---
name: contradiction-map
description: "Builds a claim-by-source conflict matrix from several sources on one topic — one row per contested claim, one column per source, and a typed label on every disagreement (direct, scope, definitional, methodological, or temporal). Use when someone hands you multiple documents, studies, or notes and wants the disagreements isolated and classified rather than reconciled into a story. Do NOT use for merging sources into one narrative, for checking a single document against itself, or for deciding which source is right."
metadata:
  decimalai-skill-id: "dbbc1d03-63da-4db5-b818-074db1a228c8"
  decimalai-version-id: "bb01ba1b-cde5-49d5-990c-e04715d2b4e2"
  decimalai-version: "1"
  decimalai-body-sha256: "6ce64c3bbf41eea7b624dcb6692f2a5625350ae5af018d466dd8c8f50380ad7e"
  decimalai-name: "contradiction-map"
  decimalai-format: "1"
  decimalai-invocation: "model"
  decimalai-category: "research"
  decimalai-source: "https://app.decimal.ai/s/contradiction-map@1/SKILL.md"
---
# Contradiction Map

## What this does

Takes several sources on one topic — studies, analyst notes, filings, retrieved passages — and produces a **matrix of exactly where they disagree**. One row per contested claim, one column per source, one cell per source's position on that claim, and on every row a **typed label** saying *what kind* of disagreement it is.

The deliverable is the map, not a verdict. It does not pick a winner, average the numbers, or weave the sources into one account. It isolates the conflicts and names their type, so the reader can see at a glance which disagreements are real contradictions and which only look like conflicts because the sources are measuring different things.

The base model, handed conflicting sources and asked "where do these disagree?", does three things this skill overrides. It **narrates** ("Source A says X, but Source B says Y, so on balance…") instead of tabulating. It treats every gap as a flat contradiction, missing that two different numbers can both be true when the sources cover different populations, periods, or definitions. And it **resolves** — declaring one source more reliable and siding with it — when the job was to map the conflict, not settle it.

## When to use / when NOT to

**Use when:** the input is two or more sources on a shared topic and the ask is to isolate and classify the disagreements — "where exactly do these conflict", "give me a conflict matrix", "which of these actually contradict and which just look like they do", "table the disagreements, don't reconcile them".

**Do NOT use when:**
- The ask is to **merge** the sources into one theme-organized narrative of what they collectively say — that is synthesis, a different job.
- There is **one** document and the question is whether it agrees with itself.
- The ask is to **decide which source is correct** or to rank source reliability — this skill maps the conflict and stops short of adjudicating it.

## The core discipline

**1. Output a matrix, not prose.** Rows are contested claims; columns are sources. Each cell states that source's position on that claim in a few words, or `—` if the source is silent on it. Never render the comparison as running paragraphs.

**2. One row per genuinely contested claim.** A claim earns a row only if at least two sources take *positions that differ* on it. Points where all sources agree, and claims only one source makes, do not belong in the matrix (note them in a one-line aside if useful, but the matrix is the disagreements).

**3. Type every disagreement.** Assign each row exactly one type from the taxonomy below. The type is the heart of the deliverable — it tells the reader whether the sources actually contradict.

**4. Do not resolve.** Never pick a winner, average conflicting figures into one number, silently drop an outlier, or rank the sources by trustworthiness. You may add a short **"what would settle it"** note per row (the missing fact that would resolve the conflict), but you state the conflict — you do not adjudicate it.

**5. Do not import outside facts.** Every cell traces to one of the given sources. Do not add figures or claims none of the sources stated.

## The disagreement taxonomy

Every contested row gets exactly one of these five types. The distinction that carries the value: only **direct** disagreements are genuine contradictions where the sources cannot both be true. The other four are *apparent* conflicts — the sources can all be correct, because they are not actually talking about the same thing.

| Type | The sources… | Can both be true? | Tell it apart by |
|---|---|---|---|
| **direct** | make the same claim, over the same scope, with incompatible values or opposite conclusions | **No** | same population, same period, same definition, same method — yet the answers conflict |
| **scope** | measure different populations, segments, geographies, or conditions | Yes | the "who/where/under what conditions" differs between sources |
| **definitional** | use the same term to mean different things | Yes | the same word ("active user", "revenue", "churn") is defined differently |
| **methodological** | use different measurement methods, instruments, or study designs | Yes | the disagreement is about *how* it was measured, not *what* was found |
| **temporal** | measured at different times or over different periods | Yes | the "as of when" differs; the gap may reflect real change over time |

When two types both seem to apply, pick the one that is the **primary** driver of the gap and name the secondary one in the row's note. Do not leave a row untyped, and do not invent a sixth type.

### Worked distinctions

- Two employment studies of the **same** minimum-wage raise, **same** period, **same** definition of "job" reach opposite conclusions → **direct**. They cannot both be right.
- One study finds remote work raises output 13% (call-center agents); another finds it flat (mixed knowledge-work roles) → **scope**. Different worker populations; both can hold.
- One report says "churn is 4%", another "churn is 9%" — the first counts logo churn, the second revenue churn → **definitional**. Same word, different metric.
- A consumer test finds EV range drops 20% in cold; the manufacturer's white paper finds 8% → **methodological**. Different test protocols, not a factual contradiction about one protocol.
- One survey (2019) puts adoption at 12%; another (2024) at 40% → **temporal**. Both can be correct as of their own date; the gap is likely real change.

## Output format

Lead with the matrix, then the typed rows. A compact shape:

**Conflict matrix** — claims down the side, sources across the top:

| Contested claim | Source 1 | Source 2 | Source 3 |
|---|---|---|---|
| Cold-weather range loss | ~20% | ~8% | — |
| … | … | … | … |

**Typed conflicts** — one line per contested claim:

- **Cold-weather range loss** — *methodological*. S1 (consumer road test) vs. S2 (manufacturer bench protocol). Both measure real range loss under different test conditions. *Would settle it:* running both vehicles under one shared protocol.

Keep each typed line to: the claim, the type, who-says-what by source, one clause of why that type, and an optional "would settle it". No concluding paragraph that reconciles the sources — the map is the end of the job.

## Common failures this skill prevents

- **Narrating instead of mapping** — emitting "Source A argues… whereas Source B contends…" prose with no matrix.
- **Flattening every gap to a contradiction** — labeling a scope or temporal difference as though the sources contradict, when both can be true.
- **Resolving the conflict** — declaring one source correct, averaging two figures, or dropping the outlier, when the task was to map.
- **Manufacturing conflict** — putting a row in the matrix where the sources actually agree, or where only one source speaks.
- **Importing outside facts** — filling a cell with a figure none of the sources stated.