_null, reword description to mention codes); replace the `status` accepted_values block with:
+
+```yaml
+      - name: status_code
+        description: GIAS EstablishmentStatus code (1 = Open, 3 = Open but proposed to close)
+        tests:
+          - accepted_values:
+              values: [1, 3]
+```
+
+Add warn-severity accepted_values for the other codes, values copied from the seed (school_type/religious/admissions lists are long — paste the full code list from `gias_code_names.csv` for each):
+
+```yaml
+      - name: school_type_code
+        tests:
+          - accepted_values:
+              severity: warn
+              values: [<all school_type codes from the seed>]
+      - name: religious_character_code
+        tests:
+          - accepted_values:
+              severity: warn
+              values: [<all religious_character codes from the seed>]
+      - name: admissions_policy_code
+        tests:
+          - accepted_values:
+              severity: warn
+              values: [<all admissions_policy codes from the seed>]
+```
+
+(`<...>` here means: paste the actual comma-separated integers from the seed file — the lists exist by the time this task runs. Leaving a literal `<...>` in the yml is a task failure.)
+
+`has_sixth_form` tests stay unchanged.
+
+- [ ] **Step 4: Write the drift test**
+
+Create `pipeline/transform/tests/assert_gias_code_names_match_seed.sql`:
+
+```sql
+-- Warn when the live GIAS CSV carries a (code, name) pair we don't have in
+-- the dictionary seed — i.e. DfE added or renamed a value. Fix by rerunning
+-- pipeline/scripts/generate_gias_codes.py and committing the regenerated
+-- dictionaries + seed together.
+{{ config(severity='warn') }}
+
+with raw_pairs as (
+    {% for field_key, code_col, name_col in [
+        ('school_type',          'TypeOfEstablishment (code)',  'TypeOfEstablishment (name)'),
+        ('establishment_status', 'EstablishmentStatus (code)',  'EstablishmentStatus (name)'),
+        ('phase_of_education',   'PhaseOfEducation (code)',     'PhaseOfEducation (name)'),
+        ('official_sixth_form',  'OfficialSixthForm (code)',    'OfficialSixthForm (name)'),
+        ('religious_character',  'ReligiousCharacter (code)',   'ReligiousCharacter (name)'),
+        ('admissions_policy',    'AdmissionsPolicy (code)',     'AdmissionsPolicy (name)')
+    ] %}
+    select distinct
+        '{{ field_key }}' as field,
+        cast(nullif(trim("{{ code_col }}"), '') as integer) as code,
+        nullif(trim("{{ name_col }}"), '') as name
+    from {{ source('raw', 'gias_establishments') }}
+    where nullif(trim("{{ code_col }}"), '') is not null
+      and nullif(trim("{{ name_col }}"), '') is not null
+    {% if not loop.last %}union all{% endif %}
+    {% endfor %}
+)
+
+select r.*
+from raw_pairs r
+left join {{ ref('gias_code_names') }} s
+    on s.field = r.field
+   and s.code = r.code
+   and s.name = r.name
+where s.field is null
+```
+
+- [ ] **Step 5: Verify statically**
+
+Run:
+```bash
+cd /Users/tudor/projects/school_compare && \
+uv run --with pyyaml python -c "import yaml; yaml.safe_load(open('pipeline/transform/models/marts/_marts_schema.yml')); print('yml OK')" && \
+grep -c "_code" pipeline/transform/models/marts/dim_school.sql && \
+grep -n "status_code in (1, 3)" pipeline/transform/models/marts/dim_school.sql pipeline/transform/models/marts/dim_location.sql && \
+grep -rn "s\.status\b\|s\.phase\b\|s\.school_type\b\|s\.religious_character\b\|s\.admissions_policy\b\|official_sixth_form\b" pipeline/transform/models/marts/dim_school.sql | grep -v "_code"; echo "stale-name grep exit=$? (want 1)"
+```
+Expected: `yml OK`, both filters matched, and no stale name-column references (final grep exits 1).
+
+- [ ] **Step 6: Commit**
+
+```bash
+git add pipeline/transform/models/marts/dim_school.sql pipeline/transform/models/marts/dim_location.sql pipeline/transform/models/marts/_marts_schema.yml pipeline/transform/tests/assert_gias_code_names_match_seed.sql
+git commit -m "feat(pipeline): dim_school/dim_location store GIAS codes; seed drift test
+
+Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>"
+```
+
+---
+
+### Task 4: Backend translates at the API boundary
+
+**Files:**
+- Modify: `backend/models.py` (DimSchool columns)
+- Modify: `backend/data_loader.py` (`_MAIN_QUERY` + translation)
+- Test: `backend/tests/test_gias_translation.py` (new)
+
+**Interfaces:**
+- Consumes: `backend/gias_codes.py` dictionaries + `translate` (Task 1); mart code columns (Task 3).
+- Produces: `translate_gias_code_columns(df) -> df` in `backend/data_loader.py`; after `load_school_data_as_dataframe()` the DataFrame carries today's name columns (`phase`, `school_type`, `status`, `religious_denomination`, `admissions_policy`) — every downstream consumer unchanged.
+
+- [ ] **Step 1: Write the failing tests**
+
+Create `backend/tests/test_gias_translation.py`:
+
+```python
+"""API-boundary translation: marts now carry GIAS codes; the DataFrame the
+rest of the backend sees must carry today's name strings."""
+
+import numpy as np
+import pandas as pd
+
+from backend.data_loader import translate_gias_code_columns
+from backend.gias_codes import ESTABLISHMENT_STATUS, PHASE_OF_EDUCATION
+
+
+def _code_for(mapping, name):
+    return next(c for c, n in mapping.items() if n == name)
+
+
+def test_codes_become_todays_names():
+    df = pd.DataFrame([{
+        "urn": 1,
+        "phase_code": float(_code_for(PHASE_OF_EDUCATION, "Primary")),
+        "school_type_code": np.nan,
+        "status_code": float(_code_for(ESTABLISHMENT_STATUS, "Open, but proposed to close")),
+        "religious_character_code": np.nan,
+        "admissions_policy_code": np.nan,
+    }])
+    out = translate_gias_code_columns(df)
+    row = out.iloc[0]
+    assert row["phase"] == "Primary"
+    assert row["status"] == "Open, but proposed to close"
+    assert row["school_type"] is None
+    assert row["religious_denomination"] is None
+    assert row["admissions_policy"] is None
+
+
+def test_unknown_code_degrades_not_blanks():
+    df = pd.DataFrame([{"urn": 1, "phase_code": 9999.0}])
+    out = translate_gias_code_columns(df)
+    assert out.iloc[0]["phase"] == "Unknown (9999)"
+
+
+def test_missing_code_columns_are_a_noop():
+    """Old-schema DataFrames (tests, pre-pipeline DBs) pass through untouched."""
+    df = pd.DataFrame([{"urn": 1, "phase": "Primary", "status": "Open"}])
+    out = translate_gias_code_columns(df)
+    assert out.iloc[0]["phase"] == "Primary"
+    assert out.iloc[0]["status"] == "Open"
+```
+
+- [ ] **Step 2: Run to verify failure**
+
+Run: `cd /Users/tudor/projects/school_compare && uv run --with-requirements requirements.txt --with pytest --with "httpx==0.27.0" python -m pytest backend/tests/test_gias_translation.py -v`
+Expected: FAIL — `ImportError: cannot import name 'translate_gias_code_columns'`.
+
+- [ ] **Step 3: Implement translation in data_loader.py**
+
+Add near the top of `backend/data_loader.py` (after existing imports):
+
+```python
+from .gias_codes import (
+    ADMISSIONS_POLICY,
+    ESTABLISHMENT_STATUS,
+    PHASE_OF_EDUCATION,
+    RELIGIOUS_CHARACTER,
+    SCHOOL_TYPE,
+    translate,
+)
+
+# mart code column -> (API name column, dictionary)
+_GIAS_CODE_COLUMNS = {
+    "phase_code": ("phase", PHASE_OF_EDUCATION),
+    "school_type_code": ("school_type", SCHOOL_TYPE),
+    "status_code": ("status", ESTABLISHMENT_STATUS),
+    "religious_character_code": ("religious_denomination", RELIGIOUS_CHARACTER),
+    "admissions_policy_code": ("admissions_policy", ADMISSIONS_POLICY),
+}
+
+
+def translate_gias_code_columns(df: pd.DataFrame) -> pd.DataFrame:
+    """Map GIAS code columns to today's name columns (API contract).
+
+    Runs immediately after pd.read_sql so every downstream consumer —
+    filters, PHASE_GROUPS, payloads, /api/filters — keeps seeing names.
+    DataFrames without the code columns (old schema, test fixtures) pass
+    through unchanged.
+    """
+    for code_col, (name_col, mapping) in _GIAS_CODE_COLUMNS.items():
+        if code_col in df.columns:
+            df[name_col] = df[code_col].map(lambda c: translate(c, mapping))
+    return df
+```
+
+- [ ] **Step 4: Switch `_MAIN_QUERY` to code columns and call the translation**
+
+In `_MAIN_QUERY` replace:
+`s.phase,` → `s.phase_code,` · `s.school_type,` → `s.school_type_code,` · `s.religious_character AS religious_denomination,` → `s.religious_character_code,` · `s.admissions_policy,` → `s.admissions_policy_code,` · `s.status,` → `s.status_code,`
+
+In `load_school_data_as_dataframe()`, insert the call immediately after the empty-check and **before** the existing `normalize_school_type` line:
+
+```python
+    if df.empty:
+        return df
+
+    df = translate_gias_code_columns(df)
+
+    # Build address string
+    ...
+    # Normalize school type  (existing line — now normalises the translated name)
+    df["school_type"] = df["school_type"].apply(normalize_school_type)
+```
+
+- [ ] **Step 5: Update DimSchool in models.py**
+
+Replace `phase = Column(String(100))`, `school_type = Column(String(100))`, `religious_character = Column(String(100))`, `admissions_policy = Column(String(50))`, `status = Column(String(50))` with:
+
+```python
+    phase_code = Column(Integer)
+    school_type_code = Column(Integer)
+    religious_character_code = Column(Integer)
+    admissions_policy_code = Column(Integer)
+    status_code = Column(Integer)
+```
+
+Then check nothing else references the removed attributes:
+```bash
+grep -rn "\.phase\b\|\.school_type\b\|\.religious_character\b\|\.admissions_policy\b\|\.status\b" backend/*.py | grep -i "dimschool\|DimSchool" 
+```
+Expected: no hits (the backend reads via `_MAIN_QUERY`, not ORM attributes). If there are hits, update them to the `_code` columns + translation and note it in your report.
+
+- [ ] **Step 6: Run the new tests and the whole backend suite**
+
+Run: `cd /Users/tudor/projects/school_compare && uv run --with-requirements requirements.txt --with pytest --with "httpx==0.27.0" python -m pytest backend/tests -v`
+Expected: all pass — 3 new + all pre-existing (their fixtures carry name columns; translation is a no-op on them).
+
+- [ ] **Step 7: Commit**
+
+```bash
+git add backend/models.py backend/data_loader.py backend/tests/test_gias_translation.py
+git commit -m "feat(api): translate GIAS codes to names at the query boundary
+
+Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>"
+```
+
+---
+
+### Task 5: Typesense sync translates before indexing
+
+**Files:**
+- Modify: `pipeline/scripts/sync_typesense.py`
+
+**Interfaces:**
+- Consumes: `pipeline/scripts/gias_codes.py` (Task 1), mart code columns (Task 3).
+- Produces: identical Typesense documents to today (facet values are names).
+
+- [ ] **Step 1: Switch the SELECT and translate**
+
+In `sync_typesense.py`: add at the top (the DAG runs `python scripts/sync_typesense.py`, so `scripts/` is `sys.path[0]` and a plain import works):
+
+```python
+from gias_codes import PHASE_OF_EDUCATION, RELIGIOUS_CHARACTER, SCHOOL_TYPE, translate
+```
+
+In the SQL, replace `s.phase,` → `s.phase_code,`, `s.school_type,` → `s.school_type_code,`, `s.religious_character,` → `s.religious_character_code,`.
+
+In the document builder, replace:
+
+```python
+        "phase": row["phase"] or "",
+        "school_type": row["school_type"] or "",
+```
+with:
+```python
+        "phase": translate(row["phase_code"], PHASE_OF_EDUCATION) or "",
+        "school_type": translate(row["school_type_code"], SCHOOL_TYPE) or "",
+```
+and:
+```python
+    if row.get("religious_character"):
+        doc["religious_character"] = row["religious_character"]
+```
+with:
+```python
+    religious_character = translate(row.get("religious_character_code"), RELIGIOUS_CHARACTER)
+    if religious_character:
+        doc["religious_character"] = religious_character
+```
+
+- [ ] **Step 2: Verify statically**
+
+Run:
+```bash
+cd /Users/tudor/projects/school_compare && \
+python3 -c "import ast; ast.parse(open('pipeline/scripts/sync_typesense.py').read()); print('sync OK')" && \
+grep -n "row\[\"phase\"\]\|row\[\"school_type\"\]\|row\[\"religious_character\"\]" pipeline/scripts/sync_typesense.py; echo "stale grep exit=$? (want 1)"
+```
+Expected: `sync OK`, no stale name-column row accesses.
+
+- [ ] **Step 3: Commit**
+
+```bash
+git add pipeline/scripts/sync_typesense.py
+git commit -m "feat(pipeline): typesense sync translates GIAS codes before indexing
+
+Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>"
+```
+
+---
+
+### Task 6: Spec status, PR, deploy runbook
+
+**Files:**
+- Modify: `docs/superpowers/specs/2026-07-09-gias-code-dictionaries-design.md` (status line)
+
+- [ ] **Step 1: Mark the spec implemented**
+
+Change `**Status:** Approved design` to `**Status:** Implemented 2026-07-09 — see docs/superpowers/plans/2026-07-09-gias-code-dictionaries.md`.
+
+- [ ] **Step 2: Commit and push**
+
+```bash
+git add docs/superpowers/specs/2026-07-09-gias-code-dictionaries-design.md
+git commit -m "docs: mark GIAS code dictionaries spec implemented
+
+Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>"
+git push -u origin feat/gias-code-dictionaries
+```
+
+- [ ] **Step 3: Open the PR (Gitea API via git credential fill — token-header auth 401s)**
+
+Title: `feat: GIAS classification fields stored as codes, translated in code`
+Body must include: (1) API contract unchanged — names still served, translation at the query boundary; (2) the **deploy runbook: merge → deploy → trigger `school_data_daily` immediately** (accepted empty-API window until the marts rebuild — spec §7); (3) dictionary maintenance loop (dbt drift test warns → rerun `generate_gias_codes.py` → commit regenerated files); (4) no frontend/e2e changes. End with the standard generation footer.
+
+- [ ] **Step 4: Watch CI**
+
+All PR checks must pass. Do not merge — merging triggers the deploy window; the human runs the runbook.