Update README.md: humanise voice, narrative, captions

README.md

@@ -14,30 +14,13 @@ short_description: Play as the on-call engineer in our incident env
 
 # DevOps Pipeline Demo
 
-*Click action buttons. Watch the simulator respond. See what the trained agent has to learn.*
-
-You are the on-call engineer. Five microservices in front of you. Health is masked until you investigate. Each step is role-gated (DEV / SRE / OPS) and the role rotates. Try to clear the incident — or watch yourself break things.
-
-This is the same env our trained Qwen3-1.7B + SFT agent operates in. Every action you take goes through the exact same FastAPI server, hits the exact same deterministic reward grader.
+You are the on-call engineer. Five microservices sit in front of you across 6 tasks, with 9 actions split across 3 roles (DEV, SRE, OPS) that rotate each step. Pick a task, click **Reset**, then click action buttons to clear the incident or watch yourself break things. This is the same env our trained Qwen3-1.7B + SFT agent operates in. The agent scored -0.044 on `judgment_call` against -1.200 for the untrained Qwen2.5-7B baseline (delta +1.156). Every action you take hits the same FastAPI server and the same deterministic reward grader the agent saw during training.
 
 ## Links
 
-- **The env (graded artifact):** [yashash045/devops-pipeline-gym](https://huggingface.co/spaces/yashash045/devops-pipeline-gym)
-- **The trained adapter:** [yashash045/devops-pipeline-gym-sft-adapter](https://huggingface.co/yashash045/devops-pipeline-gym-sft-adapter)
+- **Env (graded artifact):** [yashash045/devops-pipeline-gym](https://huggingface.co/spaces/yashash045/devops-pipeline-gym)
+- **Trained adapter:** [yashash045/devops-pipeline-gym-sft-adapter](https://huggingface.co/yashash045/devops-pipeline-gym-sft-adapter)
 - **Code:** [Yashash4/devops-pipeline-gym](https://github.com/Yashash4/devops-pipeline-gym)
 - **Blog:** [BLOG.md](https://huggingface.co/spaces/yashash045/devops-pipeline-gym/blob/main/BLOG.md)
 
-## How to use
-
-1. Pick a task from the dropdown (start with `clean_deploy` if you want easy mode, `judgment_call` for the hardest one)
-2. Click **Reset** to spin up a fresh incident
-3. Click action buttons grouped by role. Watch the services table + reward chart update live.
-4. Try to reach a clean `approve` (terminal +2.0 reward) before max steps run out.
-
-## What you'll notice
-
-- Acting outside your current role costs `-0.15` and the action is dropped (silently). The Gradio UI lets you click any role's button so you can experience this teaching moment.
-- Some actions look right but are wrong because root cause hides downstream. That's the env teaching `investigate before act`.
-- Reward bookkeeping is live — same reward function the trained agent sees during RL.
-
 Built for the Meta PyTorch OpenEnv Hackathon Grand Finale 2026. Team Tripod (Yashash, Gajanand, Likith).