7. Deliverability Rescue (climb to 95)
Reconstruction for teaching. Fictional newsletter (
The Tradewind Brief), synthetic rubric + drafts; the receipts are generated, not from a real run.
Pattern: metric-climb with an independent grader · Primitive: /goal · Domain: non-coding
Use when
Copy must clear a spam/quality threshold and you want the agent to improve it iteratively — scored by a separate grader so it can't mark its own homework.
The loop (copy-paste)
This is the library card for this example. Copy the contract and fill the brackets:
Goal: Rewrite <asset> until the independent scorer rates it >= <gate>/100.
Context: The draft; a FROZEN rubric; a separate scorer that never sees the writer's reasoning.
Constraints: The writer may not edit the rubric or the scorer (writer != scorer).
Done-when: The separate scorer returns a score >= <gate> on the frozen rubric.
Evidence: A score ledger (one row per pass, with the signal fixed); before/after drafts.
If-blocked: After <K> passes without improvement, stop and surface the stuck signal.
Verify
A separate scorer — not the writer — evaluates the final draft against the frozen rubric and must return a score at or above the gate. The score ledger must show a non-decreasing climb.
Steps
- Score the baseline draft with the independent scorer.
- Fix the lowest-scoring signal; re-score; repeat.
- Stop when the score clears the gate (or stalls).
What happened
The re-engagement email started at a spam score of 41 (before.html: ALL-CAPS subject, image-only body, no unsubscribe) and climbed pass by pass — fixing one signal each time — to 96, clearing the 95 gate (after.html). Red SPAM to green INBOX, with a separate scorer deciding each step. (Illustrative — as of June 2026, verify before relying.)

The receipts
- Score ledger — 41 → 62 → 78 → 88 → 93 → 96, signal fixed each pass.
- Frozen rubric — the independent grader's rules.
- Diptych: before (SPAM 41) vs after (INBOX 96).
- Loop log · cost ledger · all artifacts.
Notes
The discipline is writer ≠ scorer on a frozen rubric: the writer never edits the grader, so the climb is real and not the model talking itself into a higher number.