add README.md

README.md

@@ -22,7 +22,7 @@ picking the next action for a robot: same call, same model.
 
 - **Quality.** Mean accuracy 0.801 across 17 public suites (Seb-9B 0.793, hosted Jev 1.13 0.786, Laya 0.694). Biggest gains over v1:
   BANKING77 0.973, jailbreak detection 0.940, model routing 0.980. In our 3D robot-arm demos all 300 decisions matched the safety
-  rules. Against general LLMs on 420 public decision rows it is level with gpt-6-luna (high) and behind Claude Sonnet and Opus (table below). One step back: v1 is still better calibrated (ECE 0.088 vs 0.110).
+  rules. Against general LLMs on 420 public decision rows it is level with gpt-6-luna (high) and Claude Haiku (medium) and behind Claude Sonnet and Opus (table below). One step back: v1 is still better calibrated (ECE 0.088 vs 0.110).
 - **Speed.** About 22 ms per decision on one RTX PRO 6000 (p95 36 ms) and about 88 ms on an Apple M5 Max (p95 209 ms), up to about
   87 decisions per second on a single GPU.
 - **Decision Index.** 59.16 on the 0.3 public index, against 57.96 for hosted Jev 1.13 and 60.70 for Jev 1.14. We are ahead of
@@ -76,17 +76,19 @@ picking the next action for a robot: same call, same model.
 Same 420 public decision rows (20 from each of 21 public suites, fixed seed), same prompt content (state, question, options), each model answering blind;
 gold is the suite's own label. The LLMs return an answer; Seb-9B-v2 also returns a probability for every option.
 
-| | Correct | Accuracy | 95% interval |
-|---|---|---|---|
-| Claude Opus | 359 / 420 | 0.855 | 0.82 - 0.89 |
-| Claude Sonnet | 349 / 420 | 0.831 | 0.79 - 0.86 |
-| **Seb-9B-v2** | 335 / 420 | **0.798** | 0.76 - 0.83 |
-| gpt-6-luna (high) | 327 / 420 | 0.779 | 0.74 - 0.82 |
-
-Claude models are Claude Code's `opus` and `sonnet` aliases on 2026-10-22; gpt-6-luna ran through our JSON-schema labeling harness (`scripts/label_with_codex.py`, high effort).
-Seb-9B-v2 vs gpt-6-luna is within noise (26 rows only Seb right, 18 only Luna, sign test p = 0.29). Claude Sonnet is ahead by 3 points (not
-significant at 0.05, p = 0.054) and Opus by 6 points (p = 0.0003). The general models are much larger and need a hosted API call per decision;
-Seb-9B-v2 runs locally. Rows and scripts: `opus_eval/` in the demo folder.
+| | Correct | Accuracy | 95% interval | vs Seb-9B-v2 |
+|---|---|---|---|---|
+| Claude Opus | 359 / 420 | 0.855 | 0.82 - 0.89 | ahead (p = 0.0003) |
+| Claude Sonnet (default effort) | 349 / 420 | 0.831 | 0.79 - 0.86 | ahead by 3 points (p = 0.054) |
+| Claude Sonnet (low effort) | 343 / 420 | 0.817 | 0.78 - 0.85 | ahead by 2 points (p = 0.26) |
+| **Seb-9B-v2** | 335 / 420 | **0.798** | 0.76 - 0.83 | |
+| Claude Haiku (medium effort) | 334 / 420 | 0.795 | 0.75 - 0.83 | level (p = 1.0) |
+| gpt-6-luna (high) | 327 / 420 | 0.779 | 0.74 - 0.82 | level (p = 0.29) |
+
+Claude models are Claude Code's `opus`, `sonnet` and `haiku` aliases on 2026-10-22 (one Haiku answer was missing and counts as wrong); gpt-6-luna ran through our
+JSON-schema labeling harness (`scripts/label_with_codex.py`, high effort). p-values are exact sign tests on the rows where exactly one of the two was right.
+Seb-9B-v2 is level with Claude Haiku (medium) and gpt-6-luna and statistically level with Sonnet at low effort; Sonnet at default effort is 3 points ahead and Opus 6 points
+ahead (significant). The general models are much larger and need a hosted API call per decision; Seb-9B-v2 runs locally. Rows and scripts: `opus_eval/` in the demo folder.
 
 ## Demo results