Jev Forge

One question expands into one row per candidate sharing state and question plus a candidate path; one scorer rates every row and only that question's logits are normalized. That keeps candidate scores comparable without language-model decoding.

The first proving ground is page interaction: next click, navigation target, form control, route, or escalation from currently available elements. General routing, tool selection, and triage are roadmap items, not done.

A Mac demo script fetches the 0.8B checkpoint, selects MPS, starts a local Jev-compatible API, and opens a live page showing latency and the full candidate distribution before clicking. Serving otherwise takes a checkpoint directory and port.

The 0.8B and 0.6B models share data and protocol with Jev-1.13 as a separate reference arm, and the trained arms beat the raw backbone while trading blows with the reference across choice, yes/no, and score metrics. The raw-backbone row is a zero-shot slice, not a full arm.

On the fixed website-disjoint splits, JevForge 0.8B reaches 0.579 test choice top-1 and 0.637 OOD, against the raw Qwen3.5-0.8B at 0.235 and 0.340, with Jev-1.13 as a separate reference arm at 0.543 and 0.610. For noul, the 0.8B reports test 0.826 accuracy and 0.128 Brier; score MAE is 0.367 test and 0.390 OOD. The raw-backbone row is a zero-shot reference over the first 200 records of each split, while the trained and reference arms use their recorded full evaluation sets.

A preliminary calibrated-decision reinforcement learning path for browser choices, combining grouped action sampling, browser-task utility, a Brier calibration objective and KL anchoring to the supervised checkpoint. On full splits it lifted test choice top-1 from 0.5787 to 0.5863, OOD from 0.6373 to 0.6477, and choice mass on valid actions from 0.4491 to 0.4881 on test and 0.4790 to 0.5328 OOD. It can add gold-preserving hard-negative views without changing the frozen dev, test or OOD records.

The API quick start creates a venv, installs requirements, then runs python -m jevforge.serve with a checkpoint directory and port, exposing a Jev-compatible /v1/systemone that accepts state plus choice, noul or score questions. On an Apple Silicon Mac, scripts/run_mac_demo.sh downloads the 0.8B checkpoint, selects MPS, starts the local API and opens a live page. The full recipe, scripts/qwen35_08b_pipeline.sh, builds website-disjoint records from Mind2Web, fine-tunes with cross-entropy and Brier objectives, calibrates and evaluates.