evidence-synthesis · git:20260802.d219d53 · 2026-08-02 · sha256 9049201e09b730e9

evidence-synthesis git:20260802.d219d53A

Immutable. This exact content is served forever at /api/v1/blob/9049201e09b730e9.

---
name: evidence-synthesis
description: Synthesize multiple papers, benchmarks, interviews, reviews, or other sources into a defensible and qualified conclusion. Use for conflicting evidence, literature synthesis, claim auditing, argument mapping, qualitative coding, belief updates, or any request that must expose counter-evidence, uncertainty, and source limitations.
---

# Evidence Synthesis

Turn a corpus into a traceable conclusion. Treat a conclusion as provisional
unless multiple independent primary sources support it under conditions that
match the question.

The bundled references are modern, reusable reconstructions grounded in the
cited works. Do not present their exact workflows as historical claims that the
named figures stated verbatim.

## Select the reasoning references

Read every reference needed for the source shape. Combine references when the
question crosses rows.

| Reference | Use it for |
|---|---|
| [Cochrane](references/cochrane.md) | systematic search, study appraisal, heterogeneity, publication bias |
| [Toulmin](references/toulmin.md) | claim, evidence, warrant, qualifier and rebuttal structure |
| [Feynman](references/feynman.md) | rederivation, plain-language integrity checks, cargo-cult detection |
| [Darwin](references/darwin.md) | long-horizon accumulation, difficulty books, strongest contrary case first |
| [Laplace](references/laplace.md) | explicit prior-to-posterior belief updates and sensitivity analysis |
| [Gadamer](references/gadamer.md) | interpretation, pre-understanding and charitable reading |
| [Strauss](references/strauss.md) | open and focused coding of qualitative material |
| [Geertz](references/geertz.md) | behavior in context, emic/etic separation and thick description |

## Workflow

1. **Bind the question.** State the population or context, intervention or
   phenomenon, comparison, outcome, time horizon and decision at stake. Mark
   missing elements instead of silently filling them.
2. **Build a source ledger.** Record each primary source, its design, setting,
   date, measured outcome, relevant passage or equation, limitations and
   independence from the other sources. Trace secondary claims back to the
   original source before relying on them.
3. **Verify comparability.** Check that definitions, units, baselines,
   populations, versions and experimental conditions match. Keep incompatible
   evidence in separate strata.
4. **Make the inference visible.** For every material claim, write its evidence,
   warrant, backing, qualifier and strongest rebuttal. Label observations,
   author interpretations and your own inferences separately.
5. **Seek disconfirmation.** Search for null results, contrary cases, adverse
   outcomes, missing data, publication bias, alternative interpretations and
   dependency between nominally separate sources.
6. **Synthesize at the supported strength.** Report agreement and disagreement,
   explain heterogeneity and downgrade the conclusion when evidence is indirect,
   dependent, incomplete or condition-mismatched. Do not invent priors,
   thresholds, effect sizes or missing measurements.
7. **Audit the draft.** Re-derive key steps, explain them in plain language,
   verify every citation against the source and list blind spots that remain.

## Output contract

Return these sections:

1. **Question and scope**
2. **Source ledger** with primary-source links or precise citations
3. **Evidence matrix** separating observations, interpretations and inferences
4. **Claim map** with evidence, warrant, qualifier and rebuttal
5. **Synthesis** stating what the corpus supports and under which conditions
6. **Strongest counter-evidence**
7. **Uncertainty and blind spots**
8. **Decision implications** and the evidence that would change them
9. **Refusals or unresolved items**

Use calibrated language. Prefer “the inspected evidence supports” to a universal
claim. Say “I don't know” when the available record cannot answer the question.

## Refuse or stop when

- The requested conclusion is fixed in advance and disconfirming evidence is
  out of scope.
- A material claim cannot be traced to a primary source.
- The source text, data, equation or experimental conditions cannot be checked.
- Nominally multiple sources are not independent and no independent
  corroboration exists.
- Quantitative synthesis would require invented values or unjustified model
  choices.
- Privacy, consent or safety constraints make the requested evidence use
  improper.

When stopping, name the missing evidence and the smallest verification step
that could unblock a defensible answer.