More clarifications in README

README.md

@@ -31,7 +31,7 @@ flowchart TD
 
     D -->|No| E[SimdJSON DOM<br/>Easy to use<br/>Good performance]
 
-    D -->|Yes| F[SimdJSON On-Demand<br/>Fastest option<br/>Forward-only traversal]
+    D -->|Yes| F[SimdJSON On-Demand<br/>Very fast<br/>Forward-only traversal]
 ```
 
 ## When to Use What
@@ -40,19 +40,22 @@ flowchart TD
 - You want to write `obj["key"]` and have it just work
 - JSON files are reasonably sized (but you still want decent performance)
 - You care about getting things done
+- You need full document validation upfront
 
 ### Use **SimdJSON On-Demand** when:
 - Performance is critical and your data fits in memory
 - You can work with forward-only traversal (no random access or backtracking)
 - You need maximum speed but still want a usable API
 - Your JSON structure is consistent and you control both generation and parsing
+- You're okay with partial validation (only validates parts you actually access)
+- Your JSON keys don't contain escape sequences (OnDemand matches raw keys without unescaping)
 
 ### Use **WeaselJSON** when:
 - You absolutely cannot load the whole JSON into memory
 - You're processing huge JSON files (multiple gigabytes)
 - You want to parse just part of a JSON file without reading the rest
 - You need to convert JSON into some other format as you parse it
-- You can't risk accidentally slow performance
+- You need predictable performance characteristics
 - Writing stateful callbacks is your idea of fun
 
 ## The Reality
@@ -70,7 +73,7 @@ The parser represents excellent engineering work and occupies a useful niche. It
 - No memory allocations during parsing
 - O(1) memory usage regardless of input size
 - Streaming API - no need to buffer the entire document in memory. Parsing is resumed when more data is available
-- Strings are unescaped in place before they're presented. No unicode normalization is performed
+- Strings are unescaped in place before they're presented (modifies your input buffer unless WeaselJsonRaw flag is used). No unicode normalization is performed
 - Robust to crashes with untrusted input
 - SIMD optimizations for string scanning and validation