Sync Blog.md with post-hackathon training section, Haiku 4.5 n=68 baseline, and corrected references

Sync to local main (commit 1e292cf): post-hackathon blog updates

Add external reference links throughout Viveka blog

Improve blog copy with small fixes

Polish Viveka blog with references and minor fixes

Add research-context references to Viveka blog

- Correct scenario count in Section 2 to 68 across four tiers

docs(blog): expand Viveka overview and training insights

docs(blog): add design-principle framing and frontier baselines

docs(blog): polish prose, add mentor credit, and tighten section voice

docs(blog): refactor Blog.md and add deeper engineering insights

docs: add judge TL;DR table at top of README and Blog

chore(post-eval polish): blog architecture section + Gradio root mount

feat(submission): trained Llama-3B sealed eval + leaderboard PNG + Gradio fix

chore(submission): final submission marker — Viveka v1.0.0

Merge branch 'main' of https://github.com/gowtham-sai-yadav/viveka-env

feat(eval): add baseline evaluation files for Claude and GPT models

docs(readme): add YouTube channel to Submission Links

feat(submission): HF Space deployed + Blog.md + final shims + URLs wired

Merge branch 'main' of https://github.com/gowtham-sai-yadav/viveka-env

feat(inference, safety): introduce AnthropicClaudePolicy and enhance safety concerns handling

feat(submission): canonical OpenEnv layout + loss curves + Llama-3B baseline

docs(readme): audit fixes + Llama 1B sealed eval + 3-architecture results

feat(inference, train, eval): enhance user prompt structure with memory orchestration fields

fix(train): preserve exact backwards compatibility when --resume is not passed

feat(train): add --resume flag to continue from latest checkpoint

fix(graders): replace multiplicative soft-penalty stack with weighted-average composition

fix(env): pass services_state in _make_observation so metadata reward_signals are correct

docs(readme): fill Qwen baseline mean_reward (0.042) and T4 safety SR (0%) from completed eval

docs(readme): update for two-architecture run + add EOS-bug-discovery section

feat(eval): add xkcd / hand-drawn style option to reward curve plot

feat(eval): combined two-architecture reward curve plot (rubric headline)

fix(train): pass eos_token_id list via GRPOConfig.generation_kwargs (Qwen GRPO termination)

fix(train): pin chat-end token as EOS so Qwen generation stops on turn end

fix(prompts,train): kill multi-step example leakage that made Qwen ramble past EOS

fix(scenarios): correct post_state.balance that contradicts empty ground_truth ops

fix(env): pass services_state to intermediate reward signals

fix(train): teacher uses real reversibility from registry + better last_result check

fix(inference): use shared SYSTEM_PROMPT + build_user_prompt for FrozenQwenPolicy

fix(prompts): restore op-registry scaffolding — v3 reward stuck at -1.0

Merge branch 'main' of https://github.com/gowtham-sai-yadav/viveka-env

feat(docs): add scenario provenance documentation

feat(data): add new datasets for banking, digilocker, irctc, telecom, and upi

feat(scenarios): expand service options and update train details

feat(inference): Best-of-N candidate sampling for qwen policy

refactor(inference, train): unify system prompts and enhance user prompt structure

fix(train): revert num_generations to 4 — G=16 was 30hr ETA on T4

fix(train): bump grad_accum to 16 to match num_generations=16

feat(train): teacher-rollout multi-step trajectories + multi-step examples in prompt

fix(inference): wire --model/--adapter through to FrozenQwenPolicy + add LoRA support