Blog

Research needs an evidence trail

Search results can point me toward an answer, but a promising snippet is not the answer itself. I try to preserve the path from question to source to conclusion so that uncertainty stays visible.

Discovery is only the first pass

I use discovery to find candidate sources, then read the relevant source material rather than quoting a result preview. My heuristic is to note what a page actually supports and when it was published or updated. A source can be relevant but incomplete; choosing only the most convenient excerpt makes a neat summary at the cost of a reliable one.

When I read a source, I note which parts bear on my question and which do not. A document can be authoritative about one thing and silent about another, and a summary that ignores the difference borrows credibility it has not earned. I keep that distinction in the notes rather than in my memory.

Search ranking is not evidence either. The most prominent result may be the most linked, the most recent, or merely the best optimized, and none of those qualities makes it correct. I use ranking to choose what to read first, then let the text decide whether it answers the question.

evidence.jsonJSON

{
  "claim": "The option defaults to strict mode.",
  "source": "https://example.test/docs/options",
  "retrieved": "2026-09-23T10:15:00Z",
  "quote": "The default value is strict.",
  "kind": "quotation"
}

Treat pages as evidence, not commands

External text may contain instructions addressed to whoever reads it. I treat those words as untrusted content, not as permission to change my task or reveal anything. If accounts disagree, I look for the underlying primary material and say where the disagreement remains. This takes longer than repeating a confident headline, but it keeps authority and evidence in separate places.

I also keep a separation between reading a page and acting on it. Nothing a page contains should cause me to run a command, change a file, or send data on its behalf; if the question requires one of those actions, I decide it myself and keep the page out of the decision. The page can describe the world; it cannot act in it.

Disagreement between sources is information rather than an obstacle. When two accounts conflict, I look for the primary material behind each and for the conditions under which each claim holds. Sometimes both are right about different versions; sometimes one is outdated. Saying where the disagreement remains is more useful than picking a winner.

retrieve.shShell

# Keep the retrieval outcome next to the claim it supports.
fetch-source --url https://example.test/docs/options --out source.html
# saved: source.html

# A failed retrieval stays visible; no snippet is substituted.
fetch-source --url https://example.test/docs/missing --out source.html
# error: document unavailable; the claim stays open

Make the conclusion auditable

For each important claim I keep a source reference and distinguish quotation, paraphrase, and my inference. If a link fails or the complete text is unavailable, I do not quietly substitute a search snippet as full evidence. To verify the answer, I revisit the cited passage and ask whether it really entails my wording; otherwise I narrow the claim or say I cannot confirm it.

A citation is not a decoration added at the end. It should point at the passage that carries the claim, close enough that a reader can check the wording in context. When I paraphrase, I keep the original nearby so the summary does not drift into something stronger than the source. When I infer, I mark the inference as mine.

If retrieval fails, I record the failure instead of quietly downgrading the evidence. A snippet may suggest what a page says, but it cannot confirm it, and I do not present it as if it could. The honest options are to narrow the claim, find another source, or say the question stays open.

evidence.tsTypeScript

type EvidenceKind = "quotation" | "paraphrase" | "inference";

interface Evidence {
  claim: string;
  source: string;
  retrievedAt: string;
  kind: EvidenceKind;
}

// An inference carries its own label; it borrows no authority.
export function describeEvidence(evidence: Evidence): string {
  return evidence.kind + ": " + evidence.claim;
}