diff --git a/Cargo.lock b/Cargo.lock
new file mode 100644
index 00000000..8b1e8423
--- /dev/null
+++ b/Cargo.lock
@@ -0,0 +1,7 @@
+# This file is automatically @generated by Cargo.
+# It is not intended for manual editing.
+version = 4
+
+[[package]]
+name = "asap-control-core"
+version = "0.1.0"
diff --git a/Cargo.toml b/Cargo.toml
new file mode 100644
index 00000000..c3365c5b
--- /dev/null
+++ b/Cargo.toml
@@ -0,0 +1,5 @@
+[workspace]
+members = [
+    "crates/core",
+]
+resolver = "2"
diff --git a/crates/core/Cargo.toml b/crates/core/Cargo.toml
new file mode 100644
index 00000000..99972465
--- /dev/null
+++ b/crates/core/Cargo.toml
@@ -0,0 +1,4 @@
+[package]
+name = "asap-control-core"
+version = "0.1.0"
+edition = "2021"
diff --git a/crates/core/src/intent_algebra/expr.rs b/crates/core/src/intent_algebra/expr.rs
new file mode 100644
index 00000000..6860d6fc
--- /dev/null
+++ b/crates/core/src/intent_algebra/expr.rs
@@ -0,0 +1,305 @@
+use std::rc::Rc;
+use std::time::Duration;
+
+use crate::types::AccuracyTarget;
+use super::schema::{HasSchema, L3Schema, SchemaCatalog};
+
+// ── Stub leaf / supporting types ──────────────────────────────────────────────
+// Full definitions will be added as the respective layers are implemented.
+
+/// A row-level filter predicate (WHERE clause / PromQL label matcher).
+#[derive(Debug, Clone)] pub struct Predicate;
+/// One item in a SELECT projection list.
+#[derive(Debug, Clone)] pub struct ProjectItem;
+/// A GROUP BY key reference.
+#[derive(Debug, Clone)] pub struct GroupKey;
+/// A reference to a column by name.
+#[derive(Debug, Clone)] pub struct ColumnRef;
+/// A set of partitioning keys (sharding hint for L5 stage allocator).
+#[derive(Debug, Clone)] pub struct PartitionKeys;
+/// One key in an ORDER BY clause.
+#[derive(Debug, Clone)] pub struct SortKey;
+/// An analytic window frame (ROWS / RANGE BETWEEN …).
+#[derive(Debug, Clone)] pub struct WindowFrame;
+/// PromQL vector-match modifiers (`on`/`ignoring` + `group_left`/`group_right`).
+#[derive(Debug, Clone)] pub struct VectorMatch;
+/// Reference to a metric by name (PromQL / OTLP).
+#[derive(Debug, Clone)] pub struct MetricRef;
+/// Closed time interval for a time-series scan.
+#[derive(Debug, Clone)] pub struct TimeRange;
+/// Label matchers applied to a time-series scan.
+#[derive(Debug, Clone)] pub struct LabelFilter;
+/// Reference to a relational table by name.
+#[derive(Debug, Clone)] pub struct TableRef;
+/// Join key specification (USING / ON column reference).
+#[derive(Debug, Clone)] pub struct JoinKey;
+
+// ── Enum supporting types ─────────────────────────────────────────────────────
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum JoinKind {
+    Inner,
+    Left,
+    Right,
+    Full,
+    Cross,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum SetOpKind {
+    Union,
+    Intersect,
+    Except,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum BinaryOpKind {
+    // Arithmetic
+    Add, Sub, Mul, Div, Mod,
+    // Comparison
+    Eq, NotEq, Lt, LtEq, Gt, GtEq,
+    // Boolean / PromQL set operators
+    And, Or, Unless,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum WindowFuncKind {
+    RowNumber, Rank, DenseRank,
+    Lag, Lead,
+    FirstValue, LastValue,
+    NthValue(u64),
+    Sum, Avg, Count, Min, Max,
+}
+
+/// Which data model a `Source` or `AggIntent` operates over.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum DataModel {
+    TimeSeries,
+    Tabular,
+    /// Agnostic — works over either data model.
+    Any,
+}
+
+// ── Time window kind ──────────────────────────────────────────────────────────
+
+/// The lifecycle / flush semantics of a streaming time window.
+/// Used by `QueryExpr::TimeWindow`; distinct from SQL analytic frames
+/// (`QueryExpr::WindowFunc`).
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum TimeWindowKind {
+    /// Non-overlapping fixed-size windows.
+    Tumbling,
+    /// Overlapping windows advancing by `slide` interval.
+    Sliding,
+    /// Windows that open on activity and close after a gap of inactivity.
+    Session,
+}
+
+// ── Leaf data source ──────────────────────────────────────────────────────────
+
+/// The leaf data source of a query. Carried by `QueryExpr::Scan` to keep
+/// L3 data-model-agnostic: everything above `Scan` (`Filter`, `Aggregate`,
+/// etc.) works identically regardless of `Source` variant.
+#[derive(Debug, Clone)]
+pub enum Source {
+    /// Time-series input — deployment-model-asapquery / asaplifecycle shape.
+    TimeSeries {
+        metric: MetricRef,
+        time: TimeRange,
+        labels: LabelFilter,
+    },
+    /// Tabular input — deployment-model-asapfusion / future-OLAP shape.
+    Table {
+        table_ref: TableRef,
+        columns: Vec<ColumnRef>,
+    },
+    /// Recursive join over sources (multi-table tabular queries).
+    Join {
+        left: Box<Source>,
+        right: Box<Source>,
+        on: JoinKey,
+    },
+}
+
+impl Source {
+    pub fn data_model(&self) -> DataModel {
+        match self {
+            Source::TimeSeries { .. } => DataModel::TimeSeries,
+            Source::Table { .. } | Source::Join { .. } => DataModel::Tabular,
+        }
+    }
+}
+
+// ── Aggregation intent ────────────────────────────────────────────────────────
+
+/// What to compute, not how. Sketch type and parameters are chosen by L4
+/// rules; `AggIntent` is the L3 statement of intent only.
+///
+/// Heavy-hitter top-k (`TopK`) is a first-class intent because dedicated
+/// sketch primitives (SpaceSaving, CMS-with-heap) compute it in one pass.
+/// Generic ordering+limit stays as `QueryExpr::Sort + QueryExpr::Limit`.
+#[derive(Debug, Clone)]
+pub enum AggIntent {
+    // ── Data-model-agnostic ───────────────────────────────────────────────────
+    Count { accuracy: AccuracyTarget },
+    Sum,
+    Min,
+    Max,
+    Quantile { q: f64, accuracy: AccuracyTarget },
+    /// Heavy-hitter top-k. Distinct from generic `Sort + Limit` — a
+    /// dedicated sketch (SpaceSaving, CMS-with-heap) computes it as a single
+    /// primitive. Recognised by L1→L2→L3 lowering on `ORDER BY count DESC
+    /// LIMIT k` / PromQL `topk(k, …)`.
+    TopK { k: usize, by: Vec<ColumnRef>, accuracy: AccuracyTarget },
+    Cardinality { accuracy: AccuracyTarget },
+
+    // ── Time-series streaming derivatives ────────────────────────────────────
+    // Include PromQL counter-reset adjustment; not equivalent to Sum/Count
+    // over a Window. Kept distinct so delta-set aggregators bind directly.
+    Rate { window: Duration },
+    Increase { window: Duration },
+}
+
+impl AggIntent {
+    /// Which data model this intent semantically requires. L4 rules consult
+    /// this to skip non-applicable intents (e.g. `Rate` over a `Source::Table`).
+    pub fn requires(&self) -> DataModel {
+        todo!()
+    }
+
+    /// Output column type — used by L3 schema derivation for `Aggregate`.
+    pub fn output_type(&self, _input: &super::schema::L3Field) -> super::schema::L3DataType {
+        todo!()
+    }
+}
+
+// ── L3 DAG node ───────────────────────────────────────────────────────────────
+
+/// A node in the L3 DAG. Wraps the expression and its derived output schema
+/// so that every edge implicitly carries a typed schema: holding an
+/// `Rc<L3Node>` gives you both the child expression and the schema of the
+/// data flowing on that edge.
+#[derive(Debug, Clone)]
+pub struct L3Node {
+    pub expr: QueryExpr,
+    /// Output schema of `expr` — the schema of the data flowing on the edge
+    /// leading *from* this node to its parent(s).
+    pub schema: L3Schema,
+}
+
+// ── L3 intent algebra IR ──────────────────────────────────────────────────────
+
+/// Language- and deployment-independent intent-only IR. No sketch types,
+/// no sketch parameters, no language-specific operators. Traversing from
+/// the root node yields a DAG; shared sub-expressions appear as multiple
+/// `Rc` references to the same `L3Node`.
+#[derive(Debug, Clone)]
+pub enum QueryExpr {
+    // ── Base relations ────────────────────────────────────────────────────────
+    /// Outermost leaf. `source` carries the data-model-specific leaf shape.
+    Scan { source: Source, predicates: Vec<Predicate> },
+    /// Reference to a named `LetBinding` sub-expression; resolved at plan time.
+    Ref(String),
+
+    // ── Filtering & projection ────────────────────────────────────────────────
+    /// σ — row-level filter. Output schema = child schema (unchanged).
+    Filter { child: Rc<L3Node>, pred: Predicate },
+    /// π — column projection. Output schema = child schema projected to `cols`.
+    Project { child: Rc<L3Node>, cols: Vec<ProjectItem> },
+
+    // ── Aggregation ───────────────────────────────────────────────────────────
+    /// γ + α — GROUP BY + aggregate intents. Concrete operator (HashAgg /
+    /// SortAgg / SketchAgg) chosen by L4; `aggs` carry intent only.
+    Aggregate {
+        child: Rc<L3Node>,
+        by: Vec<GroupKey>,
+        aggs: Vec<AggIntent>,
+        having: Option<Predicate>,
+    },
+
+    // ── Time / streaming windows ──────────────────────────────────────────────
+    /// ψ — tumbling / sliding / session window over the time axis. Defines
+    /// the flush / reset lifecycle for aggregates in its sub-DAG. SQL analytic
+    /// frames are a different node (`WindowFunc`).
+    TimeWindow {
+        child: Rc<L3Node>,
+        kind: TimeWindowKind,
+        size: Duration,
+        slide: Option<Duration>,
+    },
+
+    // ── Distributed-execution structure ───────────────────────────────────────
+    /// Logical-only partitioning marker. Output schema = child schema.
+    /// Carries a sharding hint for the L5 stage allocator.
+    Partition { child: Rc<L3Node>, keys: PartitionKeys },
+    /// δ — SQL `DISTINCT` / row deduplication.
+    Distinct { child: Rc<L3Node>, cols: Vec<ColumnRef> },
+    /// ⊕ — exact union of sub-results from independent stages or shards.
+    /// Sketch unions are a separate node in `SummaryExpr` because they carry
+    /// sketch-family / params type constraints.
+    Merge { children: Vec<Rc<L3Node>> },
+
+    // ── Joins ─────────────────────────────────────────────────────────────────
+    /// Logical join. L4 picks the physical alternative (HashJoin /
+    /// SortMergeJoin / SketchJoin) based on selectivity, memory budget, and
+    /// accuracy target.
+    Join {
+        kind: JoinKind,
+        left: Rc<L3Node>,
+        right: Rc<L3Node>,
+        pred: Option<Predicate>,
+    },
+
+    // ── Set operators ─────────────────────────────────────────────────────────
+    SetOp {
+        kind: SetOpKind,
+        all: bool,
+        left: Rc<L3Node>,
+        right: Rc<L3Node>,
+    },
+
+    // ── Ordering & limiting ───────────────────────────────────────────────────
+    /// Generic order-by for non-heavy-hitter cases (`ORDER BY name LIMIT 10`).
+    /// Heavy-hitter shapes lower to `AggIntent::TopK` instead.
+    Sort { child: Rc<L3Node>, keys: Vec<SortKey> },
+    Limit { child: Rc<L3Node>, n: u64, offset: u64 },
+
+    // ── Subquery / CTE ────────────────────────────────────────────────────────
+    Subquery { child: Rc<L3Node>, alias: String },
+    /// SQL `WITH name AS (expr) … body`; lowering target for PromQL
+    /// recording-rule bindings. The `expr` sub-DAG may be referenced N times
+    /// via `Ref(name)` in `body`, giving the DAG its fan-in.
+    LetBinding {
+        name: String,
+        expr: Rc<L3Node>,
+        body: Rc<L3Node>,
+    },
+
+    // ── Analytic window functions ─────────────────────────────────────────────
+    /// SQL `OVER (PARTITION BY … ORDER BY … ROWS BETWEEN …)`.
+    /// Distinct from `TimeWindow` — that is a streaming window over the time
+    /// axis; this is an analytic frame over already-grouped rows.
+    WindowFunc {
+        child: Rc<L3Node>,
+        func: WindowFuncKind,
+        partition_by: Vec<GroupKey>,
+        order_by: Vec<SortKey>,
+        frame: Option<WindowFrame>,
+    },
+
+    // ── Binary composition ────────────────────────────────────────────────────
+    /// Arithmetic / comparison / boolean composition (PromQL binary ops
+    /// including `and` / `or` / `unless`, SQL boolean composition).
+    BinaryOp {
+        op: BinaryOpKind,
+        lhs: Rc<L3Node>,
+        rhs: Rc<L3Node>,
+        vector_match: Option<VectorMatch>,
+    },
+}
+
+impl HasSchema for QueryExpr {
+    fn output_schema(&self, _input_schemas: &[&L3Schema], _catalog: &SchemaCatalog) -> L3Schema {
+        todo!()
+    }
+}
diff --git a/crates/core/src/intent_algebra/mod.rs b/crates/core/src/intent_algebra/mod.rs
new file mode 100644
index 00000000..a5117cb3
--- /dev/null
+++ b/crates/core/src/intent_algebra/mod.rs
@@ -0,0 +1,10 @@
+pub mod expr;
+pub mod schema;
+
+pub use expr::{
+    AggIntent, BinaryOpKind, ColumnRef, DataModel, GroupKey, JoinKey, JoinKind, L3Node,
+    LabelFilter, MetricRef, PartitionKeys, Predicate, ProjectItem, QueryExpr, SetOpKind,
+    SortKey, Source, TableRef, TimeRange, TimeWindowKind, VectorMatch, WindowFrame,
+    WindowFuncKind,
+};
+pub use schema::{HasSchema, L3DataType, L3Field, L3Schema, SchemaCatalog};
diff --git a/crates/core/src/intent_algebra/schema.rs b/crates/core/src/intent_algebra/schema.rs
new file mode 100644
index 00000000..4f6d1fc0
--- /dev/null
+++ b/crates/core/src/intent_algebra/schema.rs
@@ -0,0 +1,53 @@
+/// Opaque handle to the external data-source catalog (Prometheus metric
+/// metadata, SQL `information_schema`, DataFusion catalog). Used only by
+/// `Scan` schema derivation to resolve leaf column types; all other nodes
+/// derive their output schemas purely from their input schemas.
+pub struct SchemaCatalog;
+
+// ── Data types ────────────────────────────────────────────────────────────────
+
+/// Column types that may appear on an L3 DAG edge.
+/// L4 extends this set with `L4DataType::Sketch`; L3 edges never carry
+/// sketch-state columns.
+#[derive(Debug, Clone, PartialEq)]
+pub enum L3DataType {
+    Int64,
+    Float64,
+    Utf8,
+    Boolean,
+    Timestamp,
+    Duration,
+    /// Key→Value map (e.g. PromQL label set encoded as a column).
+    Map(Box<L3DataType>, Box<L3DataType>),
+    List(Box<L3DataType>),
+}
+
+// ── Schema ────────────────────────────────────────────────────────────────────
+
+#[derive(Debug, Clone)]
+pub struct L3Field {
+    pub name: String,
+    pub dtype: L3DataType,
+    pub nullable: bool,
+}
+
+/// Schema carried on every edge of the L3 DAG. Describes the columns
+/// flowing between two operators. Type-checked at plan construction time:
+/// a node whose predicate references a column absent from its child's
+/// `L3Schema` is a plan-time error.
+#[derive(Debug, Clone)]
+pub struct L3Schema {
+    pub fields: Vec<L3Field>,
+    /// Index into `fields` for the time axis, if any.
+    /// PromQL `Scan` leaves always carry one; SQL leaves may or may not.
+    pub time_index: Option<usize>,
+}
+
+// ── Schema derivation trait ───────────────────────────────────────────────────
+
+/// Implemented by `QueryExpr` to compute the output schema of a node given
+/// its children's output schemas. The `L3Node` wrapper stores the derived
+/// schema so derivation runs once at construction, not on every traversal.
+pub trait HasSchema {
+    fn output_schema(&self, input_schemas: &[&L3Schema], catalog: &SchemaCatalog) -> L3Schema;
+}
diff --git a/crates/core/src/lib.rs b/crates/core/src/lib.rs
new file mode 100644
index 00000000..5d62afa9
--- /dev/null
+++ b/crates/core/src/lib.rs
@@ -0,0 +1,4 @@
+pub mod intent_algebra;
+pub mod sketch_algebra;
+pub mod types;
+pub mod workload;
diff --git a/crates/core/src/sketch_algebra/expr.rs b/crates/core/src/sketch_algebra/expr.rs
new file mode 100644
index 00000000..09772a8e
--- /dev/null
+++ b/crates/core/src/sketch_algebra/expr.rs
@@ -0,0 +1,96 @@
+use std::rc::Rc;
+
+use crate::intent_algebra::{ColumnRef, GroupKey, L3Node};
+use super::schema::L4Schema;
+use super::sketch::{SummaryKind, SummaryParams, SketchQuery};
+
+// ── L4 DAG node ───────────────────────────────────────────────────────────────
+
+/// A node in the L4 DAG. Mirrors `L3Node`: wraps the expression and its
+/// derived output schema so every edge carries a typed schema.
+/// `L4Schema` may contain `L4DataType::Sketch` columns; `L3Schema` cannot.
+#[derive(Debug, Clone)]
+pub struct L4Node {
+    pub expr: SummaryExpr,
+    /// Output schema of `expr` — the schema of the data flowing on the edge
+    /// leading *from* this node to its parent(s).
+    pub schema: L4Schema,
+}
+
+// ── L4 sketch-bound IR ────────────────────────────────────────────────────────
+
+/// Sketch-bound IR produced by L4 optimizer rules. L4 rules selectively
+/// replace logical aggregates and joins in the L3 `QueryExpr` with their
+/// sketch-bound counterparts; everything not rewritten passes through as
+/// `Logical(Rc<L3Node>)`.
+///
+/// Traversing from the root node yields a DAG; shared sub-expressions appear
+/// as multiple `Rc` references to the same `L4Node` or `L3Node`.
+#[derive(Debug, Clone)]
+pub enum SummaryExpr {
+    /// Any L3 node that no L4 rule rewrote (e.g. `Filter`, `Project`, `Sort`).
+    /// Output schema is the inner L3 node's schema, lifted to `L4Schema`
+    /// with all fields as `L4DataType::Primitive`.
+    Logical(Rc<L3Node>),
+
+    /// Sketch aggregation. L4 chose `sketch` + `params` from the catalog
+    /// for `AggIntent` under `DeploymentConstraints`.
+    /// Output schema: `by` columns (verbatim) + one `Sketch(sketch, params)`
+    /// field carrying partial sketch state per group.
+    SummaryAgg {
+        child: Rc<L4Node>,
+        sketch: SummaryKind,
+        params: SummaryParams,
+        /// The column being summarised (fed into the sketch).
+        col: ColumnRef,
+        /// GROUP BY keys carried through to the output schema.
+        by: Vec<GroupKey>,
+    },
+
+    /// Sketch-aware join (KMV / theta for join-cardinality; join-sample for
+    /// sampling). Emitted only when a `Bind*OnJoin` rule fires.
+    /// Output schema: one `Sketch(sketch, params)` field read by a downstream
+    /// `SummaryEstimate`.
+    SummaryJoin {
+        outer: Rc<L4Node>,
+        inner: Rc<L4Node>,
+        key: ColumnRef,
+        sketch: SummaryKind,
+        params: SummaryParams,
+    },
+
+    /// Subtract one sketch from another. Valid only for sketches with a
+    /// linear-inverse property (CMS, theta, count-based). Catalog flag
+    /// `subtractable` must be true for the sketch family.
+    /// Output schema: one `Sketch(s, p)` field (same family + params as inputs).
+    SummarySubtract {
+        left: Rc<L4Node>,
+        right: Rc<L4Node>,
+    },
+
+    /// Delete a key from a sketch (CMS update with −1, deletable Bloom
+    /// filter). Catalog flag `deletable` must be true. Output schema =
+    /// input schema unchanged in type (same `Sketch(s, p)` field).
+    SummaryDelete {
+        sketch_input: Rc<L4Node>,
+        key: ColumnRef,
+    },
+
+    /// Read out a query result from a built sketch. The `Sketch(…)` field
+    /// type does *not* propagate downstream of an estimate — the output
+    /// schema is a regular row-shaped schema (Float64 for quantile, Int64
+    /// for count/cardinality, `[(key, count)]` for top-k).
+    SummaryEstimate {
+        sketch_input: Rc<L4Node>,
+        query: SketchQuery,
+    },
+
+    /// ⊕ — union of sketches across stages / shards. Distinct from L3
+    /// `Merge` because sketch union has type constraints: all inputs must
+    /// agree on `(sketch, params)` and the catalog flag `mergeable` must
+    /// be true. Inserted by the L5 stage allocator on cut edges.
+    /// Output schema: one `Sketch(s, p)` field (same family + params as inputs).
+    SummaryMerge {
+        children: Vec<Rc<L4Node>>,
+    },
+}
diff --git a/crates/core/src/sketch_algebra/mod.rs b/crates/core/src/sketch_algebra/mod.rs
new file mode 100644
index 00000000..a2f589e4
--- /dev/null
+++ b/crates/core/src/sketch_algebra/mod.rs
@@ -0,0 +1,7 @@
+pub mod expr;
+pub mod schema;
+pub mod sketch;
+
+pub use expr::{L4Node, SummaryExpr};
+pub use schema::{L4DataType, L4Field, L4Schema};
+pub use sketch::{SummaryKind, SummaryParams, SketchQuery};
diff --git a/crates/core/src/sketch_algebra/schema.rs b/crates/core/src/sketch_algebra/schema.rs
new file mode 100644
index 00000000..b6d5d14e
--- /dev/null
+++ b/crates/core/src/sketch_algebra/schema.rs
@@ -0,0 +1,42 @@
+use crate::intent_algebra::L3DataType;
+use super::sketch::{SummaryKind, SummaryParams};
+
+// ── L4 data types ─────────────────────────────────────────────────────────────
+
+/// Column types that may appear on an L4 DAG edge. A strict superset of
+/// `L3DataType`: L4 adds `Sketch` for edges that carry partial sketch state
+/// between a `SummaryAgg` and a downstream `SummaryEstimate` or `SummaryMerge`.
+///
+/// The `Sketch` variant carries `(kind, params)` so the type system can reject
+/// merges of incompatible sketches at plan construction time — a `SummaryMerge`
+/// over `Sketch(Kll, …)` and `Sketch(Cms, …)` inputs is a plan-time error.
+#[derive(Debug, Clone, PartialEq)]
+pub enum L4DataType {
+    /// Any base L3 column type — passed through unchanged from L3 edges.
+    Primitive(L3DataType),
+    /// Opaque summary state (exact accumulator or approximate sketch).
+    /// The `(kind, params)` pair is the type identity: two summary columns
+    /// are compatible only if both match exactly.
+    Sketch(SummaryKind, SummaryParams),
+}
+
+// ── L4 schema ─────────────────────────────────────────────────────────────────
+
+#[derive(Debug, Clone)]
+pub struct L4Field {
+    pub name: String,
+    pub dtype: L4DataType,
+    pub nullable: bool,
+}
+
+/// Schema carried on every edge of the L4 DAG. Extends `L3Schema` with the
+/// ability to express sketch-state columns. L3 and L4 schemas are separate
+/// types so L3 nodes structurally cannot carry `Sketch`-typed columns — any
+/// attempt to do so is a compile-time type error.
+#[derive(Debug, Clone)]
+pub struct L4Schema {
+    pub fields: Vec<L4Field>,
+    /// Index into `fields` for the time axis, if any (same semantics as
+    /// `L3Schema::time_index`).
+    pub time_index: Option<usize>,
+}
diff --git a/crates/core/src/sketch_algebra/sketch.rs b/crates/core/src/sketch_algebra/sketch.rs
new file mode 100644
index 00000000..36aea623
--- /dev/null
+++ b/crates/core/src/sketch_algebra/sketch.rs
@@ -0,0 +1,75 @@
+use crate::intent_algebra::ColumnRef;
+
+// ── Summary kind identifiers ───────────────────────────────────────────────────
+
+/// Identifies a precomputed aggregation family — either an exact accumulator
+/// or an approximate sketch. Used as a type tag in `L4DataType` and as the
+/// binding choice recorded in `SummaryExpr` nodes.
+#[derive(Debug, Clone, PartialEq, Eq, Hash)]
+pub enum SummaryKind {
+    // ── Exact accumulators (no approximation error) ──────────────────────────
+    /// Exact sum accumulator (mergeable by addition).
+    Sum,
+    /// Exact count accumulator (mergeable by addition).
+    Count,
+    /// Exact min/max accumulator (mergeable by comparison).
+    MinMax,
+    /// Exact increase accumulator (counter-reset-aware delta).
+    Increase,
+    /// Rate accumulator (increase / time window duration).
+    Rate,
+    // ── Approximate sketches ─────────────────────────────────────────────────
+    /// KLL quantile sketch (mergeable, ε-accurate rank queries).
+    Kll,
+    /// Count-Min Sketch (mergeable, (ε,δ)-accurate frequency queries).
+    Cms,
+    /// HyperLogLog (mergeable, (ε,δ)-accurate cardinality).
+    Hll,
+    /// DDSketch (mergeable, relative-error quantile queries).
+    DDSketch,
+    /// CMS augmented with a min-heap for top-k / heavy-hitter queries.
+    CmsWithHeap,
+    /// K-Minimum Values sketch (mergeable, join-cardinality estimation).
+    Kmv,
+    /// Theta sketch (mergeable, set operations + cardinality).
+    Theta,
+}
+
+// ── Sketch parameters ─────────────────────────────────────────────────────────
+
+/// Concrete, catalog-validated parameters for a specific summary instance.
+/// The variant must correspond to the associated `SummaryKind`; mismatches
+/// are caught at L4 bind time before L5 ever sees the plan.
+/// Exact-accumulator variants carry no parameters (their semantics are fixed).
+#[derive(Debug, Clone, PartialEq)]
+pub enum SummaryParams {
+    // Exact accumulators — no tuning parameters
+    Sum,
+    Count,
+    MinMax,
+    Increase,
+    Rate,
+    // Approximate sketches — algorithm-specific tuning parameters
+    Kll { k: u32 },
+    Cms { width: u32, depth: u32 },
+    Hll { precision: u8 },
+    DDSketch { alpha: f64 },
+    CmsWithHeap { width: u32, depth: u32, heap_size: u32 },
+    Kmv { k: u32 },
+    Theta { k: u32 },
+}
+
+// ── Sketch read-out queries ───────────────────────────────────────────────────
+
+/// What to extract from a built sketch. Carried by `SummaryEstimate`.
+#[derive(Debug, Clone)]
+pub enum SketchQuery {
+    /// Extract the value at quantile rank `q` ∈ (0, 1].
+    Quantile { q: f64 },
+    /// Estimated count / frequency for a specific key.
+    PointCount { key: ColumnRef },
+    /// Estimated number of distinct elements.
+    Cardinality,
+    /// Top-k most frequent (key, count) pairs.
+    TopK { k: usize },
+}
diff --git a/crates/core/src/types.rs b/crates/core/src/types.rs
new file mode 100644
index 00000000..4dfad52c
--- /dev/null
+++ b/crates/core/src/types.rs
@@ -0,0 +1,10 @@
+/// Accuracy requirement that a query result must satisfy.
+#[derive(Debug, Clone, PartialEq)]
+pub enum AccuracyTarget {
+    /// Additive error bound ε: |estimate − true| ≤ ε · (domain size).
+    Epsilon(f64),
+    /// Probabilistic (ε, δ) guarantee: error ≤ ε with probability ≥ 1 − δ.
+    EpsilonDelta { epsilon: f64, delta: f64 },
+    /// No approximation permitted; result must be exact.
+    Exact,
+}
diff --git a/crates/core/src/workload.rs b/crates/core/src/workload.rs
new file mode 100644
index 00000000..ee63b07a
--- /dev/null
+++ b/crates/core/src/workload.rs
@@ -0,0 +1,119 @@
+use crate::types::AccuracyTarget;
+
+// ── Query surface ─────────────────────────────────────────────────────────────
+
+/// A raw query string in its source language, before any parsing.
+#[derive(Debug, Clone, PartialEq, Eq, Hash)]
+pub struct Query(pub String);
+
+/// How often a repeating query fires, in milliseconds.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash)]
+pub struct RepetitionInterval(pub u32);
+
+/// SQL dialect variant — different dialects have different syntax and
+/// function sets that affect how the query string is parsed at L1.
+#[derive(Debug, Clone, PartialEq, Eq, Hash)]
+pub enum SqlDialect {
+    DataFusionSQL,
+    ClickhouseSQL,
+    ElasticSQL,
+}
+
+/// Source language of every query in the workload.
+/// All queries in a single `QueryWorkload` share the same language.
+#[derive(Debug, Clone, PartialEq, Eq, Hash)]
+pub enum QueryLanguage {
+    PromQL,
+    SQL(SqlDialect),
+    DataFusion,
+    ElasticDSL,
+}
+
+// ── Per-query requirements ────────────────────────────────────────────────────
+
+/// SLA constraints attached to a single query.
+/// Both fields are optional: an absent bound means "no constraint on this axis."
+#[derive(Debug, Clone)]
+pub struct QueryRequirements {
+    /// Maximum acceptable approximation error.
+    pub accuracy: Option<AccuracyTarget>,
+    /// Maximum acceptable end-to-end query latency in milliseconds.
+    pub latency_ms: Option<f64>,
+}
+
+// ── Workload entries ──────────────────────────────────────────────────────────
+
+/// One entry in a one-shot batch: a query plus its optional SLA constraints.
+#[derive(Debug, Clone)]
+pub struct BatchEntry {
+    pub query: Query,
+    pub requirements: Option<QueryRequirements>,
+}
+
+/// One entry in a repeating (streaming) workload: a query that fires
+/// every `interval` milliseconds.
+#[derive(Debug, Clone)]
+pub struct RepeatingEntry {
+    pub query: Query,
+    /// How often the query fires, in milliseconds.
+    pub interval: RepetitionInterval,
+    pub requirements: Option<QueryRequirements>,
+}
+
+// ── Data characteristics ──────────────────────────────────────────────────────
+
+/// Statistical distribution of keys in the incoming data stream.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub enum DataDistribution {
+    /// Zipf-distributed keys (s ≈ 1.1). A small number of keys dominate,
+    /// so only a fraction of sketch cells are touched per window. Typical
+    /// production case.
+    #[default]
+    Zipf,
+    /// All keys are equally probable. Every window fills the sketch more
+    /// uniformly; delta-compression benefit is lower.
+    Uniform,
+    /// Traffic arrives in bursts with a concentrated key set. Effective
+    /// fill rate is lower on average but spikes can reach Uniform levels.
+    Bursty,
+}
+
+/// Characteristics of the data arriving at the ingestion layer.
+/// Used by the cost model and sketch-parameter binder to size sketches
+/// and estimate transmission cost without running the query.
+#[derive(Debug, Clone)]
+pub struct DataCharacteristics {
+    /// Number of distinct active time series for this metric.
+    pub series_count: u64,
+    /// Sample rate per series at the SDK / agent (Hz).
+    pub samples_per_sec_per_series: f64,
+    /// Wire size of one raw OTLP metric data point after protobuf encoding
+    /// (bytes). Typical range: 50–200 bytes.
+    pub bytes_per_raw_sample: u32,
+    /// Known distinct key values per flush period for frequency / cardinality
+    /// sketches. `None` → inferred analytically from inserts and distribution.
+    pub distinct_keys_per_window: Option<u64>,
+    /// Statistical distribution of keys in the stream.
+    pub data_distribution: DataDistribution,
+}
+
+// ── Top-level workload ────────────────────────────────────────────────────────
+
+/// The single normalised input type accepted by every entry point into the
+/// controller (HTTP POST /plan, YAML file, query-log replay, OpAMP callback).
+///
+/// `query_batch` and `repeating_queries` are mutually exclusive today; both
+/// may be present in the future when mixed batch+streaming workloads are
+/// supported.
+#[derive(Debug, Clone)]
+pub struct QueryWorkload {
+    /// Source language shared by all queries in this workload.
+    pub language: QueryLanguage,
+    /// One-shot queries executed together as a batch.
+    pub query_batch: Option<Vec<BatchEntry>>,
+    /// Queries that repeat on a fixed interval (streaming / continuous).
+    pub repeating_queries: Option<Vec<RepeatingEntry>>,
+    /// Workload-level data characteristics used for sketch sizing and cost
+    /// estimation. Applies to all queries in this workload.
+    pub data_characteristics: Option<DataCharacteristics>,
+}