Citations and verification.

LADLE's cited web search returns answers with source links. That's useful, but it's not the same as those sources having been verified. A citation is a pointer — it tells you where the model got the claim. Whether the source itself is accurate, current, or being read correctly is a separate question.

For most casual use, this distinction doesn't matter — you're checking a general fact and the citation is close enough. For anything that would go into a client deliverable, a filing, a published article, or a decision that costs money, the verification workflow is real work.

The three levels of citation

**Level 1 — Reference for orientation.** You want to know roughly what year X happened. LADLE cites a Wikipedia article that says 2018. Good enough for the moment. Verification: none needed unless it becomes load-bearing.

**Level 2 — Support for a claim in your work.** You're writing a memo and cite "as of 2024, the mortgage default rate was X%". LADLE provides a source. Verification: open the source, confirm the number, confirm the year, note that memo says "as of 2024" and the source is dated in 2024.

**Level 3 — Load-bearing claim in high-stakes work.** You're filing a brief, publishing an article, or making a $10M decision on this fact. LADLE cites a paper. Verification: read the source, understand its methodology and limitations, and if the claim doesn't have primary-source backing, either find the primary source or don't make the claim.

Common failure modes

**Citation drift.** The model summarizes what a source says, but the summary shifts the meaning slightly. The specific numbers, conditions, or scope of the original claim may not match the summary. Always compare the summary to the source's actual language for anything load-bearing.

**Stale sources.** LADLE's search returns current results, but the sources themselves can be dated. A 2019 article cited authoritatively for a 2026 claim is often wrong — things change. Check the date of the source, not just the search date.

**Secondary-source loops.** A common failure: source A cites source B cites source C, and source C doesn't actually say what A claims. Use the "trace a claim to its original source" prompt (in /prompts) to walk the citation chain when the claim matters.

**Confidence mismatch.** LADLE will sometimes state something with confidence when the source itself is hedged. If the source says "some evidence suggests X may increase Y in certain populations", and LADLE renders that as "X increases Y", that's a hedging error you need to correct.

The verification workflow

For work that has to be right:

1. Ask the question with an explicit instruction: "Provide sources; do not state claims that aren't in the cited sources." 2. Open each cited source. Read the specific passage the claim comes from. 3. Compare the summary in LADLE's response to the source's actual language. Note any drift. 4. If the source is secondary, ask LADLE to identify the primary source and repeat. 5. Any claim that doesn't survive this — either soften it in your writing to match the source's actual confidence level, or don't include it.

This sounds heavy. In practice, once you've done it a few times, it takes 5-10 minutes per major claim, which is proportionate for anything that matters.

What LADLE will do to help

The model has been prompted to hedge appropriately by default and to distinguish "I found a source that says X" from "X is definitely true". If a claim you're asking about doesn't have a source, LADLE will usually say so rather than fabricate. This is not perfect — hallucinated citations still happen, though rarely on recent Sonnet — which is exactly why the verification workflow exists.