diff --git a/doc.go b/doc.go
index 392838b..e34fb70 100644
--- a/doc.go
+++ b/doc.go
@@ -16,8 +16,7 @@
 // In feedkit today, that maps to:
 //
 //	Collect:     sources.Source + scheduler.Scheduler
-//	Normalize:   (today: domain code typically does this inside Source.Poll;
-//	            future: a normalization Processor is a good fit)
+//	Normalize:   (optional) normalize.Processor (or domain code inside Source.Poll)
 //	Policy:      pipeline.Pipeline (Processor chain; dedupe/ratelimit are planned)
 //	Emit:        dispatch.Dispatcher + dispatch.Fanout
 //	Sinks:       sinks.Sink (+ sinks.Registry to build from config)
@@ -76,6 +75,147 @@
 //
 //   - dedupe/ratelimit processors are placeholders (planned).
 //
+//   - normalize
+//     Optional normalization hook for splitting "fetch" from "transform".
+//
+//     Many domains (like weather) ingest multiple upstream providers whose payloads
+//     differ. A common evolution is to keep sources small and focused on polling,
+//     and move mapping/normalization into a dedicated stage.
+//
+//     feedkit provides this as an OPTIONAL pipeline processor:
+//
+//   - normalize.Normalizer: domain-implemented mapping logic
+//
+//   - normalize.Registry: holds normalizers and selects one by Match()
+//
+//   - normalize.Processor: adapts Registry into a pipeline.Processor
+//
+//     Normalization is NOT required:
+//
+//   - If you do all normalization inside Source.Poll, you can ignore this package.
+//
+//   - If normalize.Processor is not installed in your pipeline, nothing changes.
+//
+//   - If normalize.Processor is installed but no Normalizer matches an event,
+//     the event passes through unchanged.
+//
+//     The key types:
+//
+//     type Normalizer interface {
+//     // Match returns true if this normalizer should handle the event.
+//     // Matching is intentionally flexible: match on Schema, Kind, Source,
+//     // or any combination.
+//     Match(e event.Event) bool
+//
+//     // Normalize converts the incoming event into a new (or modified) event.
+//     //
+//     // Return values:
+//     //   - (out, nil) where out != nil: emit the normalized event
+//     //   - (nil, nil): drop the event (policy drop)
+//     //   - (nil, err): fail the pipeline
+//     Normalize(ctx context.Context, in event.Event) (*event.Event, error)
+//     }
+//
+//     type Registry struct { ... }
+//
+//     func (r *Registry) Register(n Normalizer)
+//
+//     // Normalize finds the first matching normalizer (in registration order) and applies it.
+//     // If none match, it returns the input event unchanged.
+//     func (r *Registry) Normalize(ctx context.Context, in event.Event) (*event.Event, error)
+//
+//     // Processor implements pipeline.Processor and calls into the Registry.
+//     // Optional behavior:
+//     //   - If Registry is nil, Processor is a no-op pass-through.
+//     //   - If RequireMatch is false (default), non-matching events pass through.
+//     //   - If RequireMatch is true, non-matching events are treated as errors.
+//     type Processor struct {
+//     Registry     *Registry
+//     RequireMatch bool
+//     }
+//
+//     "First match wins":
+//     Registry applies the first Normalizer whose Match() returns true.
+//     This is intentional: normalization is usually a single mapping step from a
+//     raw schema into a canonical schema. If you want multiple sequential transforms,
+//     model them as multiple pipeline processors.
+//
+//     Recommended convention: match by Event.Schema
+//     ------------------------------------------------
+//     Schema gives you a versionable selector that doesn't depend on source names.
+//
+//     A common pattern is:
+//
+//   - sources emit "raw" events with Schema like:
+//     "raw.openweather.current.v1"
+//     "raw.openmeteo.current.v1"
+//     "raw.nws.observation.v1"
+//
+//   - normalizers transform them into canonical domain schemas like:
+//     "weather.observation.v1"
+//     "weather.forecast.v1"
+//     "weather.alert.v1"
+//
+//     What is a "raw event"?
+//     ------------------------------------------------
+//     feedkit does not prescribe the raw payload representation.
+//     A raw payload is typically one of:
+//
+//   - json.RawMessage (recommended for JSON APIs)
+//
+//   - []byte (raw bytes)
+//
+//   - map[string]any (already-decoded but untyped JSON)
+//
+//     The only hard requirement enforced by feedkit is Event.Validate():
+//
+//   - ID, Kind, Source, EmittedAt must be set
+//
+//   - Payload must be non-nil
+//
+//     If you use raw events, you still must provide Event.Kind.
+//     Typical approaches:
+//
+//   - set Kind to the intended canonical kind (e.g. "observation") even before normalization
+//
+//   - or set Kind to a domain-defined "raw_*" kind and normalize it later
+//
+//     The simplest approach is: set Kind to the final kind early, and use Schema
+//     to describe the raw-vs-normalized payload shape.
+//
+//     Wiring example (daemon main.go)
+//     ------------------------------------------------
+//     Install normalize.Processor at the front of your pipeline:
+//
+//     normReg := &normalize.Registry{}
+//
+//     normReg.Register(normalize.Func{
+//     Name: "openweather current -> weather.observation.v1",
+//     MatchFn: func(e event.Event) bool {
+//     return e.Schema == "raw.openweather.current.v1"
+//     },
+//     NormalizeFn: func(ctx context.Context, in event.Event) (*event.Event, error) {
+//     // 1) interpret in.Payload (json.RawMessage / []byte / map)
+//     // 2) build canonical domain payload
+//     // 3) return updated event
+//
+//     out := in
+//     out.Schema = "weather.observation.v1"
+//     // Optionally adjust Kind, EffectiveAt, etc.
+//     out.Payload = /* canonical weather observation struct */
+//     return &out, nil
+//     },
+//     })
+//
+//     p := &pipeline.Pipeline{
+//     Processors: []pipeline.Processor{
+//     normalize.Processor{Registry: normReg}, // optional stage
+//     // dedupe.New(...), ratelimit.New(...), ...
+//     },
+//     }
+//
+//     If the event does not match any normalizer, it passes through unmodified.
+//
 //   - sinks
 //     Extension point for output adapters.
 //
@@ -141,13 +281,24 @@
 //	// Event bus.
 //	bus := make(chan event.Event, 256)
 //
+//	// Optional normalization registry + pipeline.
+//	normReg := &normalize.Registry{}
+//	// domain registers normalizers into normReg...
+//
+//	p := &pipeline.Pipeline{
+//	    Processors: []pipeline.Processor{
+//	        normalize.Processor{Registry: normReg}, // optional
+//	        // dedupe/ratelimit/etc...
+//	    },
+//	}
+//
 //	// Scheduler.
 //	s := &scheduler.Scheduler{Jobs: jobs, Out: bus, Logf: logf}
 //
 //	// Dispatcher.
 //	d := &dispatch.Dispatcher{
 //	    In:       bus,
-//	    Pipeline: &pipeline.Pipeline{Processors: nil},
+//	    Pipeline: p,
 //	    Sinks:    builtSinks,
 //	    Routes:   routes,
 //	}
@@ -167,13 +318,12 @@
 // All blocking or I/O work should honor ctx.Done():
 //   - sources.Source.Poll should pass ctx to HTTP calls, etc.
 //   - sinks.Sink.Consume should honor ctx (Fanout timeouts only help if sinks cooperate).
+//   - normalizers should honor ctx if they do expensive work (rare; usually pure transforms).
 //
 // Future additions (likely)
 //
 //   - A small Runner helper that performs the standard wiring (load config,
 //     build sources/sinks/routes, run scheduler+dispatcher, handle shutdown).
-//   - A normalization hook (a Pipeline Processor + registry) that allows sources
-//     to emit "raw" payloads and defer normalization to a dedicated stage.
 //
 // # Non-goals
 //
diff --git a/normalize/doc.go b/normalize/doc.go
new file mode 100644
index 0000000..f6dc773
--- /dev/null
+++ b/normalize/doc.go
@@ -0,0 +1,17 @@
+// Package normalize provides an OPTIONAL normalization hook for feedkit pipelines.
+//
+// Motivation:
+// Many daemons have sources that:
+//  1. fetch raw upstream data (often JSON), and
+//  2. transform it into a domain's normalized payload format.
+//
+// Doing both steps inside Source.Poll works, but tends to make sources large and
+// encourages duplication (unit conversions, common mapping helpers, etc.).
+//
+// This package lets a source emit a "raw" event (e.g., Schema="raw.openweather.current.v1",
+// Payload=json.RawMessage), and then a normalization processor can convert it into a
+// normalized event (e.g., Schema="weather.observation.v1", Payload=WeatherObservation{}).
+//
+// Key property: normalization is optional.
+// If no registered Normalizer matches an event, it passes through unchanged.
+package normalize
diff --git a/normalize/normalize.go b/normalize/normalize.go
new file mode 100644
index 0000000..2f75b0e
--- /dev/null
+++ b/normalize/normalize.go
@@ -0,0 +1,76 @@
+package normalize
+
+import (
+	"context"
+	"fmt"
+
+	"gitea.maximumdirect.net/ejr/feedkit/event"
+)
+
+// Normalizer converts one event shape into another.
+//
+// A Normalizer is typically domain-owned code (weatherfeeder/newsfeeder/...)
+// that knows how to interpret a specific upstream payload and produce a
+// normalized payload.
+//
+// Normalizers are selected via Match(). The matching strategy is intentionally
+// flexible: implementations may match on Schema, Kind, Source, or any other
+// Event fields.
+type Normalizer interface {
+	// Match reports whether this normalizer applies to the given event.
+	//
+	// Common patterns:
+	//   - match on e.Schema (recommended for versioning)
+	//   - match on e.Source (useful if Schema is empty)
+	//   - match on (e.Kind + e.Source), etc.
+	Match(e event.Event) bool
+
+	// Normalize transforms the incoming event into a new (or modified) event.
+	//
+	// Return values:
+	//   - (out, nil) where out != nil: emit the normalized event
+	//   - (nil, nil): drop the event (treat as policy drop)
+	//   - (nil, err): fail the pipeline
+	//
+	// Note: If you simply want to pass the event through unchanged, return &in.
+	Normalize(ctx context.Context, in event.Event) (*event.Event, error)
+}
+
+// Func is an ergonomic adapter that lets you define a Normalizer with functions.
+//
+// Example:
+//
+//	n := normalize.Func{
+//	  MatchFn: func(e event.Event) bool { return e.Schema == "raw.openweather.current.v1" },
+//	  NormalizeFn: func(ctx context.Context, in event.Event) (*event.Event, error) {
+//	    // ... map in.Payload -> normalized payload ...
+//	  },
+//	}
+type Func struct {
+	MatchFn     func(e event.Event) bool
+	NormalizeFn func(ctx context.Context, in event.Event) (*event.Event, error)
+
+	// Optional: helps produce nicer panic/error messages if something goes wrong.
+	Name string
+}
+
+func (f Func) Match(e event.Event) bool {
+	if f.MatchFn == nil {
+		return false
+	}
+	return f.MatchFn(e)
+}
+
+func (f Func) Normalize(ctx context.Context, in event.Event) (*event.Event, error) {
+	if f.NormalizeFn == nil {
+		return nil, fmt.Errorf("normalize.Func(%s): NormalizeFn is nil", f.safeName())
+	}
+	return f.NormalizeFn(ctx, in)
+}
+
+func (f Func) safeName() string {
+	if f.Name == "" {
+		return "<unnamed>"
+	}
+	return f.Name
+}
diff --git a/normalize/registry.go b/normalize/registry.go
new file mode 100644
index 0000000..054159b
--- /dev/null
+++ b/normalize/registry.go
@@ -0,0 +1,140 @@
+package normalize
+
+import (
+	"context"
+	"fmt"
+	"sync"
+
+	"gitea.maximumdirect.net/ejr/feedkit/event"
+)
+
+// Registry holds a set of Normalizers and selects one for a given event.
+//
+// Selection rule (simple + predictable):
+//   - iterate in registration order
+//   - the FIRST Normalizer whose Match(e) returns true is used
+//
+// If none match, the event passes through unchanged.
+//
+// Why "first match wins"?
+// Normalization is usually a single mapping step from a raw schema/version into
+// a normalized schema/version. If you want multiple transformation steps,
+// model them as multiple pipeline processors (which feedkit already supports).
+type Registry struct {
+	mu sync.RWMutex
+	ns []Normalizer
+}
+
+// Register adds a normalizer to the registry.
+//
+// Register panics if n is nil; this is a programmer error and should fail fast.
+func (r *Registry) Register(n Normalizer) {
+	if n == nil {
+		panic("normalize.Registry.Register: normalizer cannot be nil")
+	}
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.ns = append(r.ns, n)
+}
+
+// Normalize finds the first matching Normalizer and applies it.
+//
+// If no normalizer matches, it returns the input event unchanged.
+//
+// If a normalizer returns (nil, nil), the event is dropped.
+func (r *Registry) Normalize(ctx context.Context, in event.Event) (*event.Event, error) {
+	if r == nil {
+		// Nil registry is a valid "feature off" state.
+		out := in
+		return &out, nil
+	}
+
+	r.mu.RLock()
+	ns := append([]Normalizer(nil), r.ns...) // copy for safe iteration outside lock
+	r.mu.RUnlock()
+
+	for _, n := range ns {
+		if n == nil {
+			// Shouldn't happen (Register panics), but guard anyway.
+			continue
+		}
+		if !n.Match(in) {
+			continue
+		}
+
+		out, err := n.Normalize(ctx, in)
+		if err != nil {
+			return nil, fmt.Errorf("normalize: normalizer failed: %w", err)
+		}
+		// out may be nil to signal "drop".
+		return out, nil
+	}
+
+	// No match: pass through unchanged.
+	out := in
+	return &out, nil
+}
+
+// Processor adapts a Registry into a pipeline Processor.
+//
+// It implements:
+//
+//	Process(ctx context.Context, in event.Event) (*event.Event, error)
+//
+// which matches feedkit/pipeline.Processor.
+//
+// Optionality:
+//   - If Registry is nil, Processor becomes a no-op pass-through.
+//   - If Registry has no matching normalizer for an event, that event passes through unchanged.
+type Processor struct {
+	Registry *Registry
+
+	// If true, events that do not match any normalizer cause an error.
+	// Default is false (pass-through), which is the behavior you asked for.
+	RequireMatch bool
+}
+
+// Process implements the pipeline.Processor interface.
+func (p Processor) Process(ctx context.Context, in event.Event) (*event.Event, error) {
+	// "Feature off": no registry means no normalization.
+	if p.Registry == nil {
+		out := in
+		return &out, nil
+	}
+
+	out, err := p.Registry.Normalize(ctx, in)
+	if err != nil {
+		return nil, err
+	}
+
+	if out == nil {
+		// Dropped by normalization policy.
+		return nil, nil
+	}
+
+	if p.RequireMatch {
+		// Detect "no-op pass-through due to no match" by checking whether a match existed.
+		// We do this with a cheap second pass to avoid changing Normalize()'s signature.
+		// (This is rare to enable; correctness/clarity > micro-optimization.)
+		if !p.Registry.hasMatch(in) {
+			return nil, fmt.Errorf("normalize: no normalizer matched event (id=%s kind=%s source=%s schema=%q)",
+				in.ID, in.Kind, in.Source, in.Schema)
+		}
+	}
+
+	return out, nil
+}
+
+func (r *Registry) hasMatch(in event.Event) bool {
+	if r == nil {
+		return false
+	}
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+	for _, n := range r.ns {
+		if n != nil && n.Match(in) {
+			return true
+		}
+	}
+	return false
+}