Reduce inter-judge agreement on funnel analysis Prompt via manual grid search over temperature+system

You are running an automated prompt-optimization campaign using manual grid search over temperature+system to improve inter-judge agreement for a funnel analysis prompt on GPT-4.1.

## Inputs you have
- A current prompt (system + user template) that the team is shipping.
- A train set of 500 labeled funnel analysis examples.
- A held-out dev set of 200 examples (never used by the optimizer).
- A held-out test set of 200 examples (opened exactly once at the end).
- A judge based on regex match checks with a documented rubric.
- A per-example ground truth OR per-example rubric score.

## Your job

### Phase 1 — Baselines
Measure the current prompt on the dev set. Report inter-judge agreement and the guardrail metrics. This is your floor. Any optimizer output that can't beat this floor on the dev set is a failure.

### Phase 2 — Configure manual grid search over temperature+system
Document:
- Search space (what does manual grid search over temperature+system vary? instructions, few-shots, format hints, decoding params?)
- Budget (number of candidate prompts, number of eval runs per candidate, total tokens, total wallclock)
- Stopping rule (e.g., no improvement for N rounds)
- Seed handling
- Safety constraint: any candidate that increases safety-violation rate is auto-rejected even if it boosts inter-judge agreement.

### Phase 3 — Run
Run manual grid search over temperature+system. Log every candidate prompt with its score, cost, and any guardrail violations. Keep the top-K.

### Phase 4 — Human review
Take the top 3 candidate prompts and read them by hand. Reject any that:
- Hard-code dev-set patterns into the prompt (overfitting via instructions).
- Remove or weaken safety language.
- Exploit judge-specific shortcuts (e.g., adding fluff tokens the regex match checks happens to reward).
- Are incomprehensible (even if they score well — unmaintainable prompts are a liability).

### Phase 5 — Final eval
Run the surviving candidate(s) on the never-touched test set. Report:
- inter-judge agreement delta vs. baseline (with CI)
- Guardrail deltas
- Cost & latency deltas
- Qualitative delta vs. baseline on 10 sampled outputs

### Phase 6 — Ship / hold / rethink
Write a one-page decision. Include what you'd do differently next round, what the optimizer revealed about the task, and whether the improvement was instruction-level or few-shot-level.

## Constraints
- Never use the test set during optimization. Never.
- Never let the optimizer edit the safety section of the system prompt.
- Never accept a candidate that the on-call engineer can't explain in one sentence.
- Report negative results. Failed campaigns are still knowledge.

Output the full campaign report in Markdown.