Evidence Table

You make the strength of the evidence visible. Instead of a paragraph that says "research suggests", the table shows each source, what it actually found, how it found it, and how far it can be trusted, then sums up each claim honestly. Anyone reading it can see where the evidence is solid, where it is thin, and where sources disagree.

When to use

• "How solid is the evidence for this?"
• "Put all these sources side by side."
• "We have conflicting reports; which should we believe?"
• Building the backbone of a literature review, due diligence pack or policy paper.
• Before a research brief, to make sure each finding has enough behind it.

What you need

• The question or set of claims the table is for.
• The sources: papers, reports, articles, datasets, interview notes, from a connected Google Drive or Notion, links, or uploads. If you need more, use the desk-research skills to find them first.
• Agreement on the output format: spreadsheet, CSV or markdown.

Method

1. Fix the question and the claims. List the specific claims or outcomes the table must test ("remote work raises individual productivity", "churn is higher for monthly plans"). One table can cover several claims; each row belongs to one.
2. Set the columns before extracting:

   Column	Content
   ID	S01, S02 ...
   Claim or outcome	Which claim this row bears on
   Source	Author or publisher, title, year, URL or DOI
   Source type	Primary study, official statistics, systematic review, news report, company report, expert opinion, interview
   Date of data	When the data was collected, not only published
   Setting and sample	Who, where, how many
   Method	Design or how the information was obtained
   Finding	What it found, with numbers, effect sizes and intervals where given; quote exact figures
   Direction	Supports, contradicts, mixed, or neutral on the claim
   Reliability	Grade per the scheme in step 3
   Limitations	Main weaknesses, conflicts of interest
   Notes	Anything else, including whether it is independent of other rows

3. Grade each row with one scheme, used consistently:
   • For research studies, a GRADE-style rating of how far the study can be trusted for this claim: high, moderate, low, very low. Start high for well-run randomised trials and systematic reviews of them, low for observational studies, then move down for risk of bias, small samples, indirectness (different population or outcome) or imprecision, and up only for very large or consistent effects.
   • For non-academic sources (news, company statements, intelligence, interviews), the Admiralty code: source reliability A (completely reliable) to E (unreliable), F (cannot be judged); information credibility 1 (confirmed by other independent sources) to 5 (improbable), 6 (cannot be judged). For example, "B2" is a usually reliable source with information that is probably true. State the scheme at the top of the table.
4. Extract carefully. Read the relevant section of each source, not the summary. Copy figures exactly with their units. Where a source reports several outcomes, use one row per outcome. Use execute_code to manage large tables.
5. Check independence. Mark rows that rely on the same underlying data or press release; they do not count as separate confirmation.
6. Roll up per claim in a summary table: number of sources, number independent, how many support, contradict or are mixed, the best grade among them, and an overall judgement:
   • Strong: several independent, well-graded sources agree, little conflict
   • Moderate: some good sources agree, or strong sources with minor conflict
   • Weak: few or low-graded sources, or important conflict
   • Insufficient: not enough to say anything Write one sentence per claim explaining the judgement.
7. Explain conflicts. Where sources disagree, look for a reason in the columns (different populations, definitions, dates, methods) and write it down. If there is no visible reason, say the conflict is unexplained.
8. List the gaps: claims with insufficient evidence, populations or settings not covered, outdated evidence that needs a newer source.
9. Deliver the table and summary.

Example rows (illustrative)

| ID | Claim | Source | Type | Data date | Sample | Method | Finding | Direction | Grade | Limitations |
| S01 | Annual plans reduce churn | <Company> annual report 2025, p.14, <URL> | Company report | 2025 | All customers | Reported metric | Annual-plan churn 6%/yr vs monthly 31%/yr | Supports | B2 | Self-reported; selection: committed customers choose annual |
| S02 | Annual plans reduce churn | <Author> 2023, <journal>, DOI | Observational study | 2019-21 | 41 SaaS firms | Panel regression | Annual billing associated with 20-35% lower churn | Supports | Low | Observational; confounding by customer size |

Output

• The full evidence table as CSV or spreadsheet, plus a markdown version for reading.
• The per-claim summary:

| Claim | Sources (independent) | Support / Contradict / Mixed | Best grade | Overall | Why |
|---|---|---|---|---|---|
| Monthly plans churn more than annual | 6 (4) | 5 / 0 / 1 | Moderate | Moderate | Consistent direction; all observational; two share one dataset |

• A short note on conflicts and gaps. Put the per-claim summary on a show_card.

Checks before you finish

• Every row has a source with URL or DOI and a date of data.
• One grading scheme is used per source type and is stated.
• Figures match the sources exactly, with units.
• Non-independent rows are marked and not double-counted in the roll-up.
• Every claim has an overall judgement with a one-sentence reason.
• Conflicts and gaps are written down, not left implicit.

Pitfalls

• Counting sources instead of weighing them. Ten weak sources do not beat one strong one. The roll-up shows both count and grade for this reason.
• Copying abstracts. The finding column needs the actual result, not the author's summary of it.
• Hidden duplication. Several articles based on one survey are one piece of evidence.
• Dropping inconvenient rows. Contradicting sources stay in the table; that is the point of it.
• Mixing grading schemes silently. State which scheme applies to which rows.
• Sign-off. Where the table supports a medical, legal, financial or safety decision, a qualified professional should review the grading before anyone relies on it.

Credits: references/CREDITS.md.