prompt-regression

Prompt Regression

Compare two prompts head-to-head and determine, with statistical rigor, whether the candidate is better, worse, or tied on each evaluation dimension.

When to Activate

• You changed the system prompt and want to verify it's actually better
• You're iterating on RAG answer templates
• You're optimizing agent step-by-step instructions
• You want data to support a prompt change decision

Checklist

You MUST create a task for each item and complete them in order:

1. Load and analyze prompts — diff the baseline vs candidate
2. Derive comparison dimensions — from the prompt changes + task type
3. Select graders per dimension — pairwise, judge, or rule
4. Run position-debiased comparison — swap-aggregate to eliminate order bias
5. Compute statistics — win rates + bootstrap 95% CI per dimension
6. Present results — per-dimension verdict with confidence intervals

Fast path: run the bundled script

Don't hand-write the win-rate + bootstrap math (the swap-aggregation and CI are easy to get wrong). Run the bundled, tested script (scripts/pairwise.py, standard library only, no OpenJudge dependency):

python scripts/pairwise.py --comparisons comparisons.jsonl --candidate candidate --baseline baseline

Each comparison row: {"id","model_a","model_b","score","dimension"?} where score >= 0.5 means model_a won. Emit two rows per query with A/B swapped to debias position. The script reports per-dimension candidate/baseline/tie rates, bootstrap 95% CI, and a verdict (BETTER / WORSE / TIED / INSUFFICIENT_EVIDENCE / INCONCLUSIVE; exit 0 only if better). --self-test to verify it.

Steps below explain how to derive dimensions and produce the comparisons (with OpenJudge or any judge); the inline snippets are the reference behind the script.

Step 1: Load and Analyze Prompts

Read the baseline and candidate prompts. Identify:

• Task type: chatbot / RAG generation / code review / translation / summarization / agent instruction / other
• What changed: added constraints, changed tone, new examples, different output format, expanded/shortened instructions
• Intent of change: what problem was the user trying to fix?

Step 2: Derive Comparison Dimensions

Based on the task type and what changed, derive 3-5 comparison dimensions.

Dimension templates by task type

Chatbot / Conversational:

• Answer relevance — does it address the user's question?
• Tone appropriateness — does the tone match context?
• Factual accuracy — no fabricated information
• Conciseness — doesn't ramble or over-explain
• Instruction following — obeys system prompt constraints

RAG Generation:

• Faithfulness — grounded in retrieved documents
• Citation accuracy — correctly references sources
• Completeness — covers all aspects of the query
• No hallucination — no claims beyond documents

Code Review / Generation:

• Bug detection — finds real issues
• False positive rate — doesn't flag correct code
• Actionability — suggestions are specific and implementable
• Code style — follows conventions

Agent Instructions:

• Tool selection — picks the right tool
• Step efficiency — minimal steps to goal
• Error recovery — handles failures gracefully
• Output format — follows specified structure

Each dimension gets:

• An id (slug)
• A one-sentence description
• A grader type: pairwise or judge or rule

Step 3: Select Graders

Decision priority:

1. Can a rule check this? → FunctionGrader or StringMatchGrader. Free, deterministic. Example: output length, keyword presence, JSON validity.
2. Is there a reference answer? → pairwise against reference.
3. Subjective quality, no reference? → pairwise A/B comparison.
4. Single-output judgment needed? → judge (binary pass/fail per output).

Step 4: Run Position-Debiased Comparison

Pairwise comparison with swap-aggregate

LLM judges have position bias — the first response shown wins 5-15% more often. Swap-aggregate eliminates this: run each comparison twice with swapped positions, keep only consistent wins:

from openjudge.graders.llm_grader import LLMGrader
from openjudge.graders.schema import GraderMode
from openjudge.runner.grading_runner import GradingRunner
from openjudge.analyzer.pairwise_analyzer import PairwiseAnalyzer

# Judge prompt for relevance comparison
relevance_judge = LLMGrader(
    model=model,
    name="relevance_compare",
    mode=GraderMode.POINTWISE,
    template="""
Compare Response A and Response B for the query below.
Which response better addresses the user's question?

Query: {query}
Response A: {response_a}
Response B: {response_b}

Score 1.0 if A is better, 0.0 if B is better, 0.5 if tied.
Respond in JSON: {{"score": <float>, "reason": "<explanation>"}}
""",
)

# Build pairwise dataset with position swap
dataset = []
for sample in test_samples:
    # Original order
    dataset.append({
        "query": sample["query"],
        "response_a": baseline_outputs[sample["id"]],
        "response_b": candidate_outputs[sample["id"]],
        "metadata": {"model_a": "baseline", "model_b": "candidate"},
    })
    # Swapped order — critical for debiasing
    dataset.append({
        "query": sample["query"],
        "response_a": candidate_outputs[sample["id"]],
        "response_b": baseline_outputs[sample["id"]],
        "metadata": {"model_a": "candidate", "model_b": "baseline"},
    })

runner = GradingRunner(
    grader_configs={"relevance": relevance_judge},
    max_concurrency=8,
)
results = await runner.arun(dataset)

# Analyze with PairwiseAnalyzer
analyzer = PairwiseAnalyzer(model_names=["baseline", "candidate"])
analysis = analyzer.analyze(dataset, results["relevance"])

print(f"Win rates: {analysis.win_rates}")
# → {'baseline': 0.35, 'candidate': 0.55} → candidate wins 55% of comparisons
print(f"Best model: {analysis.best_model}")

Why swap-aggregate? Without it, if the judge prefers the first response shown, and you always show baseline first, you'll systematically underrate the candidate.

Step 5: Compute Statistics

For each dimension, report:

• Candidate win rate, baseline win rate, tie rate
• Bootstrap 95% confidence interval
• Verdict: better / worse / tied / inconclusive

PairwiseAnalyzer.analyze interprets each comparison as score >= 0.5 → model_a wins, using the row's metadata.model_a / metadata.model_b. So derive a per-comparison winner list from dataset + results, then bootstrap over that list — never index the PairwiseAnalysisResult object (it has no per-sample rows).

import numpy as np
from openjudge.graders.schema import GraderScore

def per_comparison_winners(dataset, grader_results):
    """One named winner per comparison row (handles swapped order via metadata)."""
    winners = []
    for sample, result in zip(dataset, grader_results):
        if not isinstance(result, GraderScore):
            continue  # skip errors
        meta = sample.get("metadata", {})
        winners.append(meta["model_a"] if result.score >= 0.5 else meta["model_b"])
    return winners

def bootstrap_win_rate(winners, target, n_iter=1000):
    n = len(winners)
    rates = []
    for _ in range(n_iter):
        idx = np.random.choice(n, n, replace=True)
        rates.append(sum(1 for i in idx if winners[i] == target) / n)
    return float(np.percentile(rates, 2.5)), float(np.percentile(rates, 97.5))

winners = per_comparison_winners(dataset, results["relevance"])
n = len(winners)
candidate_rate = sum(1 for w in winners if w == "candidate") / n
baseline_rate = sum(1 for w in winners if w == "baseline") / n
ci_low, ci_high = bootstrap_win_rate(winners, target="candidate")

if ci_low > 0.5:
    verdict = "candidate BETTER"
elif ci_high < 0.5:
    verdict = "candidate WORSE"
elif (ci_high - ci_low) < 0.3:
    verdict = "TIED (CI brackets 0.5, narrow)"
else:
    verdict = "INCONCLUSIVE (CI too wide — need more samples)"

print({"candidate_win_rate": candidate_rate, "baseline_win_rate": baseline_rate,
       "ci_95": [ci_low, ci_high], "verdict": verdict})

Note: with swap-aggregate each query produces 2 comparison rows. Bootstrapping over rows (above) is the simple approach; for a tighter estimate, bootstrap over queries and average the 2 swapped rows per query so position pairs stay together.

Step 6: Present Results

Prompt Regression: v1 (baseline) vs v2 (candidate)
Task: Customer support chatbot
Samples: 50

Dimension            Candidate  Baseline  Tie   95% CI         Verdict
===========================================================================
Answer relevance        58%       32%      10%   [51%, 65%]   ✓ BETTER
Factual accuracy        48%       44%       8%   [41%, 55%]   = TIED
Tone appropriateness    38%       52%      10%   [31%, 45%]   ✗ WORSE
Conciseness             62%       28%      10%   [55%, 69%]   ✓ BETTER

Summary: v2 is significantly better on relevance and conciseness,
but worse on tone appropriateness. The tone regression likely comes
from the new "be direct" instruction — consider softening it.

Top 3 tone failures (candidate worse):
  1. Query: "I'm really frustrated..." → v2 response too curt
  2. Query: "This is my first time..." → v2 missing empathetic opening
  3. Query: "Can you help me understand..." → v2 skipped explanation

Common Mistakes

• Not doing position swap. Position bias in LLM judges is 5-15%. Without swap-aggregate, results are systematically skewed.
• Comparing with < 10 samples. Bootstrap CI at n=10 is ±15%+ half-width. At n=5 it's ±25%+. Results are noise, not signal. Minimum 10, prefer 30+.
• Single "overall" comparison without dimensions. "V2 is 55% better" hides that it's +20% on relevance but -15% on tone. Always report per-dimension.
• Accepting ties as "no difference." A true tie and insufficient data look identical without CI. Always report confidence intervals.
• Not pinning model versions. If baseline and candidate are run on different model versions (even same model, different date), model drift contaminates the prompt comparison. Same model, same version, same temperature.

Next Skills

After 06-prompt-regression:

• 03-align-human: Calibrate the pairwise judge against human preferences.
• 02-metric-design: Turn validated dimensions into permanent graders.
• 04-eval-report: Include prompt comparison results in a comprehensive report.