Share Savings

Share Savings

Turn a measured Understudy savings result into an anonymous, metrics-only report for Understudy. This is for the lower-Anthropic-bill path and future leaderboard receipts.

Safety Gates

• Do not upload prompts, completions, traces, repo names, company names, emails, domains, contact details, source paths, or private notes.
• Do not post without showing the exact JSON payload and API URL, then getting explicit user approval for that submission.
• Prefer .understudy/value/value-report.json with claim_status: claim-supported. If the report says claim-packet-required, call it a scenario lead, not proven savings, and ask before submitting.
• Send only bounded metrics: monthly baseline/candidate/savings, savings percent, request volume, generic providers/models, intervention labels, evidence level, claim status, optional claim hash, and holdout/sample metadata.
• The leaderboard is coming soon. Do not promise ranking, publication, or placement from a submission.

Flow

1. Locate the report:

   test -f .understudy/value/value-report.json && echo .understudy/value/value-report.json
   

   If missing, create one with the value-report workflow before sharing:

   understudy value report --workload-card <path> --route-decision <path> --requests-per-month <n>
   

2. Build a dry-run payload. Add intervention labels that match the real fix:

   node skills/share-savings/scripts/share-savings.mjs \
     --from .understudy/value/value-report.json \
     --intervention prompt-cache \
     --intervention sonnet-gepa \
     --dry-run
   

   Useful intervention labels: prompt-cache, batch, max-tokens, retry-reduction, sonnet-gepa, haiku, openai-route, open-weight-fireworks, local-open-weight.

3. If there is a claim packet, include only its hash:

   node skills/share-savings/scripts/share-savings.mjs \
     --from .understudy/value/value-report.json \
     --claim .understudy/experiments/<id>/claim.json \
     --intervention open-weight-fireworks \
     --dry-run
   

4. Show the exact dry-run payload and endpoint to the user. State whether the report is claim-supported or scenario-only.

5. After explicit approval, submit:

   node skills/share-savings/scripts/share-savings.mjs \
     --from .understudy/value/value-report.json \
     --claim .understudy/experiments/<id>/claim.json \
     --intervention open-weight-fireworks \
     --post --confirm
   

   Override the endpoint only when needed:

   UNDERSTUDY_SAVINGS_API_URL=https://lowermyanthropicbill.com/api/lower-my-anthropic-bill/savings \
   node skills/share-savings/scripts/share-savings.mjs --from .understudy/value/value-report.json --dry-run
   

Output

End with:

• source artifact inspected;
• claim status and evidence level;
• exact interventions shared;
• whether the API call was dry-run or submitted;
• returned anonymous report id, if submitted;
• reminder that leaderboard publication is not live yet.