Add FinCalc-NLI eval (financial arithmetic) + results for all checkpoints (#4)

Add image Decisions serving and explicit option probabilities

Remove qwen3.5-0.8b-nli-v5 checkpoint

add qwen3.5-4b-nli-v5-nvfp4 (ModelOpt NVFP4 PTQ + QAD) with SGLang bench

card: links to the 4B / 2B / 0.8B v5 checkpoints

add qwen3.5-0.8b-nli-v5

add qwen3.5-2b-nli-v5

card: openjev-4B v5 (JevBench 0.814 public, judging/faithfulness/IF tables, determinism, injection limits, contamination manifest)

add contamination manifest (train+test splits used for v5 panel)

add RESULTS-v5.md

code: gen_ifeval.py

code: think_probe.py

code: eval.py

code: train.py

code: distill_teacher.py

code: hard_gen.py

code: data_mix.py

code: webql_null.py

code: eval_agentic.py

code: eval_extra.py

code: eval_security.py

code: order_test.py

code: eval_jevbench.py

code: modeling_openjev.py

code: openjev_decide.py

4B v5: typed-decision cross-encoder (panel + gpt-5.5 hard items + teacher distillation)

SGLang serving: external model package, launcher, client, benchmark; card section

Document the 0.8B v2s-long checkpoint served by the Space

upload openjev-0.8B v2s-long (stage3, 4k ctx, faith/ifollow/false-premise + long-doc)

Share Qwen3.5 prefix cache across NLI hypotheses (#1)

upload qwen3.5-35b-a3b-nli

upload qwen3.5-4b-nli-v2

openjev v2: results, videos (Doom v2, Minecraft), code, card

add results/webql_geminiprompt_4b.json

add results/webql_bench_4b.json

add results/webql_scaffold_4b.json

add results/webql_mlp_4b.json

add results/webql_fulldoc_35b.json

add results/webql_mlp_35b.json

add results/webql_bench_35b.json

add results/webql_fulldoc_4b.json

add results/webql_geminiprompt_35b.json

add results/latent_mlp_35b.json

add results/qwen35b_all.json

add assets/radar_openjev_35b.png

add code/webql_gemini_prompt.py

add code/webql_fulldoc.py

add code/webql_mlp.py

add code/webql_scaffold.py

add code/webql_bench.py