diff --git a/web/src/engine/predictive-text/worker-thread/src/main/correction/context-state.ts b/web/src/engine/predictive-text/worker-thread/src/main/correction/context-state.ts
index 8c3a8f9931..305f2319a1 100644
--- a/web/src/engine/predictive-text/worker-thread/src/main/correction/context-state.ts
+++ b/web/src/engine/predictive-text/worker-thread/src/main/correction/context-state.ts
@@ -15,7 +15,7 @@ import { ContextTokenization } from './context-tokenization.js';
 import { ContextTransition } from './context-transition.js';
 import { determineModelTokenizer } from '#./model-helpers.js';
 import { tokenizeAndFilterDistribution } from './transform-tokenization.js';
-import { applyTransform, buildMergedTransform } from '@keymanapp/models-templates';
+import { applyTransform } from '@keymanapp/models-templates';
 
 import Context = LexicalModelTypes.Context;
 import Distribution = LexicalModelTypes.Distribution;
@@ -281,7 +281,12 @@ export class ContextState {
 
       for(let i of transformKeys) {
         const primaryInput = transformSequenceDistribution[0].sample.get(i);
-        preservationTransform = preservationTransform ? buildMergedTransform(preservationTransform, primaryInput): primaryInput;
+        if(!preservationTransform) {
+          preservationTransform = primaryInput;
+        } else {
+          preservationTransform.insert += primaryInput.insert;
+          preservationTransform.deleteLeft += primaryInput.deleteLeft;
+        }
       }
     }
 
diff --git a/web/src/engine/predictive-text/worker-thread/src/main/predict-helpers.ts b/web/src/engine/predictive-text/worker-thread/src/main/predict-helpers.ts
index 87c7c68dbd..43fd2deb02 100644
--- a/web/src/engine/predictive-text/worker-thread/src/main/predict-helpers.ts
+++ b/web/src/engine/predictive-text/worker-thread/src/main/predict-helpers.ts
@@ -194,7 +194,6 @@ export async function correctAndEnumerate(
   const postContext = models.applyTransform(inputTransform, context);
   let rawPredictions: CorrectionPredictionTuple[] = [];
 
-  const inputIsBksp = TransformUtils.isBackspace(inputTransform);
   // If `this.contextTracker` does not exist, we don't have the
   // `LexiconTraversal` pattern available to us.  We're unable to efficiently
   // iterate through the lexicon as a result, so we use a far lazier pattern -
@@ -207,6 +206,7 @@ export async function correctAndEnumerate(
 
     // Only allow new-word suggestions if space was the most likely keypress.
     const allowSpace = TransformUtils.isWhitespace(inputTransform);
+    const allowBksp = TransformUtils.isBackspace(inputTransform);
 
     // Generates raw prediction distributions for each valid input.  Can only 'correct'
     // against the final input.
@@ -222,7 +222,7 @@ export async function correctAndEnumerate(
         // Filter out special keys unless they're expected.
         if(TransformUtils.isWhitespace(transform) && !allowSpace) {
           return null;
-        } else if(TransformUtils.isBackspace(transform) && !inputIsBksp) {
+        } else if(TransformUtils.isBackspace(transform) && !allowBksp) {
           return null;
         }
 
@@ -348,7 +348,7 @@ export async function correctAndEnumerate(
   const alignment = postContextState.tokenization.alignment;
 
   // If the context now has more tokens, the token we'll be 'predicting' didn't originally exist.
-  if(transition.preservationTransform && !inputIsBksp) {
+  if(transition.preservationTransform && alignment?.canAlign && alignment.tailTokenShift > 0) {
     // As the word/token being corrected/predicted didn't originally exist, there's no
     // part of it to 'replace'.  (Suggestions are applied to the pre-transform state.)
     deleteLeft = 0;
@@ -921,7 +921,13 @@ export function finalizeSuggestions(
     //
     // Note:  may need adjustment if/when supporting phrase-level correction.
     if(tuple.preservationTransform) {
-      let mergedTransform = models.buildMergedTransform(tuple.preservationTransform, prediction.sample.transform);
+      const presDL = tuple.preservationTransform.deleteLeft;
+      const mergedTransform = models.buildMergedTransform(tuple.preservationTransform, prediction.sample.transform);
+      // Any preserved delete-left is applied early because it directly affects the suggestion
+      // root; we need to remove that preserved delete-left here.
+      if(presDL > 0) {
+        mergedTransform.deleteLeft -= presDL;
+      }
       mergedTransform.id = prediction.sample.transformId;
 
       // Temporarily and locally drops 'readonly' semantics so that we can reassign the transform.
diff --git a/web/src/test/auto/headless/engine/predictive-text/worker-thread/worker-model-compositor.tests.ts b/web/src/test/auto/headless/engine/predictive-text/worker-thread/worker-model-compositor.tests.ts
index d780e84971..1134f5b0a7 100644
--- a/web/src/test/auto/headless/engine/predictive-text/worker-thread/worker-model-compositor.tests.ts
+++ b/web/src/test/auto/headless/engine/predictive-text/worker-thread/worker-model-compositor.tests.ts
@@ -159,6 +159,40 @@ describe('ModelCompositor', function() {
         };
 
         let suggestions = await compositor.predict(inputTransform, context);
+        // Lots of suggestions are rooted on 'the'.  We should have more than
+        // just a single 'keep' suggestion.
+        assert.isAbove(suggestions.length, 1);
+
+        // Verify the deleteLeft length - gotta make sure we get that right.
+        suggestions.forEach(function(suggestion) {
+          // Suggestions always delete the full root of the suggestion.
+          //
+          // After a backspace, that means the text 'the' - 3 chars.
+          // Char 4 is for the original backspace, as suggestions are built
+          // based on the context state BEFORE the triggering input -
+          // here, a backspace.
+          assert.equal(suggestion.transform.deleteLeft, 4);
+        });
+      });
+
+      it('properly handles complex transforms that change the root token', async function() {
+        let compositor = new ModelCompositor(plainModel, true);
+        let context = {
+          left: 'the ', startOfBuffer: true, endOfBuffer: true,
+        };
+
+        let inputTransform = {
+          insert: 'r',
+          deleteLeft: 1
+        };
+
+        let suggestions = await compositor.predict(inputTransform, context);
+        // Lots of suggestions are rooted on 'the'.  We should have more than
+        // just a single 'keep' suggestion.
+        assert.isAbove(suggestions.length, 1);
+        assert.isOk(suggestions.find((s => s.displayAs.startsWith("ther"))));
+
+        // Verify the deleteLeft length - gotta make sure we get that right.
         suggestions.forEach(function(suggestion) {
           // Suggestions always delete the full root of the suggestion.
           //