Sarim-Hash/web-coevo-stage1-handoff

Web-CoEvo Stage 1 handoff

This bundle contains three completed LoRA adapters and the code and saved task pool needed to continue Stage 1 on another GPU machine.

Adapter directory	Saved state
adapters/curr_v1	Curriculum after iteration 1
adapters/exec_v1	Executor after iteration 1
adapters/curr_v2	Curriculum after iteration 2

Start with handoff.md for model downloads, environment setup, GPU placement and the continuation command. The bundle contains multiple adapters; load the desired adapter subdirectory rather than the repository root.

The continuation starts executor iteration 2 from exec_v1, reuses the saved 149-task judged pool, then runs iteration 3. There is no completed exec_v2 in this handoff. The executor uses fresh training rollouts; the 894 trajectories in the saved judged JSON are saved assessment trajectories used to measure task difficulty. New executor checkpoints preserve optimizer state, local random-number state and schedule position after completed schedule batches.

Adapter weights are unchanged from their saved files. The adapter configurations' base-model location was changed from a machine-local path to Qwen/Qwen3.5-4B, and their base revision was pinned to the original commit. These are the only metadata changes. See provenance.json and the checksum manifests for lineage and verification.

The data are generated tasks and synthetic page/action trajectories. Their model-judge labels can be wrong; these artifacts do not establish held-out benchmark performance or real-browser robustness. No credentials, API caches, private manuscript, or machine-specific training logs are included.