# Wrong Number Debugging

> Use the moment a computed result looks wrong, surprising, suspicious, or "off" — a total that doesn't reconcile, revenue that tripled after a join, a mean that moved, a coefficient with the wrong sign, a count that's too high or too low, a metric that disagrees with another team's. Bisects the data pipeline to find the exact step where the number went bad, tracing provenance backward instead of guessing-and-patching. Use whenever the user says "this number looks wrong", "why is this so high", "these don't match", "the totals are off", "that can't be right", or when a `data-contracts` reconciliation fails — in R, Julia, Python, or Stata.

- Skill: `lancegui/wrong-number-debugging` (Agent Skill)
- Install (CLI): `npx skillmds@latest add lancegui/wrong-number-debugging`
- Raw SKILL.md: https://api.skillmd.com/api/skills/lancegui/wrong-number-debugging/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: DevOps & Infra
- Author: lancegui (https://skillmd.com/u/lancegui)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/lancegui/wrong-number-debugging

---


# Wrong-Number Debugging

## Overview

A surprising number is data trying to tell you something. The instinct is to patch it — add a `dropna`, a `distinct`, a filter — until it looks reasonable. That instinct is how a symptom gets hidden and the real bug ships. **Routing:** code THROWS or a test fails → `superpowers:systematic-debugging`. Code runs CLEAN but the number is wrong → this skill. Either way, a remedy that changes design/sample/spec is `analysis-checkpoints` territory, not a fix.

**Core principle:** Locate the bug by bisecting the pipeline, not by guessing at fixes. The number is wrong *somewhere specific* — find where, then you'll know why.

## Why analytics debugging is its own thing

In software a bug announces itself with a stack trace pointing near the cause. In analysis there is no trace — the pipeline ran clean, and the only signal you have is that the *output* is wrong. Work backward through the chain of joins, filters, groupings, and recodes, checking the number at each stage, until you find where it stopped being right. That stage contains the bug.

## The loop

```
REPRODUCE  →  LOCATE (bisect)  →  EXPLAIN  →  FIX AT THE SOURCE  →  RE-CONTRACT
```

1. **REPRODUCE** — Pin the wrong number down to a deterministic, minimal case. Same input, same seed, same result every time. If it's intermittent, you have hidden state (ordering, randomness, a mutated global) and *that* is the bug. Shrink to the smallest subset of rows that still shows it — debugging on 50 rows beats debugging on 50 million.

   **Then state the diagnostic roadmap and get a quick nod before running scans.** A bisection is a multi-step plan — state it in 2–4 lines: the stages you'll check, in what order, where you'll start. *"Roadmap: (1) pull the flagged records, (2) check X, (3) scan the panel for the pattern, (4) trace how it's produced. Stays a diagnosis — any drop/merge comes back to you. Good, or reprioritize?"* Get agreement once, then execute autonomously — only re-stopping if a step turns into a design/sample/spec change. This is where the user's local knowledge reorders your search cheaply. Skip it for a one-or-two-step check.

2. **LOCATE by bisection** — Execute the agreed roadmap. Walk the pipeline and check the number at each intermediate stage:
   - What is the row count / total / value right after **load**? Is it already wrong? Then the bug is upstream — in the source or the extract, not your code.
   - After each **join**? (Joins are the prime suspect — check row counts before and after every one.)
   - After each **filter**? (Did a filter on a column with `NA`/`missing` silently drop rows you wanted? Null-aware filters surprise people in every language.)
   - After each **group-by / aggregate**? (Wrong grain, double-counting, a `sum` over a fanned-out join.)
   - After each **recode / type cast**? (A coercion turned `"1,000"` into `NA`, or strings into a factor with a surprise level.)

   Binary-search it: check the middle of the chain. Wrong already? Bug is in the first half. Still right? Second half. A ten-step pipeline localizes in three or four checks.

3. **EXPLAIN** — Once you've found the stage, explain *why* in one sentence before touching code. "The join fanned out because `customer_id` is not unique in the orders table." If you can't articulate the mechanism, you haven't found the bug yet — keep bisecting. A patch that merely makes the number *look* right is more dangerous than the obviously-wrong one it replaced — the bug is now hidden, not fixed.

4. **FIX AT THE SOURCE — but first answer the dividing question:** *am I restoring the analysis we agreed on, or changing it?* Answer it before you touch anything; the two branches are very different:
   - **Data-bug fix** (restores the intended computation): the join fanned out, so dedup the right table or aggregate before joining; the units were wrong, so correct them; the date parse broke, so repair it. This returns the analysis to what was already agreed — **do it, then report what you found and fixed.** (Don't slap a `distinct()` on the final output to paper over tripled rows — fix the key, not the symptom.)
   - **Analytical-design change** (changes what is being estimated): the bug is real, but the remedy would change the design, the spec, the sample, or the estimand. **This is not yours to do — STOP and route it through `analysis-checkpoints`.** Present the threat, the candidate remedies, and your recommendation, and let the user decide. Implementing the redesign and presenting it as "the fix" is exactly the behind-the-back decision to avoid.
   - **The trap:** *winsorizing, dropping outliers, adding a control, or restricting the sample* feels like cleaning but is a **sample/spec change** — it belongs on the design side (STOP), not the data-bug side. And a genuine data-bug fix that nonetheless *moves a number the user has already seen* still gets surfaced before you build on it.

5. **RE-CONTRACT** — Add a `data-contracts` check that would have caught this, and watch it bite on the broken version. A bug you found once should never silently return. (If the user approved a *redesign* in the branch above, the new spec re-enters at REPRODUCE — it's a new result to validate, not a closed bug.)

## Trace provenance backward — the usual culprits

The categories worth checking at each bisected stage are the same ones `data-contracts`' invariant catalog asserts and `result-verification`'s checklist reconciles — don't re-derive the list, bisect against it: **fan-out join** (totals inflated by a clean-ish 2×/3×; check row counts and key uniqueness), **vanishing rows** (an inner join or null-dropping filter silently discarded unmatched rows; anti-join to see what failed to match), **NA/missing poisoning**, **units/scale** (dollars vs. cents, ms vs. s), **wrong grain** (double-counting user-sessions as users), **surprise category level** (a stray `"unknown"`, mojibake, a trailing-space duplicate), **temporal** (a timezone shift or resample duplicating a period), and **upstream surgery** (the data was already sampled or filtered before it reached you).

## Language cheat-sheet

| Need | Python | R | Julia |
|---|---|---|---|
| Row count at a stage | `len(df)` / `df.shape[0]` | `nrow(df)` | `nrow(df)` |
| See unmatched join keys | `merge(..., indicator=True)` then count | `anti_join(a, b)` | `antijoin(a, b, on=:id)` |
| Count missing | `df.isna().sum()` | `colSums(is.na(df))` | `count(ismissing, col)` |
| Distinct levels | `df.col.value_counts(dropna=False)` | `table(df$col, useNA="always")` | `countmap(col)` |
| Key uniqueness | `df.id.is_unique` | `!any(duplicated(df$id))` | `allunique(df.id)` |

## Red flags — STOP

- Adding a `dropna` / `distinct` / `filter` to make a number look right, without knowing why it was wrong.
- "It's probably just duplicates" — without having checked key uniqueness.
- Debugging on the full dataset instead of shrinking to a minimal failing case.
- Fixing the final output when the bug is three joins upstream.
- Declaring it fixed without adding a check that bites on the old broken version.
- **Changing the research design, spec, sample, or estimand to make a number behave — without surfacing it to the user as their decision (`analysis-checkpoints`).**

## Common rationalizations

| Excuse | Reality |
|---|---|
| "A `distinct()` makes it match, good enough." | If you don't know which rows were duplicated and why, you don't know what else that `distinct` is silently dropping. |
| "It's a small discrepancy, probably rounding." | "Small" discrepancies are often a few leaking rows. Find the rows before you blame the floats. |
| "I'll just rebuild the query from scratch." | You'll likely reintroduce the same bug. Localize it first; understand it; then rebuild if you must. |
| "The number looks reasonable now." | "Looks reasonable" is the exact disguise a hidden bug wears. Reasonable ≠ reconciled. |

## The Process

1. **Reproduce minimally, then state the diagnostic roadmap (stages, order, where you start) and get a quick nod** — a real bisect is a multi-step plan the user should be able to reorder before you run scans; agree once, then execute autonomously. Name the mechanism in one sentence. No fix until you can.
2. **Answer the dividing question** — restoring the agreed analysis, or changing it?
3. **Data bug → fix at the source, then *invoke `data-contracts`*** to add the invariant that would have caught it and watch it bite on the broken version — never patch the symptom on the final output.
4. **Design/sample/spec/estimand change → STOP and *invoke `analysis-checkpoints`*.** Present the threat, the candidate remedies, and your recommendation; do not smuggle a redesign in as a bug fix. An approved redesign re-enters at REPRODUCE as a new result to validate.
5. **Log the lesson while the bisect is fresh** — one line to the project's `docs/LESSONS.md` (symptom, cause, the check that would have caught it; create the file if absent). You just paid for the exact failure class; thirty more seconds converts it into a standing check instead of a future re-bisect. If it would bite *any* analysis, propose folding it into the relevant skill (`result-verification` → "Capture what bit you" — skill edits need sign-off).

## The bottom line

```
Wrong number  →  reproduced minimally, bisected to the stage, mechanism named, fixed at the source, check added
Otherwise     →  a symptom patched and a bug still in the pipeline
```