diff --git a/.github/scripts/sql_dialect_linter.sh b/.github/scripts/sql_dialect_linter.sh
new file mode 100644
index 00000000..32f1a7e7
--- /dev/null
+++ b/.github/scripts/sql_dialect_linter.sh
@@ -0,0 +1,291 @@
+#!/usr/bin/env bash
+
+# Usage: lint_dbt.sh <dialect>   where dialect is `databricks` or `snowflake`.
+
+set -euo pipefail
+
+# Which warehouse are we pretending to be: databricks or snowflake?
+dialect="${1:-}"
+case "$dialect" in
+  databricks|snowflake) ;;
+  *) echo "usage: $0 <databricks|snowflake>" >&2; exit 2 ;;
+esac
+
+# Read the project's name so we can build a fake connection for it below.
+profile=$(grep "^profile:" dbt_project.yml | awk '{print $2}' | tr -d "'\"")
+# Keep each dialect's compiled output in its own folder so they don't overwrite each other.
+target_path="target/lint-${dialect}"
+profiles_dir="${target_path}/dummy_profile"
+mkdir -p "$profiles_dir"
+
+# per-dialect dummy profile.
+if [[ "$dialect" == "databricks" ]]; then
+  cat > "$profiles_dir/profiles.yml" << EOF
+$profile:
+  target: dry_run
+  outputs:
+    dry_run:
+      type: databricks
+      host: 127.0.0.1
+      http_path: /sql/1.0/warehouses/dummy
+      token: dummy
+      schema: dummy
+      catalog: dummy
+      threads: 8
+      connect_timeout: 1
+      connect_retries: 0
+      retry_all: false
+      connection_parameters:
+        _retry_stop_after_attempts_count: 1
+        _socket_timeout: 3
+EOF
+else
+  cat > "$profiles_dir/profiles.yml" << EOF
+$profile:
+  target: dry_run
+  outputs:
+    dry_run:
+      type: snowflake
+      account: dummy-account
+      user: dummy
+      password: dummy
+      role: dummy
+      database: dummy
+      warehouse: dummy
+      schema: dummy
+      threads: 8
+      connect_retries: 0
+      connect_timeout: 3
+      retry_all: false
+EOF
+fi
+
+# Each of these normally hits the real warehouse during compile, so replace
+# them with fake versions that just return an empty result:
+#   get_filtered_columns_in_relation / unpivot -> list a real table's columns
+#   get_column_values                          -> select distinct values
+#   get_intervals_between                      -> a real date-diff query
+#   is_incremental                             -> checks if the table exists
+#   union_relations                            -> also lists a table's columns
+cat > macros/_ci_lint_stub.sql << EOF
+{% macro ${dialect}__get_filtered_columns_in_relation(from, except=[]) %}
+  {{ return([]) }}
+{% endmacro %}
+
+{% macro ${dialect}__union_relations(relations, column_override=none, include=[], exclude=[], source_column_name='_dbt_source_relation', where=none) %}
+  {{ return('') }}
+{% endmacro %}
+
+{% macro ${dialect}__get_column_values(table, column, order_by='count(*) desc', max_records=none, default=none, where=none) %}
+  {{ return(default if default is not none else []) }}
+{% endmacro %}
+
+{% macro is_incremental() %}
+  {{ return(False) }}
+{% endmacro %}
+
+{% macro ${dialect}__get_intervals_between(start_date, end_date, datepart) %}
+  {{ return(1) }}
+{% endmacro %}
+
+{% macro ${dialect}__unpivot(relation=none, cast_to='varchar', exclude=none, remove=none, field_name='field_name', value_name='value', quote_identifiers=False) %}
+  select
+    {% for col in (exclude or []) %}
+    cast(null as {{ cast_to }}) as {{ col }},
+    {% endfor %}
+    cast(null as {{ cast_to }}) as {{ field_name }},
+    cast(null as {{ cast_to }}) as {{ value_name }}
+  where false
+{% endmacro %}
+
+{#- edu_edfi_source's databricks__json_flatten emits "lateral variant_explode(...)",
+but sqlfluff's databricks dialect can't parse this for some reason, even though it
+is an actual Databricks call. When a where clause follows directly after this variant_explode,
+sqlfluff errors. Adding trailing comma to force sqlfluff to look at this line instead of the
+where clause. This redefines the same macro as macros/json_flatten.sql below, which is fine
+since that file only keeps the dispatch + snowflake__ versions #}
+{% macro databricks__json_flatten(column, alias, outer) -%}
+, lateral variant_explode{% if outer %}_outer{% endif %}({{ column }}) {% if alias != '' %} as {{ alias }} {% endif %}, -- noqa: PRS
+{%- endmacro %}
+
+EOF
+
+# json_flatten's databricks__ implementation is stubbed above (to work around
+# the sqlfluff parse issue), so drop it here to avoid a duplicate macro
+# definition; keep the dispatch macro and the snowflake__ implementation.
+cat > macros/json_flatten.sql << 'EOF'
+{% macro json_flatten(column, alias='', outer=False) %}
+    {{ return(adapter.dispatch('json_flatten', 'edu_edfi_source')(column, alias, outer)) }}
+{% endmacro %}
+
+{% macro snowflake__json_flatten(column, alias, outer) -%}
+, lateral flatten(input=>{{ column }}, outer=>{{ outer }}) {% if alias != '' %} as {{ alias }} {% endif %}
+{%- endmacro %}
+EOF
+
+# extract_descriptor also runs a real query against the warehouse (via
+# run_query), which can't use a dummy the same way as above. So replace its
+# whole file with a simple version that skips the query and just returns
+# the raw descriptor code.
+cat > macros/extract_descriptor.sql << 'EOF'
+{% macro extract_descriptor(col, descriptor_name=None) -%}
+  split_part({{ col }}, '#', -1)
+{%- endmacro %}
+EOF
+
+# Generate a placeholder covering every resource name actually referenced,
+# with a fake database/schema, so source() resolves without needing a real one.
+resources=$(grep -rhoE "source_edfi3\(\s*['\"][a-zA-Z0-9_]+['\"]" \
+    models macros 2>/dev/null \
+    | sed -E "s/source_edfi3\(\s*['\"]//; s/['\"]//" | sort -u)
+
+{
+  echo "sources:"
+  echo "  - name: raw_edfi_3"
+  echo "    database: dummy"
+  echo "    schema: dummy_raw"
+  echo "    tables:"
+  echo "      - name: _deletes"
+  for r in $resources; do
+    echo "      - name: $r"
+  done
+} > models/_ci_lint_sources.yml
+
+# Generate an empty placeholder seed for every ref() that isn't defined
+# anywhere else, so ref() resolves without needing the real crosswalk data.
+mkdir -p seeds
+{
+  find . -iname "*.sql" -path "*/models/*" 2>/dev/null | sed -E 's#.*/##; s/\.sql$//'
+  find . -iname "*.csv" -path "*/seeds/*" 2>/dev/null | sed -E 's#.*/##; s/\.csv$//'
+} | sort -u > /tmp/_ci_defined_nodes.txt
+
+grep -rhoE "ref\(['\"][a-zA-Z0-9_]+['\"]" . --include="*.sql" 2>/dev/null \
+  | sed -E "s/ref\(['\"]//; s/['\"]//" | sort -u > /tmp/_ci_all_refs.txt
+
+for missing in $(comm -23 /tmp/_ci_all_refs.txt /tmp/_ci_defined_nodes.txt); do
+  printf 'placeholder\nx\n' > "seeds/${missing}.csv"
+done
+
+# Some models talk to the warehouse directly for real data or column info,
+# so they can't compile against a fake connection. Any OTHER kind of failure (a real bug, a missing
+# config) stops the script.
+#
+# tpdm, tpdmcommunity, and sedm staging models are disabled by default.
+# turn them on so they get linted too.
+models_needing_warehouse=()
+tests_needing_warehouse=()
+compile_log=$(mktemp)
+trap 'rm -f "$compile_log"' EXIT
+
+while true; do
+  exclude_flags=()
+  all_excluded=("${models_needing_warehouse[@]}" "${tests_needing_warehouse[@]}")
+  [[ ${#all_excluded[@]} -gt 0 ]] && exclude_flags=(--exclude "${all_excluded[@]}")
+
+  dbt compile --no-introspect --no-populate-cache \
+    --select package:edu_edfi_source \
+    "${exclude_flags[@]}" \
+    --vars '{"src:domain:tpdm:enabled": true, "src:domain:tpdmcommunity:enabled": true, "src:domain:sedm:enabled": true}' \
+    --profiles-dir "$profiles_dir" --target dry_run --target-path "$target_path" \
+    > "$compile_log" 2>&1 && break
+
+  # dbt's error looks like "Runtime Error in model some_model_name (path/to/file.sql)"
+  # or "...in test some_test_name (...)". Pull out both the node type and its name.
+  culprit=""
+  culprit_type=""
+  grep -q "connection never acquired for thread" "$compile_log" \
+    && culprit_type=$(grep -oE "Runtime Error in (model|test)" "$compile_log" | head -1 | awk '{print $NF}') \
+    && culprit=$(grep -oE "Runtime Error in (model|test) [a-zA-Z0-9_]+" "$compile_log" | head -1 | awk '{print $NF}')
+
+  if [[ -z "$culprit" ]] || [[ " ${all_excluded[*]} " == *" $culprit "* ]]; then
+    echo "❌ dbt compile failed:"
+    cat "$compile_log"
+    exit 1
+  fi
+
+  echo "⚠️ $culprit needs a live warehouse connection to compile, adding $culprit to the exclude list and recompiling"
+  if [[ "$culprit_type" == "test" ]]; then
+    tests_needing_warehouse+=("$culprit")
+  else
+    models_needing_warehouse+=("$culprit")
+  fi
+done
+
+# sqlfluff only catches syntax that doesn't parse, it can't tell that a
+# function or type doesn't exist in that dialect, so check for
+# these known Databricks gaps by name instead. Add new ones here as they come up.
+databricks_incompatible_patterns=(
+  'try_to_date\s*\('  # function has no Databricks equivalent
+  '\bas\s+time\b'     # Databricks has no TIME datatype
+)
+incompatible_regex=$(IFS='|'; echo "${databricks_incompatible_patterns[*]}")
+
+# Lint each compiled model/test. Generic tests (auto-generated from
+# schema .yml files, like unique/not_null) compile into a "<node>.yml/"
+mapfile -d '' -t compiled_files < <(find "$target_path/compiled" -path "*/edu_edfi_source/models/*" -name "*.sql" -print0)
+model_total=0; model_pass=0; model_fail=0; model_failed=()
+test_total=0; test_pass=0; test_fail=0; test_failed=()
+for f in "${compiled_files[@]}"; do
+  name=$(basename "$f")
+  path=${f#*/edu_edfi_source/}
+  is_test=false
+  [[ "$path" == *.yml/* ]] && is_test=true
+
+  if out=$(sqlfluff lint --config .sqlfluff --templater raw --dialect "$dialect" "$f" 2>&1); then
+    sqlfluff_ok=0
+  else
+    sqlfluff_ok=1
+  fi
+
+  bad_function=""
+  if [[ "$dialect" == "databricks" ]]; then
+    match=$(grep -inE "$incompatible_regex" "$f" | head -1 || true)
+    [[ -n "$match" ]] && bad_function="No Databricks equivalent: $match"
+  fi
+
+  if [[ $sqlfluff_ok -eq 0 && -z "$bad_function" ]]; then
+    echo "Linting $path ✅"
+    if $is_test; then test_pass=$((test_pass + 1)); else model_pass=$((model_pass + 1)); fi
+  else
+    printf '::group::Linting %s ❌\n' "$path"
+    [[ -n "$bad_function" ]] && echo "$bad_function"
+    echo "$out"
+    echo "::endgroup::"
+    if $is_test; then
+      test_fail=$((test_fail + 1)); test_failed+=("${name%.sql}")
+    else
+      model_fail=$((model_fail + 1)); model_failed+=("${name%.sql}")
+    fi
+  fi
+  if $is_test; then test_total=$((test_total + 1)); else model_total=$((model_total + 1)); fi
+done
+
+echo ""
+model_known=$((model_total + ${#models_needing_warehouse[@]}))
+printf '✅ %d/%d models compatible with %s\n' "$model_pass" "$model_known" "$dialect"
+if [[ ${#models_needing_warehouse[@]} -gt 0 ]]; then
+  printf '⚠️ %d of those %d cannot be compiled here and requires live warehouse, lint these models locally instead:\n' "${#models_needing_warehouse[@]}" "$model_known"
+  printf '    - %s\n' "${models_needing_warehouse[@]}"
+fi
+
+echo ""
+test_known=$((test_total + ${#tests_needing_warehouse[@]}))
+printf '✅ %d/%d tests compatible with %s\n' "$test_pass" "$test_known" "$dialect"
+if [[ ${#tests_needing_warehouse[@]} -gt 0 ]]; then
+  printf '⚠️ %d of those %d cannot be compiled here and requires live warehouse, lint these tests locally instead:\n' "${#tests_needing_warehouse[@]}" "$test_known"
+  printf '    - %s\n' "${tests_needing_warehouse[@]}"
+fi
+if [[ ${#test_failed[@]} -gt 0 ]]; then
+  printf '❌ %d tests are NOT compatible with %s:\n' "${#test_failed[@]}" "$dialect"
+  printf '    - %s\n' "${test_failed[@]}"
+fi
+
+if [[ ${#model_failed[@]} -eq 0 && ${#test_failed[@]} -eq 0 ]]; then
+  exit 0
+fi
+echo ""
+if [[ ${#model_failed[@]} -gt 0 ]]; then
+  printf '❌ %d models are NOT compatible with %s:\n' "${#model_failed[@]}" "$dialect"
+  printf '    - %s\n' "${model_failed[@]}"
+fi
+exit 1
diff --git a/.github/workflows/requirements/sql-linter.txt b/.github/workflows/requirements/sql-linter.txt
new file mode 100644
index 00000000..e777449b
--- /dev/null
+++ b/.github/workflows/requirements/sql-linter.txt
@@ -0,0 +1,7 @@
+# Pinned so CI and local dev always use the same versions.
+# Install with: pip install -r .github/workflows/requirements/sql-linter.txt
+sqlfluff==4.2.2
+sqlfluff-templater-dbt==4.2.2
+dbt-core==1.10.19
+dbt-databricks==1.10.19
+dbt-snowflake==1.10.8
diff --git a/.github/workflows/sql_dialect_linter.yml b/.github/workflows/sql_dialect_linter.yml
new file mode 100644
index 00000000..cf9867dd
--- /dev/null
+++ b/.github/workflows/sql_dialect_linter.yml
@@ -0,0 +1,60 @@
+name: sql_dialect_linter
+
+on:
+  pull_request:
+  # Adds a manual "Run workflow" button in the Actions tab. GitHub only
+  # shows that button after this file is merged to main.
+  workflow_dispatch:
+
+jobs:
+  lint:
+    name: sql_dialect_linter (${{ matrix.dialect }})
+    runs-on: ubuntu-latest
+    timeout-minutes: 20
+    strategy:
+      # Run both dialects even if one fails, so we get both results.
+      fail-fast: false
+      matrix:
+        dialect: [databricks, snowflake]
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+
+      # Skip reinstalling everything if nothing's changed since last time.
+      - name: Cache virtualenv
+        id: cache-venv
+        uses: actions/cache@v6
+        with:
+          path: .venv
+          key: venv-${{ runner.os }}-${{ hashFiles('.github/workflows/requirements/sql-linter.txt') }}
+
+      # Installs the exact versions pinned in sql-linter.txt. Skipped
+      # entirely when the cache above already has a matching virtualenv.
+      - name: Install dbt + sqlfluff
+        if: steps.cache-venv.outputs.cache-hit != 'true'
+        run: |
+          python -m venv .venv
+          .venv/bin/pip install -r .github/workflows/requirements/sql-linter.txt
+
+      - name: Put virtualenv on PATH
+        run: echo "$PWD/.venv/bin" >> "$GITHUB_PATH"
+
+      # Skips the dbt deps step below entirely if packages.yml hasn't changed.
+      - name: Cache dbt packages
+        id: cache-dbt-packages
+        uses: actions/cache@v6
+        with:
+          path: dbt_packages
+          key: dbt-packages-${{ hashFiles('packages.yml') }}
+
+      # Downloads edu_edfi_source and dbt_utils, unless the cache above
+      # already has them.
+      - name: dbt deps
+        if: steps.cache-dbt-packages.outputs.cache-hit != 'true'
+        run: dbt deps
+
+      - name: Compile and lint (${{ matrix.dialect }})
+        run: bash .github/scripts/sql_dialect_linter.sh ${{ matrix.dialect }}
diff --git a/.sqlfluff b/.sqlfluff
new file mode 100644
index 00000000..8a7113b6
--- /dev/null
+++ b/.sqlfluff
@@ -0,0 +1,19 @@
+[sqlfluff]
+# global default
+templater = dbt
+dialect = snowflake
+exclude_rules = layout, convention, capitalisation, aliasing, references, ambiguous, structure
+
+# Settings for --templater dbt
+[sqlfluff:templater:dbt]
+# If dbt compilation fails on a model, skip it rather than erroring out.
+dbt_skip_compilation_error = True
+
+[sqlfluff:templater:jinja]
+apply_dbt_builtins = true
+load_macros_from_path = macros
+
+# Stubs only for dbt_utils macros, those live in dbt_packages
+[sqlfluff:templater:jinja:macros]
+generate_surrogate_key = {% macro generate_surrogate_key(fields) %}null{% endmacro %}
+deduplicate            = {% macro deduplicate(relation, partition_by, order_by) %}select * from stub_relation{% endmacro %}
diff --git a/models/staging/edfi_3/stage/stg_ef3__staff_school_associations.sql b/models/staging/edfi_3/stage/stg_ef3__staff_school_associations.sql
index 1d109e7a..dac0f490 100644
--- a/models/staging/edfi_3/stage/stg_ef3__staff_school_associations.sql
+++ b/models/staging/edfi_3/stage/stg_ef3__staff_school_associations.sql
@@ -10,7 +10,7 @@ keyed as (
         -- which means this table almost never contains year data, so we have to 
         -- add it back in where possible
         coalesce(school_year, api_year) as school_year,
-        {{ dbt_utils.star(from=ref('base_ef3__staff_school_associations'), except=['school_year']) }}
+        {{ edu_edfi_source.star(from=ref('base_ef3__staff_school_associations'), except=['school_year']) }}
         {{ extract_extension(model_name=this.name, flatten=True) }}
     from base_staff_school_assoc
 
diff --git a/models/staging/edfi_3/stage/stg_ef3__student_discipline_incident_behavior_associations.sql b/models/staging/edfi_3/stage/stg_ef3__student_discipline_incident_behavior_associations.sql
index 673e73b5..6b7b6f85 100644
--- a/models/staging/edfi_3/stage/stg_ef3__student_discipline_incident_behavior_associations.sql
+++ b/models/staging/edfi_3/stage/stg_ef3__student_discipline_incident_behavior_associations.sql
@@ -18,7 +18,7 @@ dedupe_base_student_discipline_incident  as (
 format_student_discipline_incident as (
     -- note: the deprecated model needs to be flattened to match the grain of the new model
     select 
-        {{ dbt_utils.star(ref('base_ef3__student_discipline_incident_associations'), 
+        {{ edu_edfi_source.star(ref('base_ef3__student_discipline_incident_associations'), 
             except=['student_participation_code', 'v_behaviors', 'v_ext', 'discipline_incident_reference', 'student_reference']) }},
             value:behaviorDetailedDescription::string as behavior_detailed_description,
         {{ extract_descriptor('value:behaviorDescriptor::string') }} as behavior_type,
diff --git a/models/staging/edfi_3/stage/stg_ef3__student_discipline_incident_non_offender_associations.sql b/models/staging/edfi_3/stage/stg_ef3__student_discipline_incident_non_offender_associations.sql
index 6041d97b..1cd3b551 100644
--- a/models/staging/edfi_3/stage/stg_ef3__student_discipline_incident_non_offender_associations.sql
+++ b/models/staging/edfi_3/stage/stg_ef3__student_discipline_incident_non_offender_associations.sql
@@ -17,7 +17,7 @@ dedupe_base_student_discipline_incident  as (
 -- projects should only ever have one of the two models in use
 format_student_discipline_incident as (
     select 
-        {{ dbt_utils.star(ref('base_ef3__student_discipline_incident_associations'), 
+        {{ edu_edfi_source.star(ref('base_ef3__student_discipline_incident_associations'), 
             except=['student_participation_code', 'v_behaviors', 'v_ext']) }},
         {{
             json_array_agg(
@@ -65,4 +65,4 @@ deduped as (
     }}
 )
 select * from deduped
-where not is_deleted
\ No newline at end of file
+where not is_deleted