Phase 13 Stage 1 v2: fix termination signal - reward penalties for non-termination, reduced max_tokens, stop_strings

Phase 13 Stage 1: pivot to winner's pattern - TRL 0.29.0 + fast_inference=False + DAPO loss

Phase 13 Stage 1: fix Unicode encoding artifacts in notebook comments

Phase 13 Stage 1: drop vLLM, use Kaggle native torch 2.10, upgrade trl to 0.18+ (Kaggle dep compatibility)

Phase 13 Stage 1: apply all 7 notebook audit fixes (account labels + health check + hub namespace + explicit commenting)

Phase 13 Stage 1 prep: Kaggle runbook + Run 1 config pre-populated (notebook audit flags pending approval)

docs: correct gpt-oss baseline overclaim — tiered win thresholds, cumulative is diagnostic not trophy

docs: correct overclaim in gpt-oss baseline — honest win thresholds

Phase 13 prep: ollama provider + gpt-oss:120b baseline eval (3 seeds × 6 scenarios)

Phase 13 prep: 5-seed baseline evals + Kaggle Stage 1 kickoff checklist

docs: log Phase 11 M-tier scenarios (cumul=1.988 > E-tier 1.284, discriminability three-tier)

Phase 11 fixup: deploy smoke /tasks count 3 -> 6

Phase 11: medium scenarios M1/M2/M3 + AdaptationRubric multi-drift verification

docs: clean BUILD_LOG phases list + point to TRAINING_LOG; cross-ref Section 1 partial data

docs: log Phase 10 HF Space deploy (production step_shaping=0.1000 verified)

Merge branch 'main' of https://huggingface.co/spaces/yashash045/schemashift

initial commit

Phase 10: HF Space deploy — README frontmatter + DEPLOY.md + deploy smoke tests

docs: log Phase 9 results (discriminability gap 0.348 shaped, env validated)

Phase 9: baseline eval harness (heuristic + LLM agents) + tests

docs: backfill Phase 8 BUILD_LOG entry + update README status

docs: add TRAINING_LOG.md template + training_logs dir for Phase 13 discipline

Phase 8: Env client + GRPO training skeleton (Kaggle notebook)

Phase 7: FastAPI server + Dockerfile + openenv.yaml

Phase 6: SchemaShiftEnvironment scheduler — full episode lifecycle

Phase 5: Composable rubric grader + dense step shaping

Phase 4: CRMAPI with 3 drifts + scenarios E1/E2/E3

Phase 3: CalendarAPI with 2 drifts + DriftInjector

Phase 2: BaseTool + MailAPI with 3 drift handlers

Phase 1: Pydantic models + validation tests

first commit