Alexander Ammann@agammann

Source Trail makes web answers inspectable. Ask a natural-language question, supply starting URLs and optionally a software version, then review the answer beside exact quotes, source links, detected versions and a navigation trace. The Python prototype reads static HTML and plain text within a six-page budget. It ranks passages, preserves table context and checks every quoted passage against retrieved text. Verified mode adds version and current/stale-source signals; the baseline shares traversal, prompt limits and quote checks. Missing evidence can produce an abstention, while unresolved conflicting evidence can produce a conflict result. A local browser inspector and JSON export make the evidence easy to review. We validated Source Trail on an AMD Instinct MI300X virtual function through AMD's general One-Click Tutorial service, using local vLLM and pinned Qwen3-4B-Instruct-2507 weights. All 44 recorded model requests completed. Provider usage and server counters agree at 58,604 input and 2,335 output tokens, with active GPU samples overlapping every request. A fresh Python TaskGroup query returned an exact quote retaining the asyncio.CancelledError exception qualifier. On 16 fictional development cases, baseline scored 12/16 and verified scored 11/16. Both passed all six synthetic stress-development cases, and 114 unit and loopback tests passed. All failures remain available in the public repository; the fixtures are not held out or an official score. Quote presence checks text integrity, not answer entailment. Official Match 4 interface, evaluator and entry-container validation remain pending. The MIT-licensed source, runtime records and retained results let reviewers inspect this particular AMD development run. A separate online extractive demo uses Qwen3 0.6B to select one source passage verbatim within a three-page budget; it does not replay AMD responses.

Six Matches. One leaderboard. Build on AMD, earn points and compete for prizes.