Which Prompt Line Changed the Behavior?

A recruiting agent screens the same candidate differently across two versions. Attribution traces both decisions to the exact line where the experience threshold changed.

👤

The Scenario

A recruiting platform screens candidates for Senior Engineer roles. The experience threshold changes between prompt versions.

Candidate "Sarah Chen — 4 years Python/Django, applying for Senior Backend Engineer"
v1.0.0
Reject candidate
Rule: require 5+ years for Senior
vs
v1.1.0
Advance to interview
Rule: require 3+ years for Senior
v1.0.0 Require 5+ years v1.1.0 Require 3+ years
1.0x
🔍

The Key Moment: 5 Years → 3 Years

Same candidate (Sarah Chen, 4 years experience), two different outcomes. Attribution traces both decisions to line 12 of the system prompt, where the experience threshold changed from 5 years to 3 years.

v1.0.0 rejects her (4 < 5). v1.1.0 advances her (4 ≥ 3). One line, one number, one completely different hiring decision.

What the Demo Shows

Act 1 — Versions

Line-Level Tracing

Attribution points to the specific line in the prompt that most likely influenced the agent's decision. Not just "the prompt changed" — exactly which line matched the action.

Act 2 — Compare

Side-by-Side Comparison

Same input, two prompt versions, two different actions. The comparison table makes behavioral divergence immediately visible — where did the agent start doing something different?

Act 3 — Analyze

Confidence Scoring

Each influence is scored 0.0–1.0 and labeled HIGH (≥ 0.80), MEDIUM (≥ 0.50), or LOW (< 0.50). Primary and supporting influences ranked by confidence.

Act 4 — Disclaimer

Honest About Limitations

Attribution is pattern matching, not causal proof. It points you to the right section of the prompt — it does NOT prove that a specific line caused the behavior.

Try It Yourself

Attribution runs locally with no API keys. Point it at a bundle and trace data.

pip install llmhq-promptops llmhq-releaseops
releaseops attribution explain \
  trace.json --bundle my-agent@1.0.0