Polish README wording and clarify runtime distinctions

README.md

@@ -84,7 +84,7 @@ This model is designed as a **coding agent model** for use with Claude Code, Ope
 
 ## Why IQ4_NL is Recommended Over BF16, Q8, and Q6
 
-This is not typical. Through extensive real-world testing, the **author's preferred IQ4_NL quantization with custom importance matrix consistently outperformed Q8_0, Q6_K, and full-precision BF16**:
+This is not typical. In the author's tested coding / agent workloads, the **preferred IQ4_NL quantization with custom importance matrix consistently outperformed Q8_0, Q6_K, and full-precision BF16**:
 
 1. **Quantization acts as a regularizer** — slight weight rounding prevents degenerate thinking loops that BF16 and Q6_K are prone to
 2. **Custom imatrix preserves reasoning weights** — 30% reasoning/self-verification calibration data ensures chain-of-thought and self-correction weights are preserved
@@ -106,7 +106,7 @@ The IQ4_NL with custom imatrix was the **only variant that produced working data
 
 **Supporting evidence: 5 automated coding problems (merge intervals, LCS, RPN evaluator, valid parentheses, trapping rain water) tested at both temperature 0.6 and 0.8 — IQ4_NL with custom imatrix scored 5/5 at both temperatures.** BF16 and Q6_K both entered infinite thinking loops on the same problems.
 
-For clarity: the results above come from the **ik-llama.cpp-oriented IQ4_NL build**, which remains the author's primary recommendation. A separate standard llama.cpp-compatible IQ4_NL build is now provided for portability.
+For clarity: the results above come from the **ik-llama.cpp-oriented IQ4_NL build**, which remains the author's primary recommendation. A separate standard llama.cpp-compatible IQ4_NL build is provided for portability.
 
 ## Critical: KV Cache Recommendations
 
@@ -197,12 +197,23 @@ The resulting standard llama.cpp-compatible file still uses a mixed quant layout
 
 The author still personally recommends the **ik-llama** build because that is the day-to-day driver and the one that will be exercised the most in live coding / agent workloads.
 
+Both IQ4_NL releases are mixed quants. The key difference is that the `ik-llama` build uses ik-specific tensor types, while the `llama.cpp-compatible` build uses mainline-supported tensor types.
+
 ## Recommended Parameters
 
 These are the **current best test parameters** for this repo so far:
 
 ```bash
-llama-server   -m RYS-Qwen3.5-27B-Uncensored-Splice-IQ4_NL-ik-llama.gguf   --host 0.0.0.0 --port 8001   -ngl 99 -c 262144   --cache-type-k f16 --cache-type-v f16   --cache-ram 30720   --flash-attn on   --jinja --reasoning-format deepseek   --temp 0.7 --top-p 0.95 --top-k 20   --min-p 0.0 --repeat-penalty 1.0
+llama-server \
+  -m RYS-Qwen3.5-27B-Uncensored-Splice-IQ4_NL-ik-llama.gguf \
+  --host 0.0.0.0 --port 8001 \
+  -ngl 99 -c 262144 \
+  --cache-type-k f16 --cache-type-v f16 \
+  --cache-ram 30720 \
+  --flash-attn on \
+  --jinja --reasoning-format deepseek \
+  --temp 0.7 --top-p 0.95 --top-k 20 \
+  --min-p 0.0 --repeat-penalty 1.0
 ```
 
 **Current params:**
@@ -274,7 +285,7 @@ Primary validation for this release used **ik-llama.cpp**, mainly because its **
 | Flash Attention | ON (`GGML_CUDA_FA_ALL_QUANTS`) |
 | Build Date | Apr 5, 2026 |
 
-The author uses ik-llama because graph split performs much better on multi-GPU setups. This remains the author's **personal daily driver** and the variant that will receive the most testing.
+The author uses ik-llama because graph split performs much better on multi-GPU setups. It remains the author's **personal daily driver** and the variant that will receive the most testing.
 
 ### standard llama.cpp (used for the new compatibility build)
 
@@ -299,14 +310,14 @@ llama-quantize \
   IQ4_NL
 ```
 
-Compatibility test load used:
+Compatibility validation used:
 
 - `CUDA_VISIBLE_DEVICES=0,1,2`
 - `--tensor-split 3,2,2`
 - `-c 262144`
 - `--cache-type-k f32 --cache-type-v f32`
 
-Verification prompts returned:
+Verification prompts returned correct final outputs:
 
 - `0.05`
 - `5050`