Discover forecasting papers
SkillMonitoring & opsDiscover, deduplicate, rank, and optionally dispatch review tasks for new time-series forecasting papers. Use for arXiv or Hugging Face paper scans, recurring literature monitoring, and candidate-model intake; not for claiming or completing a paper reproduction.
Available today. Use it from your connected AI after setup.
No other account needed.
Connect ahel once, and every AI you use reads what you have installed.
Then ask your AI: use the Discover forecasting papers skill
What this skill tells your AI
The instructions your AI receives, as published by diaugeia/moderntsf in .agents/skills/discover-papers/SKILL.md and read by ahel’s review.
This skill is a research-intake harness. It finds candidates and creates bounded, reviewable work items; it does not treat search relevance as permission or evidence to add an implementation.
Read references/intake.md before scanning or dispatching.
Scan
- Run
uv run tsf model list --jsonand use model-card paper URLs, titles, and public names as the deduplication baseline. - Search both arXiv and Hugging Face Papers. Cover the query lattice in the reference instead of relying on one broad phrase. Prefer source metadata and primary paper/project pages over search snippets.
- Normalize arXiv identifiers and titles, collapse cross-source duplicates, and reject papers that do not actually forecast future time-series values.
- Rank candidates by task relevance, novelty relative to the flat catalog, authoritative code availability, license clarity, recency, and implementability.
- Return a concise candidate brief for each retained paper. Clearly distinguish facts from inference; discovery does not claim implementation or verification.
Dispatch
Dispatch only when the user or the recurring-task prompt explicitly requests it.
Create at most three independent tasks in one run, one paper per task. Include the
candidate brief, primary URLs, deduplication result, expected deliverable, and the
instruction to use add-model only after paper, source, license, and runtime inputs
are resolved. Do not ask a dispatched task to merge, publish, or modify external
systems.
Without dispatch authorization, report the ranked queue in the current task. If no candidate clears the threshold, report that the scan completed with no dispatch; unchanged state is a successful monitoring result.
Integration boundary
The downstream task owns implementation. It must preserve the flat
src/models/<lowercase_module_slug>/ layout, use shared components only when
semantics match, and pass the repository's provenance and contract gates. Search
results alone never establish a local implementation or verification result.
Signals
- GitHub stars
- 65
- Forks
- 8
- Last commit
- Sep 2026
Advanced
- Catalog kind
- skill
- Gateway key
discover-papers- Source
- github.com/diaugeia/moderntsf