# Shortlist Desk

> Set up the review of a conference's call for papers before anyone scores, then rewrite the proposals blind. Paste
> the proposals as CSV or TSV with the slots, tracks, committee and what the call promised. The page picks the
> committee model and its rubric and the anonymization depth, scans every title and abstract for identity, finds
> conflicts, and sets the cut rule, waitlist, decline rung and notification dates. A metered run writes the selection
> plan or anonymizes a batch of proposals. Built on the SkillSafe skill @samber/event-talk-selection
> (samber/dev-event-organizer-skills, MIT). Not affiliated with PyCon, KubeCon or any CFP platform.

https://shortlist-desk.skillsafe.ai/

## What is free
- The pool: the submission-to-slot ratio and acceptance rate, with calibration points (PyCon US 2025 about 15% of 938; KubeCon EU 2024 9% of 2,541; DjangoCon Europe 2024 174 proposals; CHI 2024 26.4%, which does not transfer).
- The committee model (small consensus committee with a non-voting anonymizer, multi-tier cascade, distributed 0-3 pool, or solo curation - deleted under a blind-review promise) and the rubric that follows it (letter advocate scale as best-worst pairs with the one-A floor; four-criteria stars; 0-3 with three reviews), and the reading load against the committee's hours (ten minutes a read, self-set).
- Anonymization depth (named review when not promised; platform hiding by default; a dedicated anonymizer when aid requests ride on the form, after a fairness complaint, or when a small committee would recognize regulars) and a scan of every title and abstract for speaker names, employers, links, handles, emails and gendered pronouns, with highlights and a CSV.
- Conflicts by reviewer employer, organizer submissions (disclose, abstain, cap), sponsor-affiliated submissions (same review, a ceiling, no bought slots at a community event), invited talks in their own lane, duplicate speakers, panels (one-organization and composition gate), likely product talks (lesson-transfer test) and "latest" promises.
- The cut rule, the waitlist (about five per track), the decline rung (criteria recap, or per-proposal feedback for small or first-time-heavy pools), and the notification dates: acceptances and waitlist first with a confirmation deadline, declines after.

## The metered lanes
Two tasks on gpt-terra, metered per run, routed by `task`. Input string fields: `task` (plan or anonymize), `event`, `pool` (ratio, balance, grade summary and one line per proposal with its flags), `facts` (the page's picks and every finding) and `proposals` (the batch to anonymize, or none).
- `plan` replies with `summary`, `committee`, `rubric`, `anonymization`, `conflicts`, `rounds`, `program_fit`, `waitlist`, `notifications`, `decline`, `metrics` and `caveats`; the page holds the model, depth, waitlist and decline rung to its own, requires acceptances before declines and refuses numbers not in the input.
- `anonymize` replies with `items` (id, title, abstract, placeholders, kept) and `note`; the page requires every ID back, rescans each rewrite with the same leak scan, and refuses a summary in place of a rewrite.

## Links
- API tutorial: https://shortlist-desk.skillsafe.ai/api.html
- Source skill: https://skillsafe.ai/skill/@samber/event-talk-selection