Add up/down arrows to results table for quick read

Author: sandeepl337

docs/BENCHMARKS.md

@@ -58,14 +58,23 @@ Same eval slice (`training/FROZEN_EVAL_SCORED.jsonl`, 791 attacks /
 at its published default threshold and at a cross-model neutral
 `0.5`.
 
-| Model | recall@default | FPR@default | recall@0.5 | FPR@0.5 |
+Header arrows show the direction of merit (recall higher = better,
+FPR lower = better). Per cell: ↑ = top-tier on this axis, ↓ =
+bottom-tier, blank = mid.
+
+| Model | recall@default ↑ | FPR@default ↓ | recall@0.5 ↑ | FPR@0.5 ↓ |
 |---|---:|---:|---:|---:|
-| **promptpurify** | **83.94%** | **10.61%** | **87.10%** | **12.88%** |
-| ProtectAI v2 | 40.71% | 43.18% | 40.71% | 43.18% |
-| deepset | 97.22% | 59.85% | 97.22% | 59.85% |
-| fmops | 100.00% | 100.00% | 100.00% | 100.00% |
-| Meta Prompt-Guard | 67.00% | 88.64% | 67.00% | 88.64% |
-| Meta Prompt-Guard-2 | 12.77% | 1.52% | 12.77% | 1.52% |
+| **promptpurify** | **83.94% ↑** | **10.61% ↑** | **87.10% ↑** | **12.88% ↑** |
+| ProtectAI v2 | 40.71% ↓ | 43.18% ↓ | 40.71% ↓ | 43.18% ↓ |
+| deepset | 97.22% ↑ | 59.85% ↓ | 97.22% ↑ | 59.85% ↓ |
+| fmops | 100.00% ↑ | 100.00% ↓ | 100.00% ↑ | 100.00% ↓ |
+| Meta Prompt-Guard | 67.00% | 88.64% ↓ | 67.00% | 88.64% ↓ |
+| Meta Prompt-Guard-2 | 12.77% ↓ | 1.52% ↑ | 12.77% ↓ | 1.52% ↑ |
+
+`promptpurify` is the only row with ↑ on every column. `fmops` "wins"
+recall by predicting positive for every input — its FPR ↓ shows it's
+mis-calibrated, not skilled. `Meta Prompt-Guard-2` flips the trade:
+nearly-zero FPR at the cost of catching ~1 in 8 attacks.
 
 How to read this: