CSE Roadmap

Common tier · about 4 hours · 11 parts

Part 3 of 11 · 2 min read

Sources disagree, so surface it instead of hiding it

Picking one value and staying silent about the rest hides exactly the information a reader needs to judge it.

Think of a doctor who gets two different test results back and simply reports the one that looks more official, never mentioning the other result existed. A patient would want to know both numbers came back, even if one is trusted more.

In plain words: when sources disagree, choosing the most trustworthy one to answer with is right. But silently dropping the disagreeing values hides information a careful reader would want. The agent should report the chosen value and name what the other sources said.

A worked example. Among the 17 answered facts, several carry a visible disagreement that per-need search surfaced. Zephyr DB's p95 latency is reported as 42ms from a primary 2026 source, with a forum post claiming 15ms and a 2024 blog claiming 80ms. Lumen Cache's license is BSD-3 from the primary source, against a forum rumor of GPL-3. In every case the chosen value comes from ranking source kind first (primary beats secondary beats forum), then recency. But the disagreeing values are still shown, not discarded.

The key lines.

# from code/mini-research/research.py
RANK = {"primary": 3, "secondary": 2, "forum": 1}

def resolve(candidates):
    best = max(candidates, key=lambda c: (RANK[c[1]["kind"]], c[1]["date"]))
    others = sorted({v for v, _ in candidates if v != best[0]})
    return best[0], best[1]["id"], others

Read it as: the ranking picks the single best-supported value to answer with, but others is kept and returned alongside it. The disagreement does not disappear just because a decision was made. Three pages that trace to the same original claim should count as one source, not three separate confirmations of it.

Primary source, citation laundering

A primary source is the originator of a fact, such as the official documentation; secondary sources report it, and forum posts are unvetted. Citation laundering is when several sources repeat one unverified claim, making agreement look like independent confirmation when it's really one source copied several times.

Example: the 42ms/15ms/80ms three-way disagreement above, shown to the reader rather than silently resolved.