diff --git a/Makefile b/Makefile
index 202ae52..d34a060 100644
--- a/Makefile
+++ b/Makefile
@@ -1,4 +1,5 @@
 .PHONY: kind tilt-up tilt-down seed seed-fetch seed-push recon evals test \
+        run-log run-logs \
         compose-up compose-down compose-clean compose-seed compose-recon compose-evals \
         docs-graphs
 
@@ -51,6 +52,23 @@ recon:
 evals:
 	kubectl $(KCTX) $(KNS) exec deploy/api -- uv run python -m api.evals.run_evals
 
+# Pull a run's JSONL transcript from the api pod to the host. The api writes
+# every published event into /app/.data/logs/<RUN>.jsonl inside the pod;
+# this target copies it next to .data/logs/ on the host so it's easy to
+# share / grep / paste a path to.
+# Usage: make run-log RUN=<run_id>
+run-log:
+	@[ -n "$(RUN)" ] || { echo "usage: make run-log RUN=<run_id>" >&2; exit 1; }
+	@mkdir -p .data/logs
+	@POD=$$(kubectl $(KCTX) $(KNS) get pod -l app=api -o jsonpath='{.items[0].metadata.name}'); \
+	kubectl $(KCTX) $(KNS) cp $$POD:/app/.data/logs/$(RUN).jsonl .data/logs/$(RUN).jsonl \
+		&& echo "→ .data/logs/$(RUN).jsonl"
+
+# List the runs the api currently has log files for (most recent first).
+run-logs:
+	@POD=$$(kubectl $(KCTX) $(KNS) get pod -l app=api -o jsonpath='{.items[0].metadata.name}'); \
+	kubectl $(KCTX) $(KNS) exec $$POD -- sh -c 'ls -1t /app/.data/logs 2>/dev/null || echo "(no logs yet)"'
+
 # Unit tests run on the host venv (no postgres/llm calls needed).
 # Ensures dev extras are installed first; uv is idempotent on no-ops.
 test:
diff --git a/api/composer/aliases.py b/api/composer/aliases.py
new file mode 100644
index 0000000..46aff10
--- /dev/null
+++ b/api/composer/aliases.py
@@ -0,0 +1,56 @@
+"""Alias allocation for the composer.
+
+Every table referenced in a composed query gets a short, query-unique alias.
+`AliasMap` owns that allocation so the rest of the composer never threads a
+bare dict around: ask it for a table's alias (allocating on first use) or for a
+qualified `exp.Column`, and it remembers the assignment.
+
+Allocation strategy (stable — golden SQL depends on it): first letter, then the
+first two/three letters, then the full name, then a numbered `<first><n>` on
+collision. So `loan→l`, `account→a`, `district→d`.
+"""
+
+from __future__ import annotations
+
+from sqlglot import exp
+
+
+def _alloc(table: str, used: set[str]) -> str:
+    """Pick an alias for `table` not already in `used`."""
+    for base in (table[:1], table[:2], table[:3], table):
+        if base and base not in used:
+            return base
+    i = 1
+    while True:
+        cand = f"{table[:1]}{i}"
+        if cand not in used:
+            return cand
+        i += 1
+
+
+class AliasMap:
+    """Maps table name → alias, allocating on first request."""
+
+    def __init__(self) -> None:
+        self._by_table: dict[str, str] = {}
+
+    def has(self, table: str) -> bool:
+        return table in self._by_table
+
+    def alias(self, table: str) -> str:
+        """Return `table`'s alias, allocating a fresh unique one if needed."""
+        if table not in self._by_table:
+            self._by_table[table] = _alloc(table, set(self._by_table.values()))
+        return self._by_table[table]
+
+    def col(self, table: str, column: str) -> exp.Column:
+        """A quoted, alias-qualified column reference for the final tree."""
+        return exp.column(column, table=self.alias(table), quoted=True)
+
+    def aliased_table(self, table: str) -> exp.Expression:
+        """`"<table>" AS "<alias>"` for FROM / JOIN."""
+        return exp.alias_(exp.to_table(table, quoted=True), self.alias(table), quoted=True)
+
+    def tables(self) -> list[str]:
+        """Sorted names of every table that has been referenced."""
+        return sorted(self._by_table)
diff --git a/api/composer/compose.py b/api/composer/compose.py
index d21b7b5..65557e2 100644
--- a/api/composer/compose.py
+++ b/api/composer/compose.py
@@ -1,191 +1,48 @@
 """compose(pick, recon) → ComposeResult.
 
-Deterministic SQL emission from a typed Pick. The composer:
-  1. Resolves the metric → from_table + sql expression + default filter.
-  2. Resolves each group_by column → owning table (via recon.resolve_column).
-  3. Computes the join graph: join_path(metric.from_table, owner) per dim,
-     deduped, walked in declared order to pick the right relationship edge.
-  4. Renders SELECT / FROM + JOINs / WHERE / GROUP BY / ORDER BY / LIMIT
-     with every identifier quoted via sqlglot's `identify=True` round-trip
-     as a paranoid post-check.
+Deterministic SQL emission from a typed Pick. The composer assembles a single
+sqlglot expression tree and renders it once — there is no hand-built SQL string
+and no after-the-fact re-parse; the tree *is* the structure, and
+`.sql(dialect="postgres", identify=True)` (every identifier quoted) is the one
+place a string is produced.
 
-No LLM call anywhere in this module. If the Pick can't be expressed against
-recon, the composer raises `PickValidationError` — the caller surfaces it.
+The work is split across focused modules, each returning `exp` nodes:
+
+  aliases.py   table → alias allocation (AliasMap)
+  joins.py     join-graph walk → JOIN specs
+  filters.py   typed Filter → predicate node
+  metrics.py   metric YAML fragment → alias-qualified expression
+  dates.py     date_range → BETWEEN node (encoding-aware)
+  order.py     OrderBy → Ordered node
+  encodings.py default + dataset storage-encoding library
+
+No LLM call anywhere. The Pick is the constraint boundary: if it can't be
+expressed against recon the composer raises `PickValidationError`, which the
+caller surfaces — it never fabricates SQL. New capability is additive: a new
+filter shape is a branch in filters.py; a new Pick `kind` is a builder here.
+Widening what a Pick can express is a deliberate decision — see
+`def/schema-as-constraint.md` before doing so.
 """
 
 from __future__ import annotations
 
-import sqlglot
+from sqlglot import exp
 
-from api.composer.dates import resolve_date_range
-from api.composer.types import (
-    ComposeResult,
-    Filter,
-    OrderBy,
-    Pick,
-    PickValidationError,
-)
-from api.recon.types import Recon, Relationship
+from api.composer.aliases import AliasMap
+from api.composer.encodings import effective_encodings
+from api.composer.filters import render_filter
+from api.composer.joins import build_joins
+from api.composer.metrics import qualify_metric_expr
+from api.composer.order import default_order_by, render_order_by
+from api.composer.types import ComposeResult, Pick, PickValidationError
+from api.recon.types import Recon
 
 
-# ── Alias allocation ────────────────────────────────────────
-
-def _alias_for(table: str, used: dict[str, str]) -> str:
-    """Pick a short alias for `table` that's unique within the query.
-
-    First letter, then first-two, then numbered. `used` maps alias→table so
-    we can grow on collision.
-    """
-    for base in (table[:1], table[:2], table[:3], table):
-        if base and base not in used:
-            return base
-    i = 1
-    while True:
-        cand = f"{table[:1]}{i}"
-        if cand not in used:
-            return cand
-        i += 1
-
-
-# ── Join graph ──────────────────────────────────────────────
-
-def _find_edge(recon: Recon, a: str, b: str) -> Relationship:
-    """Return the relationship that joins tables a and b (either direction).
-    Raises if no declared FK connects them."""
-    for r in recon.relationships:
-        if (r.from_table == a and r.to_table == b) or (r.from_table == b and r.to_table == a):
-            return r
-    raise PickValidationError(
-        f"no declared relationship between {a!r} and {b!r}"
-    )
-
-
-def _build_joins(recon: Recon, base: str, targets: list[str],
-                 aliases: dict[str, str]) -> list[str]:
-    """Compute the union of join paths from `base` to each target table,
-    emit JOIN clauses in walk order. Allocates aliases for any intermediate
-    table that wasn't already in `aliases`."""
-    visited: set[str] = {base}
-    clauses: list[str] = []
-    for target in targets:
-        path = recon.join_path(base, target)
-        if path is None:
-            raise PickValidationError(
-                f"no join path from {base!r} to {target!r} in recon"
-            )
-        for i in range(1, len(path)):
-            prev, cur = path[i - 1], path[i]
-            if cur in visited:
-                continue
-            edge = _find_edge(recon, prev, cur)
-            # Edge direction tells us which side has which column.
-            if edge.from_table == prev:
-                left_t, left_c, right_t, right_c = edge.from_table, edge.from_column, edge.to_table, edge.to_column
-            else:
-                left_t, left_c, right_t, right_c = edge.to_table, edge.to_column, edge.from_table, edge.from_column
-            # Ensure both sides have aliases.
-            for t in (left_t, right_t):
-                if t not in aliases:
-                    aliases[t] = _alias_for(t, {a: t for t, a in aliases.items()})
-            la, ra = aliases[left_t], aliases[right_t]
-            clauses.append(
-                f'JOIN "{cur}" AS "{aliases[cur]}" '
-                f'ON "{la}"."{left_c}" = "{ra}"."{right_c}"'
-            )
-            visited.add(cur)
-    return clauses
-
-
-# ── Filter rendering ────────────────────────────────────────
-
-def _render_literal(v) -> str:
-    """Render a Python value as a SQL literal. Strings are single-quoted with
-    embedded quotes escaped. Numbers pass through."""
-    if isinstance(v, str):
-        return "'" + v.replace("'", "''") + "'"
-    if isinstance(v, bool):
-        return "TRUE" if v else "FALSE"
-    if v is None:
-        return "NULL"
-    return str(v)
-
-
-def _render_filter(f: Filter, recon: Recon, aliases: dict[str, str]) -> str:
-    """Render one Filter to a SQL predicate. Resolves the column's owning
-    table and aliases it (allocating an alias if needed — and a join later
-    if that introduces a new table)."""
-    table, col = recon.resolve_column(f.column)
-    if table not in aliases:
-        aliases[table] = _alias_for(table, {a: t for t, a in aliases.items()})
-    qualified = f'"{aliases[table]}"."{col.name}"'
-
-    kind = f.kind()
-    if kind == "equals":
-        return f"{qualified} = {_render_literal(f.equals)}"
-    if kind == "in_values":
-        if not f.in_values:
-            raise PickValidationError(
-                f"filter on {f.column!r} has empty in_values"
-            )
-        rendered = ", ".join(_render_literal(v) for v in f.in_values)
-        return f"{qualified} IN ({rendered})"
-    if kind == "between":
-        lo, hi = f.between
-        return f"{qualified} BETWEEN {_render_literal(lo)} AND {_render_literal(hi)}"
-    if kind == "date_range":
-        return resolve_date_range(f.date_range, col, qualified)
-    raise PickValidationError(f"unknown filter kind {kind!r}")
-
-
-# ── ORDER BY rendering ──────────────────────────────────────
-
-def _render_order_by(ob: OrderBy, metric_name: str, dim_select: dict[str, str]) -> str:
-    """`metric_name` is the alias of the metric SELECT expression;
-    `dim_select` maps dimension ref (column ref) → output alias."""
-    if ob.by == "metric":
-        return f'"{metric_name}" {ob.direction.upper()}'
-    # by == "dimension"
-    if ob.dimension not in dim_select:
-        raise PickValidationError(
-            f"order_by.dimension={ob.dimension!r} is not in group_by ({list(dim_select)})"
-        )
-    return f'"{dim_select[ob.dimension]}" {ob.direction.upper()}'
-
-
-# ── Metric expression rewriting ─────────────────────────────
-
-def _qualify_metric_sql(sql_fragment: str, table_alias: str) -> str:
-    """Rewrite bare column refs inside the metric's SQL expression so they
-    point at the metric table's alias.
-
-    Metric.sql in metrics.yaml is written as an unaliased expression (e.g.
-    `AVG(CASE WHEN status='B' ... END)`). We need it qualified to the
-    metric table's alias so the composer can introduce other tables via
-    joins without ambiguity. Uses sqlglot to parse and rewrite identifier
-    refs that don't already have a table prefix.
-    """
-    tree = sqlglot.parse_one(sql_fragment, dialect="postgres")
-    for col in tree.find_all(sqlglot.exp.Column):
-        if col.table:
-            continue
-        # Inject the alias as the table qualifier.
-        col.set("table", sqlglot.exp.Identifier(this=table_alias, quoted=True))
-    return tree.sql(dialect="postgres", identify=True)
-
-
-def _qualify_metric_filter(filter_fragment: str, table_alias: str) -> str:
-    """Same as _qualify_metric_sql but for the metric's optional WHERE filter.
-    metric.filter is written as a bare boolean expression."""
-    return _qualify_metric_sql(filter_fragment, table_alias)
-
-
-# ── Compose ─────────────────────────────────────────────────
-
 def compose(pick: Pick, recon: Recon) -> ComposeResult:
-    """Render `pick` to SQL against `recon`. Raises PickValidationError on
-    any reference the recon can't resolve."""
+    """Render `pick` to SQL against `recon`. Raises PickValidationError on any
+    reference the recon can't resolve."""
 
-    # 1. Metric.
+    # 1. Metric → base table.
     if pick.metric not in recon.metrics:
         raise PickValidationError(
             f"metric {pick.metric!r} not in recon (known: {sorted(recon.metrics)})"
@@ -197,92 +54,73 @@ def compose(pick: Pick, recon: Recon) -> ComposeResult:
             f"but no such table in recon"
         )
 
-    aliases: dict[str, str] = {}
-    aliases[metric.from_table] = _alias_for(metric.from_table, {})
-    base_alias = aliases[metric.from_table]
+    aliases = AliasMap()
+    base_alias = aliases.alias(metric.from_table)
+    encodings = effective_encodings(recon)
 
-    # 2. Resolve every group_by column to (table, Column) and collect target
-    #    tables that aren't the metric's table — those need joins.
-    group_bindings: list[tuple[str, str, str, str]] = []  # (ref, table, col_name, output_alias)
+    # 2. Resolve group_by columns → owning table + output alias. Tables other
+    #    than the metric's need joins.
+    group_bindings: list[tuple[str, str, str]] = []  # (table, col_name, out_alias)
     extra_tables: list[str] = []
     dim_select_alias: dict[str, str] = {}
     for ref in pick.group_by:
         table, col = recon.resolve_column(ref)
-        if table not in aliases:
+        if not aliases.has(table):
             if table != metric.from_table:
                 extra_tables.append(table)
-            aliases[table] = _alias_for(table, {a: t for t, a in aliases.items()})
+            aliases.alias(table)
         out_alias = col.name if ref == col.name else ref.replace(".", "_")
-        group_bindings.append((ref, table, col.name, out_alias))
+        group_bindings.append((table, col.name, out_alias))
         dim_select_alias[ref] = out_alias
 
-    # 3. Resolve filter columns first (they may also introduce new tables we
-    #    need to join). Rendered separately so we can keep their predicates
-    #    in WHERE.
-    filter_clauses: list[str] = []
+    # 3. Resolve filters → predicate nodes (may also introduce new tables).
+    filter_nodes: list[exp.Expression] = []
     for f in pick.where:
         table, _ = recon.resolve_column(f.column)
-        if table not in aliases:
+        if not aliases.has(table):
             if table != metric.from_table:
                 extra_tables.append(table)
-            aliases[table] = _alias_for(table, {a: t for t, a in aliases.items()})
-        filter_clauses.append(_render_filter(f, recon, aliases))
+            aliases.alias(table)
+        filter_nodes.append(render_filter(f, recon, aliases, encodings))
 
-    # 4. Build JOINs for the union of extra tables.
-    join_clauses = _build_joins(recon, metric.from_table, extra_tables, aliases)
+    # 4. JOINs for the union of extra tables.
+    join_specs = build_joins(recon, metric.from_table, extra_tables, aliases)
 
-    # 5. SELECT list: dimension columns first (in group_by order), then the
-    #    metric expression.
-    select_parts: list[str] = []
-    for ref, table, col_name, out_alias in group_bindings:
-        select_parts.append(f'"{aliases[table]}"."{col_name}" AS "{out_alias}"')
-    metric_expr = _qualify_metric_sql(metric.sql, base_alias)
-    select_parts.append(f'{metric_expr} AS "{metric.name}"')
+    # 5. SELECT list: dimensions (in group_by order), then the metric.
+    select_nodes: list[exp.Expression] = [
+        exp.alias_(aliases.col(table, col_name), out_alias, quoted=True)
+        for table, col_name, out_alias in group_bindings
+    ]
+    select_nodes.append(
+        exp.alias_(qualify_metric_expr(metric.sql, base_alias), metric.name, quoted=True)
+    )
 
     # 6. WHERE: metric's default filter (if any) ANDed with the Pick's filters.
-    where_parts: list[str] = []
+    where_nodes: list[exp.Expression] = []
     if metric.filter:
-        where_parts.append(_qualify_metric_filter(metric.filter, base_alias))
-    where_parts.extend(filter_clauses)
+        where_nodes.append(qualify_metric_expr(metric.filter, base_alias))
+    where_nodes.extend(filter_nodes)
 
-    # 7. GROUP BY (just the dim output aliases) + ORDER BY + LIMIT.
-    group_by_clause = ""
-    if group_bindings:
-        group_by_clause = "GROUP BY " + ", ".join(
-            f'"{out_alias}"' for _, _, _, out_alias in group_bindings
-        )
-
-    order_by_clause = ""
+    # 7. ORDER BY: explicit, else largest-metric-first for grouped queries.
     if pick.order_by is not None:
-        order_by_clause = "ORDER BY " + _render_order_by(
-            pick.order_by, metric.name, dim_select_alias
-        )
+        order_node = render_order_by(pick.order_by, metric.name, dim_select_alias)
     elif group_bindings:
-        # Sensible default: largest metric first.
-        order_by_clause = f'ORDER BY "{metric.name}" DESC'
+        order_node = default_order_by(metric.name)
+    else:
+        order_node = None
 
-    limit_clause = f"LIMIT {pick.limit}" if pick.limit is not None else ""
+    # 8. Assemble one Select and render once.
+    sel = exp.select(*select_nodes).from_(aliases.aliased_table(metric.from_table))
+    for table_expr, on_expr in join_specs:
+        sel = sel.join(table_expr, on=on_expr, join_type="")
+    for node in where_nodes:
+        sel = sel.where(node)
+    if group_bindings:
+        sel = sel.group_by(*(exp.column(oa, quoted=True) for _, _, oa in group_bindings))
+    if order_node is not None:
+        sel = sel.order_by(order_node)
+    if pick.limit is not None:
+        sel = sel.limit(pick.limit)
 
-    # 8. Assemble.
-    sql_parts = [
-        "SELECT " + ", ".join(select_parts),
-        f'FROM "{metric.from_table}" AS "{base_alias}"',
-    ]
-    sql_parts.extend(join_clauses)
-    if where_parts:
-        sql_parts.append("WHERE " + " AND ".join(where_parts))
-    if group_by_clause:
-        sql_parts.append(group_by_clause)
-    if order_by_clause:
-        sql_parts.append(order_by_clause)
-    if limit_clause:
-        sql_parts.append(limit_clause)
-    raw = "\n".join(sql_parts)
-
-    # 9. Paranoid re-parse: catches composer bugs by round-tripping through
-    #    sqlglot. identify=True keeps every identifier quoted.
-    parsed = sqlglot.parse_one(raw, dialect="postgres")
-    sql = parsed.sql(dialect="postgres", identify=True)
-
-    used = sorted(aliases)
-    return ComposeResult(sql=sql, used_tables=used)
+    sql = sel.sql(dialect="postgres", identify=True)
+    return ComposeResult(sql=sql, used_tables=aliases.tables())
diff --git a/api/composer/dates.py b/api/composer/dates.py
index d870160..18256aa 100644
--- a/api/composer/dates.py
+++ b/api/composer/dates.py
@@ -1,22 +1,29 @@
 """Date-range resolver for typed Filters.
 
-Given a `Filter` with `date_range="YYYY"` or `date_range="YYYY-Q[1-4]"` and
-the target column, render a SQL fragment that evaluates correctly under the
-column's actual storage type. Dispatched in two layers:
+Given a `Filter` with `date_range="YYYY"` or `date_range="YYYY-Q[1-4]"`, the
+target column, and the active encoding map, build a SQL predicate node that
+evaluates correctly under the column's actual storage. Dispatched in two layers:
 
-1. `Column.semantic_type` if set (extension point — e.g. `"date_yymmdd"` for
-   datasets that ship YYMMDD-encoded integers).
-2. Otherwise `Column.sql_type` (`DATE`, `TIMESTAMP`, …) — the default path
-   for warehouses that store dates as real date columns.
+1. `Column.semantic_type` if it names an encoding in the effective map — coerce
+   the raw column to a DATE via the encoding's `as_date` template, then compare.
+   This is the extension point: BIRD's `date_yymmdd` ships as a default
+   (see `api.composer.encodings`); a dataset can add or override encodings in
+   its own `schema_docs.yaml`.
+2. Otherwise `Column.sql_type` (`DATE`, `TIMESTAMP`, …) — compare the column
+   directly.
 
-Inclusive on both ends. Bounds quoted with single quotes (Postgres parses
-date literals from strings).
+Returns an `exp.Between` node (inclusive on both ends) so it composes into the
+query tree the composer assembles; the composer is the single place a SQL
+string is produced.
 """
 
 from __future__ import annotations
 
 import re
 
+import sqlglot
+from sqlglot import exp
+
 from api.composer.types import PickValidationError
 from api.recon.types import Column
 
@@ -47,27 +54,37 @@ def _bounds(spec: str) -> tuple[str, str]:
     )
 
 
-def resolve_date_range(spec: str, column: Column, qualified_col: str) -> str:
-    """Render the WHERE fragment for `<qualified_col> BETWEEN ...`.
+def resolve_date_range(
+    spec: str,
+    column: Column,
+    col_expr: exp.Expression,
+    encodings: dict[str, dict[str, str]],
+) -> exp.Between:
+    """Build the `<comparable> BETWEEN '<lo>' AND '<hi>'` predicate for `spec`.
 
-    `qualified_col` is the already-quoted reference the composer wants in
-    the SQL (e.g. `"l"."date"`). Returns just the predicate, no `WHERE`.
+    `col_expr` is the (already alias-qualified) column reference node;
+    `encodings` is the effective encoding map (see `api.composer.encodings`).
     """
     lo, hi = _bounds(spec)
-    semantic = (column.semantic_type or "").lower()
+    bounds = (exp.convert(lo), exp.convert(hi))
 
-    if semantic == "date_yymmdd":
-        # Column stores YYMMDD as integer (some BIRD datasets ship this way).
-        return (
-            f"TO_DATE(LPAD({qualified_col}::text, 6, '0'), 'YYMMDD') "
-            f"BETWEEN '{lo}' AND '{hi}'"
+    semantic = (column.semantic_type or "").lower()
+    enc = encodings.get(semantic) if semantic else None
+    if enc and enc.get("as_date"):
+        # Coerce the raw column to a DATE via the dataset/default template.
+        template = enc["as_date"]
+        coerced = sqlglot.parse_one(
+            template.format(col=col_expr.sql(dialect="postgres", identify=True)),
+            dialect="postgres",
         )
+        return exp.Between(this=coerced, low=bounds[0], high=bounds[1])
 
     sql_type = column.sql_type.upper()
     if sql_type.startswith("DATE") or sql_type.startswith("TIMESTAMP"):
-        return f"{qualified_col} BETWEEN '{lo}' AND '{hi}'"
+        return exp.Between(this=col_expr, low=bounds[0], high=bounds[1])
 
     raise PickValidationError(
         f"date_range on column with sql_type={column.sql_type!r} and "
-        f"semantic_type={column.semantic_type!r} is not supported"
+        f"semantic_type={column.semantic_type!r} is not supported "
+        f"(no native date type and no matching encoding)"
     )
diff --git a/api/composer/encodings.py b/api/composer/encodings.py
new file mode 100644
index 0000000..54e5225
--- /dev/null
+++ b/api/composer/encodings.py
@@ -0,0 +1,42 @@
+"""Storage-encoding library — how a column's raw storage maps to a value the
+composer can compare against.
+
+Some warehouses store logical types in surprising physical encodings (BIRD
+ships YYMMDD-encoded *integer* "dates", for instance). The composer needs a way
+to coerce such a column to a comparable value without baking any one dataset's
+quirk into the generic SQL builder.
+
+The model is **defaults with override**:
+
+  - `DEFAULT_ENCODINGS` ships the generic, reusable patterns here — keyed by the
+    `Column.semantic_type` a dataset declares in `schema_docs.yaml`.
+  - A dataset may add a novel encoding or override a default via the optional
+    `encodings:` section of its `schema_docs.yaml` (carried onto
+    `Recon.encodings`). The dataset always wins.
+
+Each encoding is plain data: a map of role → SQL template with a `{col}`
+placeholder for the (already-qualified) column reference. Today the only role
+is `as_date` — the expression that yields a DATE for `date_range` filters; new
+roles are additive.
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+if TYPE_CHECKING:
+    from api.recon.types import Recon
+
+
+DEFAULT_ENCODINGS: dict[str, dict[str, str]] = {
+    # BIRD-style YYMMDD stored as a 6-digit integer → real DATE.
+    "date_yymmdd": {"as_date": "TO_DATE(LPAD({col}::text, 6, '0'), 'YYMMDD')"},
+    # Add further reusable patterns here (e.g. epoch_seconds, julian_day) as
+    # they recur across datasets — never a single dataset's one-off quirk.
+}
+
+
+def effective_encodings(recon: "Recon") -> dict[str, dict[str, str]]:
+    """The encoding map the composer should use for `recon`: built-in defaults
+    overlaid with the dataset's own `encodings` (dataset entries win)."""
+    return {**DEFAULT_ENCODINGS, **(recon.encodings or {})}
diff --git a/api/composer/filters.py b/api/composer/filters.py
new file mode 100644
index 0000000..e0be3de
--- /dev/null
+++ b/api/composer/filters.py
@@ -0,0 +1,46 @@
+"""Filter rendering — one typed `Filter` → one SQL predicate node.
+
+Each filter shape maps to a sqlglot expression node (`exp.EQ`, `exp.In`,
+`exp.Between`, or the date-range subtree). Literals go through `exp.convert`,
+so quoting and escaping are sqlglot's job, not ours. The column is resolved to
+its owning table and qualified via the shared `AliasMap` (allocating an alias,
+and implying a join later, if the filter introduces a new table).
+
+Adding a new filter shape is a new branch here that returns an `exp` node — no
+change to the assembly in `compose.py`. The `Filter` type stays the constraint
+boundary: there is no raw-SQL escape hatch.
+"""
+
+from __future__ import annotations
+
+from sqlglot import exp
+
+from api.composer.aliases import AliasMap
+from api.composer.dates import resolve_date_range
+from api.composer.types import Filter, PickValidationError
+from api.recon.types import Recon
+
+
+def render_filter(
+    f: Filter,
+    recon: Recon,
+    aliases: AliasMap,
+    encodings: dict[str, dict[str, str]],
+) -> exp.Expression:
+    """Render one `Filter` to a predicate node bound to the column's table."""
+    table, col = recon.resolve_column(f.column)
+    col_expr = aliases.col(table, col.name)
+
+    kind = f.kind()
+    if kind == "equals":
+        return exp.EQ(this=col_expr, expression=exp.convert(f.equals))
+    if kind == "in_values":
+        if not f.in_values:
+            raise PickValidationError(f"filter on {f.column!r} has empty in_values")
+        return exp.In(this=col_expr, expressions=[exp.convert(v) for v in f.in_values])
+    if kind == "between":
+        lo, hi = f.between
+        return exp.Between(this=col_expr, low=exp.convert(lo), high=exp.convert(hi))
+    if kind == "date_range":
+        return resolve_date_range(f.date_range, col, col_expr, encodings)
+    raise PickValidationError(f"unknown filter kind {kind!r}")
diff --git a/api/composer/joins.py b/api/composer/joins.py
new file mode 100644
index 0000000..435f018
--- /dev/null
+++ b/api/composer/joins.py
@@ -0,0 +1,66 @@
+"""Join-graph walking.
+
+The composer never lets the LLM choose joins. Given the metric's base table and
+the set of other tables a Pick pulled in (via group_by / filters), `build_joins`
+walks `recon.join_path` for each target, dedupes the visited tables, and emits a
+JOIN spec per new edge — direction resolved from the declared relationship so
+the ON clause names the right column on each side.
+
+Each spec is `(aliased_table_expr, on_expr)`: the composer applies them with
+`select.join(table, on=on, join_type="")`. Aliases for intermediate tables are
+allocated through the shared `AliasMap` as they're encountered.
+"""
+
+from __future__ import annotations
+
+from sqlglot import exp
+
+from api.composer.aliases import AliasMap
+from api.composer.types import PickValidationError
+from api.recon.types import Recon, Relationship
+
+
+def _find_edge(recon: Recon, a: str, b: str) -> Relationship:
+    """The declared relationship joining tables `a` and `b` (either direction).
+    Raises if no FK connects them."""
+    for r in recon.relationships:
+        if (r.from_table == a and r.to_table == b) or (r.from_table == b and r.to_table == a):
+            return r
+    raise PickValidationError(f"no declared relationship between {a!r} and {b!r}")
+
+
+def build_joins(
+    recon: Recon,
+    base: str,
+    targets: list[str],
+    aliases: AliasMap,
+) -> list[tuple[exp.Expression, exp.Expression]]:
+    """Union of join paths from `base` to each target → list of
+    `(aliased_table, on_predicate)` specs in walk order."""
+    visited: set[str] = {base}
+    specs: list[tuple[exp.Expression, exp.Expression]] = []
+    for target in targets:
+        path = recon.join_path(base, target)
+        if path is None:
+            raise PickValidationError(f"no join path from {base!r} to {target!r} in recon")
+        for i in range(1, len(path)):
+            prev, cur = path[i - 1], path[i]
+            if cur in visited:
+                continue
+            edge = _find_edge(recon, prev, cur)
+            # Edge direction tells us which side owns which column.
+            if edge.from_table == prev:
+                left_t, left_c, right_t, right_c = (
+                    edge.from_table, edge.from_column, edge.to_table, edge.to_column,
+                )
+            else:
+                left_t, left_c, right_t, right_c = (
+                    edge.to_table, edge.to_column, edge.from_table, edge.from_column,
+                )
+            on_expr = exp.EQ(
+                this=aliases.col(left_t, left_c),
+                expression=aliases.col(right_t, right_c),
+            )
+            specs.append((aliases.aliased_table(cur), on_expr))
+            visited.add(cur)
+    return specs
diff --git a/api/composer/metrics.py b/api/composer/metrics.py
new file mode 100644
index 0000000..a56ed6d
--- /dev/null
+++ b/api/composer/metrics.py
@@ -0,0 +1,30 @@
+"""Metric-fragment qualification.
+
+Metric expressions in `metrics.yaml` are written *unaliased* — e.g.
+`AVG(CASE WHEN status = 'B' THEN 1.0 WHEN status = 'A' THEN 0.0 END)`, `A13`,
+`SUM(amount)`. Before such a fragment can sit in a multi-table query it has to
+be bound to the metric's own table alias, so the composer can introduce other
+tables via joins without ambiguity.
+
+We parse the fragment with sqlglot and inject the alias as the table qualifier
+on every bare column reference, returning the resulting expression node (not a
+string) so it splices directly into the query tree the composer assembles.
+"""
+
+from __future__ import annotations
+
+import sqlglot
+from sqlglot import exp
+
+
+def qualify_metric_expr(sql_fragment: str, table_alias: str) -> exp.Expression:
+    """Parse `sql_fragment` and qualify its bare column refs to `table_alias`.
+
+    Used for both the metric's SELECT expression and its optional WHERE filter
+    (both are written as bare expressions over the metric's table)."""
+    tree = sqlglot.parse_one(sql_fragment, dialect="postgres")
+    for col in tree.find_all(exp.Column):
+        if col.table:
+            continue
+        col.set("table", exp.Identifier(this=table_alias, quoted=True))
+    return tree
diff --git a/api/composer/order.py b/api/composer/order.py
new file mode 100644
index 0000000..2e81ff9
--- /dev/null
+++ b/api/composer/order.py
@@ -0,0 +1,38 @@
+"""ORDER BY rendering.
+
+A Pick orders by either the metric or a named dimension; both reference the
+SELECT-list *output alias* (the metric's name, or the dimension's output alias),
+not the underlying qualified column. Returns an `exp.Ordered` node.
+"""
+
+from __future__ import annotations
+
+from sqlglot import exp
+
+from api.composer.types import OrderBy, PickValidationError
+
+
+def _ordered(alias: str, direction: str) -> exp.Ordered:
+    return exp.Ordered(this=exp.column(alias, quoted=True), desc=(direction == "desc"))
+
+
+def render_order_by(
+    ob: OrderBy,
+    metric_name: str,
+    dim_select: dict[str, str],
+) -> exp.Ordered:
+    """`metric_name` is the metric SELECT alias; `dim_select` maps a dimension
+    ref → its output alias."""
+    if ob.by == "metric":
+        return _ordered(metric_name, ob.direction)
+    if ob.dimension not in dim_select:
+        raise PickValidationError(
+            f"order_by.dimension={ob.dimension!r} is not in group_by ({list(dim_select)})"
+        )
+    return _ordered(dim_select[ob.dimension], ob.direction)
+
+
+def default_order_by(metric_name: str) -> exp.Ordered:
+    """Sensible default when a grouped query gives no explicit order: largest
+    metric value first."""
+    return _ordered(metric_name, "desc")
diff --git a/api/datasets/financial/schema_docs.yaml b/api/datasets/financial/schema_docs.yaml
index 84f2c87..30eadef 100644
--- a/api/datasets/financial/schema_docs.yaml
+++ b/api/datasets/financial/schema_docs.yaml
@@ -9,8 +9,19 @@
 #   tables.<name>:           one-line description.
 #   columns.<table>.<col>:   one-line description. Sparse — only the ones
 #                            worth describing. Undescribed columns just have
-#                            no description in the recon.
+#                            no description in the recon. A column may also
+#                            carry a domain type, e.g.
+#                              loan.date: {desc: "...", type: date_yymmdd}
 #   relationships:           declared FKs (BIRD ships SQLite without them).
+#   encodings:               OPTIONAL. Storage-encoding overrides keyed by a
+#                            column's `type`. Each maps a role → SQL template
+#                            with a `{col}` placeholder. The composer ships
+#                            defaults (e.g. date_yymmdd); add an entry here only
+#                            to introduce a novel encoding or override a default:
+#                              encodings:
+#                                date_yymmdd:
+#                                  as_date: "TO_DATE(LPAD({col}::text, 6, '0'), 'YYMMDD')"
+#                            `financial` stores real DATE columns, so it needs none.
 
 schema: financial
 
diff --git a/api/main.py b/api/main.py
index d9328a6..a1b8ce6 100644
--- a/api/main.py
+++ b/api/main.py
@@ -5,6 +5,7 @@ Endpoints:
     POST /ask                     — submit a question, returns run_id
     GET  /runs/{run_id}/stream    — SSE stream of run events
     GET  /runs/{run_id}           — final state snapshot
+    GET  /runs/{run_id}/log       — JSONL transcript (every published event)
 """
 
 from __future__ import annotations
@@ -17,7 +18,7 @@ from datetime import datetime, timezone
 
 from fastapi import FastAPI, HTTPException
 from fastapi.middleware.cors import CORSMiddleware
-from fastapi.responses import StreamingResponse
+from fastapi.responses import FileResponse, StreamingResponse
 from pydantic import BaseModel
 
 from api.runtime import events
@@ -116,3 +117,13 @@ async def stream_run(run_id: str):
     if run_id not in runs:
         raise HTTPException(404, detail=f"Run {run_id} not found")
     return StreamingResponse(events.stream(run_id), media_type="text/event-stream")
+
+
+@app.get("/runs/{run_id}/log")
+async def get_run_log(run_id: str):
+    """Serve the JSONL transcript written by `events.publish`. Available even
+    for in-flight runs (the file is flushed after every write)."""
+    path = events.log_path(run_id)
+    if not path.exists():
+        raise HTTPException(404, detail=f"No log file for run {run_id}")
+    return FileResponse(path, media_type="application/x-ndjson", filename=path.name)
diff --git a/api/recon/build.py b/api/recon/build.py
index 5aa287b..68ed35c 100644
--- a/api/recon/build.py
+++ b/api/recon/build.py
@@ -137,6 +137,11 @@ def merge_into_recon(dataset: str, extracted: dict[str, Any]) -> Recon:
 
     relationships = [Relationship.parse(r) for r in (aug.get("relationships", []) or [])]
 
+    # Storage-encoding overrides (optional). Keyed by Column.semantic_type; each
+    # value maps a role (e.g. "as_date") to a SQL template with a `{col}`
+    # placeholder. The composer merges these over its built-in defaults.
+    encodings: dict[str, dict[str, str]] = aug.get("encodings", {}) or {}
+
     column_to_tables: dict[str, list[str]] = {}
     for t in tables.values():
         for c in t.columns:
@@ -150,6 +155,7 @@ def merge_into_recon(dataset: str, extracted: dict[str, Any]) -> Recon:
         metrics=metrics,
         column_to_tables=column_to_tables,
         relationships=relationships,
+        encodings=encodings,
     )
 
 
diff --git a/api/recon/types.py b/api/recon/types.py
index e61556c..b84b082 100644
--- a/api/recon/types.py
+++ b/api/recon/types.py
@@ -151,6 +151,12 @@ class Recon:
     metrics: dict[str, Metric]
     column_to_tables: dict[str, list[str]] = field(default_factory=dict)
     relationships: list[Relationship] = field(default_factory=list)
+    # Storage-encoding overrides keyed by Column.semantic_type. Each value is a
+    # map of role → SQL template with a `{col}` placeholder (e.g.
+    # {"date_yymmdd": {"as_date": "TO_DATE(LPAD({col}::text, 6, '0'), 'YYMMDD')"}}).
+    # The composer merges these over its DEFAULT_ENCODINGS — the dataset wins.
+    # Empty for datasets whose columns are stored in native types.
+    encodings: dict[str, dict[str, str]] = field(default_factory=dict)
 
     # ── Read-side helpers used by Analyses + the composer ──
 
@@ -299,6 +305,7 @@ class Recon:
             "metrics": {n: m.to_dict() for n, m in self.metrics.items()},
             "column_to_tables": self.column_to_tables,
             "relationships": [r.to_dict() for r in self.relationships],
+            "encodings": self.encodings,
         }
 
     @classmethod
@@ -309,4 +316,5 @@ class Recon:
             metrics={n: Metric.from_dict(m) for n, m in d.get("metrics", {}).items()},
             column_to_tables=d.get("column_to_tables", {}),
             relationships=[Relationship(**r) for r in d.get("relationships", [])],
+            encodings=d.get("encodings", {}) or {},
         )
diff --git a/api/runtime/events.py b/api/runtime/events.py
index c679e50..ab5e8dd 100644
--- a/api/runtime/events.py
+++ b/api/runtime/events.py
@@ -15,24 +15,59 @@ from __future__ import annotations
 
 import asyncio
 import json
-from typing import Any
+import logging
+import time
+from pathlib import Path
+from typing import Any, IO
 
 from api.analyses.types import Finding
 from api.runtime.context import current_analysis, current_run_id
 
+logger = logging.getLogger("nvi.runtime.events")
+
 # run_id -> queue of events. Events are plain dicts.
 _queues: dict[str, asyncio.Queue[dict[str, Any] | None]] = {}
 
+# run_id -> append-mode JSONL log file. Mirrors every published event so the
+# whole run is replayable from disk. Path: `<LOG_DIR>/<run_id>.jsonl`.
+_log_files: dict[str, IO[str]] = {}
+LOG_DIR = Path(".data/logs")
+
+
+def log_path(run_id: str) -> Path:
+    """Where this run's JSONL log lives. Resolved relative to cwd of the api
+    process — `/app/.data/logs/` in the pod."""
+    return LOG_DIR / f"{run_id}.jsonl"
+
 
 # ── Queue management ──
 
 def open_run(run_id: str) -> asyncio.Queue:
     q: asyncio.Queue = asyncio.Queue()
     _queues[run_id] = q
+    try:
+        LOG_DIR.mkdir(parents=True, exist_ok=True)
+        _log_files[run_id] = log_path(run_id).open("a", encoding="utf-8")
+    except OSError as e:
+        # Filesystem issues shouldn't kill the run; just lose the on-disk log.
+        logger.warning("could not open log file for run %s: %s", run_id, e)
     return q
 
 
+def _write_log(run_id: str, event: dict[str, Any]) -> None:
+    f = _log_files.get(run_id)
+    if f is None:
+        return
+    try:
+        record = {"t": time.time(), **event}
+        f.write(json.dumps(record, default=str) + "\n")
+        f.flush()  # so `kubectl exec ... -- tail -f` shows live progress
+    except (OSError, ValueError) as e:
+        logger.warning("log write failed for run %s: %s", run_id, e)
+
+
 async def publish(run_id: str, event: dict[str, Any]) -> None:
+    _write_log(run_id, event)
     q = _queues.get(run_id)
     if q is not None:
         await q.put(event)
@@ -54,6 +89,12 @@ async def close(run_id: str) -> None:
 
 def drop(run_id: str) -> None:
     _queues.pop(run_id, None)
+    f = _log_files.pop(run_id, None)
+    if f is not None:
+        try:
+            f.close()
+        except OSError:
+            pass
 
 
 # ── SSE formatting + streaming ──
diff --git a/tests/test_composer_dates.py b/tests/test_composer_dates.py
index 310a498..639d2e1 100644
--- a/tests/test_composer_dates.py
+++ b/tests/test_composer_dates.py
@@ -1,10 +1,25 @@
-"""Date-range resolver tests — bounds parsing + dispatch on column type."""
+"""Date-range resolver tests — bounds parsing + dispatch on column type/encoding.
+
+`resolve_date_range` now returns an `exp` node and takes the effective encoding
+map (built-in defaults overlaid with any dataset overrides). Tests render the
+node to SQL to assert on the predicate.
+"""
 
 import pytest
+from sqlglot import exp
 
 from api.composer.dates import _bounds, resolve_date_range
+from api.composer.encodings import DEFAULT_ENCODINGS, effective_encodings
 from api.composer.types import PickValidationError
-from api.recon.types import Column
+from api.recon.types import Column, Recon
+
+
+def _col(name: str = "date", table: str = "l") -> exp.Column:
+    return exp.column(name, table=table, quoted=True)
+
+
+def _sql(node: exp.Expression) -> str:
+    return node.sql(dialect="postgres", identify=True)
 
 
 def test_bounds_year():
@@ -26,25 +41,56 @@ def test_bounds_unsupported_raises():
 
 def test_resolve_real_date_column():
     col = Column(name="date", sql_type="DATE", nullable=False)
-    out = resolve_date_range("1996", col, '"l"."date"')
+    out = _sql(resolve_date_range("1996", col, _col(), {}))
     assert out == "\"l\".\"date\" BETWEEN '1996-01-01' AND '1996-12-31'"
 
 
 def test_resolve_timestamp_column():
     col = Column(name="created_at", sql_type="TIMESTAMP WITHOUT TIME ZONE", nullable=True)
-    out = resolve_date_range("1996-Q2", col, '"x"."created_at"')
+    out = _sql(resolve_date_range("1996-Q2", col, _col("created_at", "x"), {}))
     assert out == "\"x\".\"created_at\" BETWEEN '1996-04-01' AND '1996-06-30'"
 
 
-def test_resolve_yymmdd_semantic_type():
+def test_resolve_yymmdd_via_default_encoding():
     col = Column(name="date", sql_type="INTEGER", nullable=False,
                  semantic_type="date_yymmdd")
-    out = resolve_date_range("1996", col, '"l"."date"')
+    out = _sql(resolve_date_range("1996", col, _col(), DEFAULT_ENCODINGS))
     assert "TO_DATE(LPAD(" in out
     assert "BETWEEN '1996-01-01' AND '1996-12-31'" in out
 
 
+def test_dataset_encoding_overrides_default():
+    # A dataset can override a built-in encoding with its own coercion SQL.
+    col = Column(name="d", sql_type="INTEGER", nullable=False,
+                 semantic_type="date_yymmdd")
+    override = {"date_yymmdd": {"as_date": "MAKE_DATE(1900 + {col} / 10000, 1, 1)"}}
+    out = _sql(resolve_date_range("1996", col, _col("d"), override))
+    assert "MAKE_DATE(" in out
+    assert "TO_DATE" not in out
+
+
+def test_resolve_semantic_type_without_encoding_raises():
+    # Declared semantic_type but no matching encoding and not a native date type.
+    col = Column(name="d", sql_type="INTEGER", nullable=False,
+                 semantic_type="date_yymmdd")
+    with pytest.raises(PickValidationError, match="not supported"):
+        resolve_date_range("1996", col, _col("d"), {})
+
+
 def test_resolve_unsupported_column_type_raises():
     col = Column(name="x", sql_type="TEXT", nullable=True)
     with pytest.raises(PickValidationError, match="not supported"):
-        resolve_date_range("1996", col, '"x"."x"')
+        resolve_date_range("1996", col, _col("x", "x"), {})
+
+
+def test_effective_encodings_merges_defaults_with_dataset():
+    recon = Recon(schema="t", tables={}, metrics={},
+                  encodings={"my_epoch": {"as_date": "TO_TIMESTAMP({col})::date"}})
+    eff = effective_encodings(recon)
+    assert "date_yymmdd" in eff          # default survives
+    assert eff["my_epoch"]["as_date"].startswith("TO_TIMESTAMP")  # dataset added
+
+    # Dataset entry wins on key collision.
+    recon2 = Recon(schema="t", tables={}, metrics={},
+                   encodings={"date_yymmdd": {"as_date": "CUSTOM({col})"}})
+    assert effective_encodings(recon2)["date_yymmdd"]["as_date"] == "CUSTOM({col})"
diff --git a/ui/app/src/pages/Ask.vue b/ui/app/src/pages/Ask.vue
index 10123ae..dad5ca2 100644
--- a/ui/app/src/pages/Ask.vue
+++ b/ui/app/src/pages/Ask.vue
@@ -141,7 +141,6 @@ function nodeClasses(n: GraphNode) {
     <div class="trace-pane">
       <Panel title="Plan" :status="statusOf(status)">
         <div v-if="rationale" class="rationale">{{ rationale }}</div>
-        <div v-else class="muted small">Waiting for the planner…</div>
 
         <div class="graph">
           <div v-for="(n, idx) in graphNodes" :key="n.id" class="graph-row">