Self-contained in-browser demo: 350M GGUF loaded client-side via wllama (no server in the inference path), user types an instruction against a real page snapshot, model streams {type,index,value}, chosen element highlighted. The "runs in YOUR browser" proof, watchable and reproducible — sits next to the offline 91.2% strict action accuracy (Mind2Web held-out) as live supporting evidence.
Plan: .agents/plans/2026-07-16-live-proof-of-work.md (Track B)
Deliverables:
runtime-bench/demo.html — interactive page (sibling to harness.html), loads GGUF via wllama, uses wllama.createChatCompletion so llama.cpp applies the GGUF's embedded chat template, renders returned action + highlights chosen element/bounds.- Verification prerequisite: confirm exported GGUF carries
tokenizer.chat_template metadata and its templated output matches training/eval.py::_build_prompt byte-for-byte for a known record; fall back to inlining the LFM2.5 template in JS if not.
- Curated seed snapshots/instructions (reuse
data/samples/*.jsonl) so the demo works on open with zero setup.
README.md — "Try it live (in your browser)" section with recorded GIF + one-command serve.py invocation.
Success = a seeded task grounds to the same action the offline eval expects, and a fresh instruction against a pasted snapshot highlights the correct element, recorded as a GIF for the README.
Self-contained in-browser demo: 350M GGUF loaded client-side via wllama (no server in the inference path), user types an instruction against a real page snapshot, model streams
{type,index,value}, chosen element highlighted. The "runs in YOUR browser" proof, watchable and reproducible — sits next to the offline 91.2% strict action accuracy (Mind2Web held-out) as live supporting evidence.Plan:
.agents/plans/2026-07-16-live-proof-of-work.md(Track B)Deliverables:
runtime-bench/demo.html— interactive page (sibling toharness.html), loads GGUF via wllama, useswllama.createChatCompletionso llama.cpp applies the GGUF's embedded chat template, renders returned action + highlights chosen element/bounds.tokenizer.chat_templatemetadata and its templated output matchestraining/eval.py::_build_promptbyte-for-byte for a known record; fall back to inlining the LFM2.5 template in JS if not.data/samples/*.jsonl) so the demo works on open with zero setup.README.md— "Try it live (in your browser)" section with recorded GIF + one-commandserve.pyinvocation.Success = a seeded task grounds to the same action the offline eval expects, and a fresh instruction against a pasted snapshot highlights the correct element, recorded as a GIF for the README.