mattmireles Find Papers
Produce a small, decision-relevant candidate set with primary-source evidence. Discovery output is not a paper corpus and abstracts are not substitutes for full-text evidence.
Scope the question
Read the relevant current belief in README/Wiki/, external research in
README/Guides/, and local evidence in README/Notes/ before launching
research. Write a private brief at
README/Notes/<slug>-find-papers-brief.md containing the decision, technical
context, precise questions, source priorities, exclusions, and a maximum paper
count (default 10; hard maximum 30). Do not include credentials, private audio,
or unreleased benchmark outputs in the brief or prompt.
If README/Papers/ exists, inventory it by title and identifier to avoid
duplicates. If it does not exist, record that the repository has no paper
corpus; do not create one as a side effect of discovery.
For a single paper named by the user, retrieve and verify its primary source directly instead of running broad discovery.
Run the research workflow
From /Users/mm/Documents/GitHub/llm-workflows, run:
pnpm run research:find-papers -- \
--topic "<bounded topic>" \
--context-file /Users/mm/Documents/GitHub/kokoro-coreml/README/Notes/<slug>-find-papers-brief.md \
--target-repo kokoro-coreml \
--target-paper-directory README/Papers \
--max-papers <n> \
--agent-mode max
The workflow uses find_papers_v1. Preserve its generated run directory and
paper-candidates.md; do not hand-copy an unverified list into repository
documentation. A completed run requires nonempty candidate output and completed
metadata. Resume a failed workflow through its checkpoint rather than launching
an unrelated duplicate run.
Validate before recommending
For each retained candidate, verify its canonical identifier, title, authors, year, landing page or official PDF/full HTML, and abstract from a primary source. Merge a preprint and a published version into one record, preferring the published version when it contains the relevant method or results. Record why it matters to the exact TTS/vocoder, Core ML conversion, model architecture, or Apple hardware decision in the brief.
Exclude papers that only share generic keywords, lack a recoverable primary source, or cannot support the decision in the brief. Treat reported benchmark numbers as hypotheses until the relevant model revision, device, OS, runtime, and measurement method have been reproduced locally.
Handoff boundary
Return the verified candidate set and the retained workflow artifact. Do not
create README/Papers/, reconstruct papers from abstracts, or claim full-text
review as a discovery side effect. When the user asks to import a retained
paper, hand off to paper-ingest after primary-source validation.