Correct the r3 comparison's status and remove relative dates

README.md

@@ -168,7 +168,7 @@ reasoning, emitted no answer, and ended in repeated `P3DT4H5M` lines. Their toke
 by two, 16,391 against 16,389, which is the per-round budget overshoot described above, not a
 difference in the answer.
 
-This non-termination was first recorded on this stack two weeks before this drafter existed,
+This non-termination was first recorded on this stack before this drafter was trained,
 and across the author's earlier quality runs it tracks the KV cache rather than the
 drafter. Counting the controls below, the task passes at 16-bit and 4-bit KV in all five runs
 (2,232 to 14,726 tokens), and fails at 8-bit in seven of eight, always by exhausting the budget,
@@ -205,11 +205,10 @@ attention read, and one deterministic greedy run a cell is not a rate.
 
 - [naklitechie's ternary Bonsai 2 fine-tune](https://huggingface.co/naklitechie/Qwen3.8-27B-DFlash2-ternary-bonsai2)
   is an independent DFlash 2 fine-tune against the same target, trained on CUDA. It also beats the
-  stock drafter on both runtimes measured. Compared with ft5 under the repository's pre-registered
-  `BENCHMARK.md` (written by this card's author, who trained ft5), it accepted less on every suite.
-  On general chat at 200 tokens it was 2.41% lower (−3.83% to −1.08%) on its own runtime,
-  dflash-mlx-bonsai2, and 4.63% lower (−6.39% to −2.72%) on mlx-dspark. The raw reports will be
-  published with the evidence.
+  stock drafter on both runtimes measured. On its own runtime, dflash-mlx-bonsai2, measured under the
+  repository's pre-registered `BENCHMARK.md` (written by this card's author, who trained ft5), it
+  accepted 2.41% less than ft5 on general chat at 200 tokens (−3.83% to −1.08%). A pre-publication
+  screen on mlx-dspark, not a protocol result, found 4.63% less (−6.39% to −2.72%).
 - [nathansutton's Bonsai/DFlash2 MLX pack](https://huggingface.co/nathansutton/Qwen3.8-27B-Ternary-Bonsai-2-DFlash2-MLX)
   bundles a converted stock drafter. Its reported 512-token-prompt/128-token-decode result uses
   another runtime and template; it is not a matched comparison to these suites.
@@ -230,7 +229,7 @@ target is unsettled. The NOTICE carries Prism ML's material regardless and takes
 the question. The fine-tune changes the drafter's weights only: it does not contain, modify or
 redistribute the target, which anyone serving this drafter fetches upstream. Training prompts
 come from sahil2801/CodeAlpaca-20k (CC BY 4.0); the corpora are not redistributed here. The
-upstream licences were reviewed, and re-read unchanged just before publication.
+upstream licences were reviewed, and re-read unchanged at publication.
 
 This is an independent fine-tune, not an official Prism ML, Qwen, Inco AI or z-lab release.