diff --git a/lib/unit/system.rb b/lib/unit/system.rb
index c3fbfbf..ccd57e7 100644
--- a/lib/unit/system.rb
+++ b/lib/unit/system.rb
@@ -47,10 +47,6 @@ def load(input)
 
       @unit.each {|name, unit| validate_unit(unit[:def]) }
 
-      # Newly registered symbols may introduce glyph characters the lexer must
-      # recognize; rebuild it lazily on next parse.
-      @tokenizer = nil
-
       true
     end
 
@@ -66,7 +62,7 @@ def validate_unit(units)
 
     def parse_unit(expr)
       stack, result, implicit_mul = [], [], false
-      expr.to_s.scan(tokenizer).each do |tok|
+      expr.to_s.scan(TOKENIZER).each do |tok|
         if tok == '('
           stack << '('
           implicit_mul = false
@@ -102,24 +98,15 @@ def parse_unit(expr)
     OPERATOR_TOKENS = OPERATOR.keys.sort_by {|x| -x.size }. map {|x| Regexp.quote(x) }
     OPERATOR_CHARS = (OPERATOR.keys.join + '()').chars.uniq.freeze
 
-    # The tokenizer and symbol matcher are derived from the symbols actually
-    # registered in this system (plus an ASCII baseline), then memoized.
-    # #load clears the memo, so any unit defined later — including ones whose
-    # symbol uses a non-ASCII glyph (Ω, ℃, µ, %) — becomes lexable without
-    # touching the lexer. This replaces a frozen glyph char-class that silently
-    # dropped unrecognized glyphs to a dimensionless value.
-    def tokenizer
-      @tokenizer ||= Regexp.new(
-        (OPERATOR_TOKENS + [REAL.source[1..-2], DEC.source[1..-2], symbol_pattern, '\\(', '\\)']).join('|')
-      )
-    end
-
-    def symbol_pattern
-      glyphs = (unit_symbol.keys + factor_symbol.keys).join.chars.uniq.reject do |char|
-        char =~ /[A-Za-z0-9_]/ || char =~ /\s/ || OPERATOR_CHARS.include?(char)
-      end
-      "[A-Za-z0-9_#{glyphs.map { |char| Regexp.escape(char) }.join}]+"
-    end
+    # A symbol is any run of characters that are not whitespace, parentheses, or
+    # operator characters. Crucially this also matches *unrecognized* glyphs
+    # (e.g. a stray Greek letter): they survive lexing and then fail loudly as
+    # an "Undefined unit" during validation, exactly like an unknown ASCII
+    # symbol does, instead of being silently dropped to a dimensionless value.
+    SYMBOL = Regexp.new("[^\\s#{OPERATOR_CHARS.map { |char| Regexp.escape(char) }.join}]+")
+    TOKENIZER = Regexp.new(
+      (OPERATOR_TOKENS + [REAL.source[1..-2], DEC.source[1..-2], SYMBOL.source, '\\(', '\\)']).join('|')
+    )
 
     def lookup_symbol(symbol)
       if unit_symbol[symbol]
diff --git a/spec/system_spec.rb b/spec/system_spec.rb
index 4b24eea..c5bcd2b 100644
--- a/spec/system_spec.rb
+++ b/spec/system_spec.rb
@@ -127,4 +127,49 @@
       expect(Unit(50, '%', system)).to eq(Unit(Rational(1, 2), system))
     end
   end
+
+  describe "rejecting symbols it cannot resolve" do
+    # The lexer matches any run of non-operator, non-whitespace characters as a
+    # symbol — including glyphs that no system has registered — so an unknown
+    # symbol reaches validation and fails loudly as "Undefined unit" instead of
+    # being silently dropped to a dimensionless value. This keeps unknown
+    # glyphs consistent with unknown ASCII symbols rather than special-casing
+    # them into silence.
+    before { system.load(:si) }
+
+    it "raises for an unknown ASCII symbol" do
+      expect { Unit(1, "foo", system) }.to raise_error(TypeError, "Undefined unit foo")
+    end
+
+    it "raises for an unregistered glyph instead of silently dropping it" do
+      expect { Unit(1, "λ", system) }.to raise_error(TypeError, "Undefined unit λ")
+      expect { Unit(5, "αβγ", system) }.to raise_error(TypeError, "Undefined unit αβγ")
+    end
+
+    it "raises when an unregistered glyph is implicitly multiplied with a known unit" do
+      expect { Unit(1, "m λ", system) }.to raise_error(TypeError, "Undefined unit λ")
+      expect { Unit(1, "λ m", system) }.to raise_error(TypeError, "Undefined unit λ")
+    end
+
+    it "names the unregistered glyph rather than a neighboring operator" do
+      # The glyph used to be dropped, leaving a dangling operator that raised a
+      # misleading "Unexpected token /"; now the real culprit is reported.
+      expect { Unit(1, "m/λ", system) }.to raise_error(TypeError, "Undefined unit λ")
+      expect { Unit(1, "λ^2", system) }.to raise_error(TypeError, "Undefined unit λ")
+    end
+  end
+
+  describe "parsing a unit whose system has not been loaded" do
+    # Directly answers "what happens with unit strings that aren't loaded": the
+    # symbol is lexed but unresolved, so it fails loudly until the defining
+    # system is loaded, after which it resolves normally.
+    it "raises until the defining system is loaded, then resolves" do
+      system.load(:si)
+      # megaelectronvolt is defined in :scientific, which is not loaded yet.
+      expect { Unit(1, "MeV", system) }.to raise_error(TypeError, "Undefined unit MeV")
+
+      system.load(:scientific)
+      expect(Unit(1, "MeV", system).unit).to eq(Unit(1, "megaelectronvolt", system).unit)
+    end
+  end
 end