Revise release card benchmark notes

README.md

@@ -28,6 +28,10 @@ This is a normal `llama.cpp` GGUF quant of DavidAU's Deckard Gemma model:
 
 The goal was a small, practical creative/voice/writing model that keeps more coherence than a plain Q4-style quant while still fitting comfortably on modest GPUs.
 
+This was an unplanned release. The author was testing models for creative writing use, found a quantized GGUF of this model that showed severe looping and quality degradation, and built this version to check whether the issue came from the core finetune or from the quantization. The result was unexpectedly useful: creative, able to reason, good at following instructions, strong for roleplay, and expressive in a way that made it enjoyable to use. Because of that, it was uploaded for anyone else interested in this niche.
+
+Because this was exploratory rather than a planned release, it has not gone through the author's usual comprehensive release process. A gender-affirming finetune based on the DavidAU model may come later for people interested in that direction.
+
 ## What Is Special
 
 This is not a plain one-shot Q4 quant.
@@ -123,54 +127,33 @@ Local file sizes from the experiment:
 | IQ4_NL + attnQ8 last12 | 5,283,415,040 bytes |
 | Q6_K reference | 6,184,790,688 bytes |
 
-The `last6` mixed quant is only about 45 MB larger than the comparable b8763 IQ4_NL build, while testing better in our small creative/voice/coherence probe.
+The `last6` mixed quant is only about 45 MB larger than the comparable b8763 IQ4_NL build.
 
-## Test Results
+## Runtime Tests
 
-These are local smoke/proxy tests, not a formal benchmark suite. The prompts targeted:
+These are direct runtime probes for this exact released GGUF using standard `llama.cpp`. No custom fork was used.
 
-- short creative scene writing
-- transfeminine voice-practice instruction quality
-- coherence/repetition stress with port-city micro-scenes
+| GPU | Runtime | Context / KV | Prompt workload | Decode workload | Prompt tok/s | Decode tok/s | VRAM note |
+|---|---|---|---:|---:|---:|---:|---|
+| RTX 5060 Ti | `llama.cpp` b8763 server | `-c 120000`, f32 K/V, all layers offloaded | 1352 tokens | 256 tokens | 2816.56 | 73.54 | about 7796 MiB observed |
+| RTX 5090 | `llama.cpp` b8763 server | `-c 120000`, f32 K/V, all layers offloaded | 1122 tokens | 40 tokens | 4251.93 | 138.82 | 7159 MiB llama.cpp self allocation |
 
-Best mini-eval run:
+5060 Ti log summary:
 
 ```text
-docs/dynamic_q8_mini_20260429_112232_report.md
-```
-
-Summary from that run:
-
-| Model | Proxy score | Coverage | Worst 4-gram repeat | Decode tok/s | Prompt tok/s |
-|---|---:|---:|---:|---:|---:|
-| attnQ8_last6 | 76.407 | 0.489 | 1 | 86.579 | 1354.519 |
-| Q6_K | 73.778 | 0.422 | 1 | 72.340 | 1480.222 |
-| Q8_0 | 73.278 | 0.411 | 1 | 61.660 | 1502.476 |
-| IQ4_NL b8763 | 70.185 | 0.356 | 1 | 87.449 | 1300.931 |
-| attnQ8_last12 | 68.426 | 0.300 | 1 | 85.448 | 1372.260 |
-
-Interpretation: `attnQ8_last6` was the best quality/size tradeoff in this test. The broader `attnQ8_last12` variant was worse, so simply adding more Q8 was not helpful.
-
-## 5060 Ti Runtime Probe
-
-On one RTX 5060 Ti, using standard `llama.cpp` with f32 KV and `-c 120000`:
-
-```text
-offloaded 43/43 layers to GPU
 CUDA0 model buffer size = 2775.92 MiB
 non-SWA KV buffer size = 3752.00 MiB
 SWA KV buffer size = 60.00 MiB
 compute buffer size = 572.00 MiB
-observed VRAM = about 7796 MiB
-prompt = 2816.56 tok/s on a 1352-token prompt
-decode = 73.54 tok/s for 256 generated tokens
 ```
 
-Small-context kernel probe on the same GPU:
+5090 log summary:
 
 ```text
-prompt = 5558.54 tok/s
-decode = 105.80 tok/s
+CUDA0 self allocation = 7159 MiB
+model buffer = 2775 MiB
+context buffer = 3812 MiB
+compute buffer = 572 MiB
 ```
 
 ## Intended Use
@@ -186,11 +169,11 @@ Because the source model is an uncensored/aliterated model, apply your own safet
 3. Built an English creative/voice/logic imatrix calibration set.
 4. Quantized with `IQ4_NL`.
 5. Re-quantized selected late attention tensors in blocks `36..41` to `Q8_0`.
-6. Tested against IQ4_NL, Q6_K, Q8_0, and a wider `attnQ8_last12` candidate.
+6. Runtime-tested the released GGUF on RTX 5060 Ti and RTX 5090 with standard `llama.cpp`.
 
 ## Limitations
 
-- The reported eval is a small automatic proxy eval plus user testing, not a broad benchmark.
+- This was an unplanned exploratory release, not a broad benchmarked model release.
+- Public runtime numbers are limited to the two direct tests above.
 - The no-thinking template is recommended for creative use because it reduces visible planning/scaffold output.
 - This is a GGUF release for `llama.cpp`; it is not a safetensors checkpoint.
-