Investigating Anomalous Results Agent Skill

Investigating Anomalous Results

Overview

Random tweaks waste time and manufacture false findings. Quietly dropping the inconvenient data point, nudging the cutoff, or re-running until it "works" doesn't fix the problem — it fabricates a result.

Core principle: ALWAYS find the root cause before adjusting anything. An adjustment made before you understand the cause is, at best, noise and, at worst, fraud.

Violating the letter of this process is violating the spirit of it.

The first question for any anomaly: is this a code bug, a data issue, or a real finding? You cannot answer by guessing, and you must not "fix" it until you know — because a real finding is not something to fix.

The Iron Law

NO ADJUSTMENTS WITHOUT ROOT-CAUSE INVESTIGATION FIRST

No dropping data, changing a test, transforming a variable, re-running with new parameters, or "cleaning" until Phase 1 is complete. If you haven't found the cause, you cannot justify the adjustment.

When to Use

• A result is surprising, too good, or contradicts what's known
• An impossible value (negative count, probability > 1, age of 500)
• A pipeline errors or a model won't converge
• A result contradicts a sanity check or a simple cross-tab
• A replication or robustness check disagrees with the primary result
• Performance that's suspiciously high (often leakage)

Use this ESPECIALLY when:

• The anomaly is *convenient* (it supports your hypothesis) — that's when the urge to skip investigation is strongest
• Under time pressure
• You've already tried one quick adjustment

The Four Phases

Complete each before the next.

Phase 1: Characterize the Anomaly

Before any adjustment:

• Read the output/error carefully. The exact value, the exact message, the stack trace. Note shapes, counts, and where in the pipeline it appears.
• Reproduce it deterministically. Fixed seed, same data version. Does it happen every time? If not reproducible, gather more data — don't guess.
• Check recent changes. What changed in the code, the data version, or the environment since it last looked right? git diff, data provenance, package versions.
• Gather evidence at each pipeline stage. Instrument raw → cleaned → derived → result. Print shapes, ranges, null counts, and key summaries at each boundary. This reveals *where* the anomaly enters.

raw:     N=10342, outcome mean=0.31, nulls=0
cleaned: N=10298, outcome mean=0.31, nulls=0      <- 44 dropped, expected
derived: N=10298, feature mean=4e7                <- WRONG: unit blew up here
result:  coefficient enormous                     <- symptom; cause is upstream

• Classify the cause with the evidence: code bug, data issue, or real finding. Do not proceed assuming one without evidence.

Phase 2: Pattern Analysis

• Find a working comparison. A subset, variable, site, or time period that behaves as expected. What's different about the anomalous one?
• Compare against a reference. A known correct result, a textbook expectation, a prior estimate from the survey. Read the reference method completely — don't skim.
• List every difference between working and anomalous, however small. Don't assume "that can't matter."

Phase 3: Hypothesis and Test

• Form a single hypothesis: "The anomaly is caused by X because Y." Write it down. Be specific.
• Test minimally on the smallest slice that exhibits it. One variable at a time.
• Verify before continuing. Confirmed → Phase 4. Not confirmed → new hypothesis. Don't stack adjustments.
• When you don't know, say so. "I don't understand why the derived feature inflates" — then investigate or ask. Don't pretend.

Phase 4: Resolution

Resolve according to the cause you established:

• Code bug: Write a minimal failing check that reproduces it, fix the code at the source (not a downstream patch), re-run the affected steps, confirm the anomaly is gone and other results unchanged.
• Data issue: Handle it per a documented rule. If the rule was not pre-registered, the decision is a deviation — record it, and the affected analysis becomes exploratory. Dropping observations is legitimate ONLY when root cause shows they are invalid measurements (e.g., a sensor error), and that justification is documented. Never drop data because it's inconvenient.
• Real finding: It is a result, not a defect. Do NOT "fix" it. If it was not what you pre-registered, it is an exploratory finding — report it as such, and pre-register a fresh confirmatory test (ideally on new/held-out data) before claiming it.

If 3+ adjustments fail: STOP. The pipeline or the design may be wrong — repeated failures that each reveal a new problem indicate a structural issue, not a series of small bugs. Discuss the design with your human partner before another attempt. Re-opening the design or pre-registration must be documented as a deviation.

Red Flags - STOP and Investigate

• "Let me just drop that outlier and re-run"
• "I'll nudge the threshold until it's significant"
• "Probably a fluke, let me re-run with a different seed" (changing the seed to change the result)
• "The result I want appeared, no need to dig into why the first run didn't"
• "Quick adjustment now, understand it later"
• Proposing an adjustment before tracing where the anomaly enters the pipeline
• An adjustment that happens to make the result support your hypothesis

All of these mean: STOP. Return to Phase 1.

Your Human Partner's Signals You're Doing It Wrong

• "Is that even possible?" — you reported a value without sanity-checking it
• "Why did that point get dropped?" — you adjusted without documenting the cause
• "Did you check leakage?" — suspiciously good performance you didn't investigate
• "That seems too clean" — question the pipeline, not just the result

Common Rationalizations

| Excuse | Reality |

|--------|---------|

| "It's clearly an outlier, just remove it" | "Clearly" is a guess. Find why it's extreme; document if you remove it. |

| "Re-running with a new seed fixed it" | You changed the result by chance, not the cause. Investigate. |

| "The good result is the right one" | Convenient results need MORE scrutiny, not less. |

| "Emergency, no time to investigate" | Investigation is faster than retracting a wrong finding. |

| "Tweak the model, then understand it" | The first tweak sets a false trail. Understand first. |

| "Performance is just high, ship it" | Suspiciously high performance is usually leakage. Check. |

Quick Reference

| Phase | Activities | Success Criteria |

|-------|-----------|------------------|

| 1. Characterize | Read, reproduce, check changes, instrument stages, classify | Know WHERE and WHICH (bug/data/finding) |

| 2. Pattern | Working comparison, reference, list differences | Identify the difference |

| 3. Hypothesis | Single theory, minimal test | Confirmed or new hypothesis |

| 4. Resolution | Fix bug / document data rule / report real finding | Anomaly explained, not just hidden |

Related Skills

• science-superpowers:verifying-results-before-claiming — confirm the resolution actually holds
• science-superpowers:preregistering-analysis — a real but unregistered finding needs a fresh confirmatory pre-registration
• science-superpowers:setting-up-reproducible-analysis — deterministic reproduction depends on fixed seeds and pinned environments