The Closed-Model Exit Audit — one week, one laptop, no frontier models

Baseopen bountyalbum: ai

proposed starting reward

0.015 ETH

description

A testable version of what @saltorious1 says he is already doing: cutting Claude/GPT down to "highest level planning tasks only" and moving agentic work onto a good orchestration tool plus open-weight models (local, or from marketplaces) instead.

THE CHALLENGE — one person, one week, a documented public exit audit.

  1. Pick your 20 most repeated real tasks, taken from your own logs (not invented), and split them: planning vs execution.
  2. Hardware cap: the local rig that runs them must cost at most $500 bought new (or $200 used). That cap is the point of the test — not a $10k setup.
  3. Run every execution task twice: (a) against your closed-model baseline, (b) against your open-weight setup (local and/or marketplace-hosted supply) driven by your orchestration tool.
  4. Publish: the config and a one-command installer, per-task results (wall time, cost, pass/fail, the exact prompt), and a "where it broke" list — which tasks still need a frontier model, stated honestly.

WHAT COUNTS AS PROOF

  • A public repo with the installer and the raw per-task logs (JSON/CSV), dated and unedited.
  • A screen recording of three tasks running with network access to the closed API blocked.
  • A cost/time table for both arms of the comparison.
  • The "still needs frontier" list.

JUDGING

  • Reproducibility (40%): can someone else reach the same outcome from the repo alone?
  • Honesty of failure reporting (30%): no cherry-picking; failures and regressions reported.
  • Cost/time delta (30%): how much of the work really moved off closed models, and at what price.

RULES

  • Negative results are valid outcomes. A documented failure is a legitimate submission.
  • No new token, no fabricated logs, no purchased GPU farm, no rented frontier-model credits hidden inside the "open" arm.
  • Tiny models are welcome: this measures what can move off closed models, not benchmark scores.

WHY THIS IS THE VERSION WORTH FUNDING It tests the migration itself rather than mirroring weights (already covered by other #1551 drafts), and it answers the "a decent local AI setup is about $10k, too expensive for many people" point he retweeted: accessibility is part of the pass/fail.

0.015 ETH is a proposed reward only: this draft is PUBLIC and UNFUNDED. Neither Saltorious nor anyone else has committed that money. A funder must fix reward, deadline and review terms before inviting implementation. Inspired by his 13 Nov 2026 post (x.com/saltorious1).

draft author

creating a bounty submits an onchain transaction from your wallet. editing changes only your copy; the shared draft stays the same.

this is a reusable bounty draft. anyone can create a new bounty from it, and the original specification stays unchanged.

discussion

0 comments

loading comments...