Summary

• consume Verifiers sft-replay instead of keeping a local rollout replay env
• allow teacherless SFT only for explicit sft-replay taskset datasets
• force SFT to MITO/no renderer and drop the default zero-advantage post-batch filter unless explicitly configured
• backfill token usage for replayed message-only trajectories
• add a debug sft_replay.toml config, focused replay/config tests, and docs/skill guidance for replay-backed SFT

Verifiers dependency

deps/verifiers points at 14be5ee, the current head of PrimeIntellect-ai/verifiers#1536.

Validation

• uv run --no-project --with ruff ruff check ...
• inline config validation for teacherless SFT guardrails and configs/debug/training_modes/sft_replay.toml parsing
• inline sft-replay rollout -> token backfill -> interleave smoke
• git diff --check HEAD~2..HEAD for the docs/skill and submodule-pin sync commits
• Verifiers dependency: uv run pytest tests/test_v1_replay_harness.py tests/test_v1_config_extension.py -q