Mark O2 checkpoints as preview and remove citation

README.md

@@ -123,7 +123,7 @@ O2-9B-Preview was evaluated across eleven reasoning, tool-use, and coding benchm
       <td colspan="13">&nbsp;</td>
     </tr>
     <tr>
-      <td>O2–<br>9B<br>(Ours)</td>
+      <td>O2-9B-Preview<br>(Ours)</td>
       <td align="center">64.07</td>
       <td align="center">87.37</td>
       <td align="center">84.44</td>
@@ -138,7 +138,7 @@ O2-9B-Preview was evaluated across eleven reasoning, tool-use, and coding benchm
       <td align="center">30.00</td>
     </tr>
     <tr>
-      <td>O2–<br>27b<br>(Ours)</td>
+      <td>O2-27B-Preview<br>(Ours)</td>
       <td align="center"><strong>73.40</strong></td>
       <td align="center"><strong>90.40</strong></td>
       <td align="center"><strong>93.33</strong></td>
@@ -235,19 +235,6 @@ O2-9B-Preview can generate incorrect content, malformed tool calls, insecure cod
 
 The model has not been established as suitable for autonomous use in high-stakes domains such as medical, legal, financial, or safety-critical decision-making.
 
-## Citation
-
-If you use O2-9B-Preview, please cite the accompanying technical report:
-
-```bibtex
-@techreport{tong2026oracle,
-  title       = {Oracle: A Verifier-Scaling Paradigm for Runtime Self-Learning Agents},
-  author      = {Tong, Yongqi and Tan Li Hui, Faith and Choy Zhen Wen, Marcus and others},
-  institution = {Ant International},
-  year        = {2026}
-}
-```
-
 ## Acknowledgements
 
 O2-9B-Preview is initialized from Qwen3.5-9B. We thank the teams who built and operated the office, customer-service, software-development, verification, and evaluation runtimes that made this model possible.