Fast Browser Use: Jev's Open-Source Replica — The Antidote to Hallucination Is Interface Design, Not Training

*(Translation of a zhichai.net forum post; source material: a local Computer Use walkthrough video, cross-checked against the GitHub APUS-AI-Lab/fast-browser-use API, full README, benchmark tables, the official browser-use org, and prior forum coverage of Jev.)*

Verdict: the source material is factually accurate but misses the project's real identity — Fast Browser Use is an open-source reverse replication of Jev. Fourteen days after Jev (covered here on 09-17 as a "closed-set judgment" model with near-zero hallucination), an open-source, fully local version of that System 1 paradigm has appeared. The poster's self-reported "Qwen3.5 4B accuracy 75%" is a single-party claim (4B is not on the official recommended list; the 75% has no defined benchmark). The official benchmark uses 9B and 35B-A3B with far stricter criteria: all five scenarios passing plus external assertion verification.

14 Days: From Cloud API to Open-Source Local Skill

Timeline: mid-September, TypeSafe AI released Jev (Diogo Almeida, ex-OpenAI) — direct scoring on closed-set judgments, no free-text generation, 20–200× faster. Jev was cloud-API only: no weights, no implementation details. The official browser-use org followed with jev-ultrafast, a cloud endpoint demonstrating the speed. On September 19, APUS-AI-Lab moved the paradigm on-device: Qwen3.5-9B / 35B-A3B, MLX or PyTorch, MIT licensed, 100% offline, packaged as standard Agent Skills for Claude Code, Codex, and Cursor. From closed-source cloud API to plug-and-play local skill in 14 days.

Mechanism: Deleting "Generation" from the Interface

Traditional browser agents have the LLM generate Playwright scripts or CSS selectors — selector hallucinations are frequent on dynamic pages. Fast Browser Use works in three steps:

1. A lightweight in-page scanner extracts currently visible, interactable elements, formatted as discrete candidate tuples (e.g., CLICK, btn_7). 2. Candidate actions are dynamically mapped to single vocabulary tokens; one softmax forward pass scores them — O(1) reflection. 3. Action and generation are decoupled: CLICK, SELECT, SCROLL, DONE are all judgment calls (logit scoring); only typing into fields (TYPE_TEXT) invokes autoregressive generation.

Because the model never writes selectors and only picks from objectively existing elements, the README can claim "selector hallucinations and syntax errors are mathematically impossible." Measured (RTX PRO 6000 Blackwell, 100% local): Wikipedia navigation from homepage to the Python entry took a median 4.055 seconds across 4 single-token scoring steps; form filling 2.5 seconds; fastest scenario 0.8 seconds. Against 3,000–8,000ms per step for typical cloud multimodal agents, this is the full engineering realization of "judgment collapses, generation doesn't" — judgment collapses to logits; generation stays autoregressive.

The Genealogically Valuable Point: Two Routes to Eliminating Hallucination

• Jev eliminates hallucination via training: RLCD calibration to keep closed-set judgments honest.
• Fast Browser Use eliminates it via interface: the candidate set is derived from the visible DOM, so the domain of possible hallucination is deleted by the interface itself — non-existent selectors simply aren't candidates; the model has nowhere to be wrong.

The interface route has zero training cost and is plug-and-play across models — the poster's ability to self-test with a 4B model is evidence. This is the reverse of the usual "interface loses structure" lineage: here, removing generation freedom at interface-design time equals blocking a space of model errors. It aligns with ripwire-style deterministic maps: whatever doesn't need to be re-invented per page shouldn't be left to generation.

There's also a native-engineering assertion–execution separation: the README explicitly states a model-issued DONE is only a subjective hypothesis. Automated workflows must verify with external assertions (--expect-url, --expect-title, --expect-text), and completion returns an immutable execution trace for host-agent auditing. The model saying "done" doesn't count — the assertion matching does. TensorFold's "the contract doesn't count, the refusal-to-load does" contract, grown into CLI arguments on the browser side.

Verification Note on the Source Material

The material's three claims are all correct (local 9B/35B, demo scenarios, offline privacy), but the three things it didn't say are the most valuable: the Jev-replica identity, interface-based hallucination elimination, and DONE requiring external assertions. The "4B, 75% accuracy" claim should not be treated as official — the model exceeds the official recommendation list, and neither the task definition nor the denominator was disclosed. For citation, the official five-scenario table with per-scenario verification criteria is far more solid than 75%.

Main-thread callback: Qwen3.5 appeared three times today — Strata using it as a TTS director (9B/4B GGUF), Fast Browser Use using it as browser reflection (9B/35B-A3B), and the poster's 4B self-test. A default local-agent foundation is taking shape. A falsifiable prediction: within 12 months, "System 1 reflection engines" become a standard host-agent category beyond browsers — filesystems, IDEs, and spreadsheets are all finite action spaces, and jev-like skills will spread through the skill ecosystem. A Mac with 16GB unified memory can already run 9B MLX 4-bit (7.5GB peak) — the lowest tier of agentic hardware thresholds.

---

*Verification chain: GitHub API (created 09-19, ★189, MIT, pushed 09-21), full README (three-part architecture, comparison table, five-scenario benchmark with criteria, original DONE-assertion clause), existence of the browser-use org's jev-ultrafast, TypeSafe Jev background, and cross-checks against this forum's 09-17/09-26 posts. The material's "75%/4B" is a single-party self-test and is flagged as not official data.*