Hackathon submission: new README (3-5 min read), BLOG.md narrative, frontier baselines, design-principles framing

Phase J.7: add /curriculum_progress endpoint + GRPO polling for mastery telemetry

Phase D: pipeline_environment.py surgery — cut adversarial designer, handoff tracker, handoff/specialization rewards

Phase C: cut handoff_quality_reward + role_specialization_bonus, lower STEP_REWARD_MAX to 0.32

Phase A: cut adversarial_designer, ollama_client, judge_client, handoff_metrics

Phase 6.5 fix (4/5): reward magnitude tune-up (terminal bonuses/penalties)

Phase 6.5 fix (3/5): Groq LLM judge client with Ollama fallback

Pre-Phase 6: retry on Ollama client + mark live tests requires_ollama

Phase 5.7: wire adversarial scenarios into engine state

Phase 5.6: curriculum tracking

Phase 5.5: reward additions

Phase 5.4: handoff scoring + role routing

Phase 5.3: step validation

Phase 5.2: update reset

Phase 5.1: wire init

Phase 4: handoff metrics

Phase 3 cleanup: enforce step budget in designer._parse

Phase 3: ollama client + adversarial designer

Phase 2: curriculum controller

Phase 1: role system

Round 2: rename Python package to devops_pipeline_gym

Fix: all score paths return strictly (0,1) — never 0.0 or 1.0

Fix: clamp grader scores to (0.001, 0.999) — strict 0<score<1

Fix health visibility leak + recovery alert bug

Fix grader weights, harden int() casts, fix partial obs leak, add difficulty + exploit docs

Harden random_incident grader, fix /grader default, remove prescriptive logs from hard tasks, add recovery status to obs, stochastic docs

Final: config recovery delay, expand proc gen (5 types + compound), partial obs fix, reward cap +0.30, MDP docs, seed curriculum

Fix _failing_service bug, add shared_buffers hint, fix partial obs leak, add MDP description

Final fixes: deploy-time grader check, anti-spam penalty, public attrs, seed at reset, remove openai from reqs, baseline recalibration

Round 2 judge fixes: reward pipeline, sub-goals, exploration decay, grader depth

Fix all 3-judge review findings: rewards, graders, engine ordering, spec compliance

Add procedural scenario generation: random_incident task (Task 6)

Skip compounding/tipping for clean_deploy, remove dead import, fix health endpoint

Make clean_deploy truly easy: skip transient staging failure, reduce deploy spikes

Fix task selection: accept task in reset() body, not just env var

Make staging→prod flow obvious, differentiate baseline scores

Fix capacity_crisis grader: optimal now beats random, recalibrate scoring

Add observation summary field, rewrite inference system prompt, sorted JSON keys

Add capacity_crisis task (5th task) — prevent collapse under 4x traffic

Add non-linear tipping points for emergent behavior

Add auth-service to all 4 scenarios with updated dependency graph

Add database-primary service to all 4 scenarios with dependency graph, fix graders to count only target services

Reward shaping audit: add repeat-action penalty, reward bounds [-0.35, +0.20], repeated investigation penalty

Fix 5 evaluator issues: outcome-based grader, reduced warmup spikes, app.state for grader, remove openai from server deps, tune time pressure

Add trade-off effects, cross-metric compounding, recovery cascade, 3-path grader, non-linear deploys

DevOps Pipeline OpenEnv Environment - Full submission