Interview prep + transcription + synthesis

Expert interviews break down in three different places. The researcher walks in with generic questions. The recording is never transcribed properly. Or twenty transcripts sit in a folder and nobody reads past the fourth.

One pipeline covers all three, and the parts should stay separable. Before the call, a tailored question set built from the current deep-dive draft. During and after, a transcript with speakers and timestamps. Across the set, themes that carry the quotes and a link back to the moment.

The deliverable is not a summary of what experts said. It is a themed evidence file. Every claim traces to a person, a minute, and a sentence. A theme with no attached quotes is the researcher's memory, formatted.

Nothing here decides anything. Fourteen experts agreeing is a finding about fourteen experts. Whether that settles the question is a judgment somebody makes later, on the record.

Interviews also sit inside a protocol. If the deep dive declared how experts would be chosen, the pipeline should preserve that record. A check of the deep-dive against its declared protocol will ask for it.

Load the current deep-dive draft before generating anything. The draft is where the gaps live. A claim with one weak source. A mechanism nobody could confirm. A number that came from a single model run. Those are the questions worth an expert's hour.

Then load what is known about this specific expert. What they published, what they built, where they disagree with the field. Names surfaced by an expert finder reading publications and talks usually arrive with that material attached. Use it to ask the one question only this person can answer.

A tailored set is short. Six to eight questions, ordered so the hardest one is not last. Each question should name the draft claim it is testing. That label survives into the transcript and makes synthesis mechanical instead of interpretive.

Do not let the generator write leading questions. A model given a draft will try to confirm it. Ask for at least two questions per interview designed to make the draft wrong. Those are the ones that change the deep dive.

If the draft has no open gaps, do not run interviews yet. Work out which input would most move the answer first. That ordering comes from ranking inputs by how much uncertainty they remove.

A transcript without speaker labels is unusable for synthesis. Two people discussing a disagreement produce text where the attribution is the entire point. Timestamps matter for the same reason. A quote that cannot be relocated in the audio will not survive a challenge.

Expect the transcription to mangle technical vocabulary. Chemistry terms, plant names, acronyms, and units come back wrong in predictable ways. Load a glossary from the deep dive before the first file is processed. Correcting the same term fourteen times by hand is where the time savings disappear.

Consent and recording scope belong on the record with the file. Some experts speak on background. Some will not be quoted by name, and some cannot be recorded at all. Store that permission next to the transcript, not in an email thread. A theme file that quotes a person who refused attribution is a real problem.

Synthesis runs across the set, not within one interview. Cluster statements by the claim they address, then let a person name each cluster. Keep the quotes attached in every view. A theme panel where the quotes are one click away gets read; a summary does not.

This walkthrough is hypothetical, with invented names and figures. An industrial decarbonization team is assessing retrofit options for cement kilns. Ines Cardoso leads the deep dive. The draft claims that alternative fuel substitution is limited by permitting rather than by kiln design.

Fourteen interviews are scheduled over five weeks. Four plant operators, three regulators, four equipment engineers, and three fuel suppliers. Each gets a different question set built from the same draft claim. The regulators are asked about timelines and precedent; the engineers are asked what the kiln actually tolerates.

Synthesis returns nine clusters across the set. The largest holds twenty-two statements about permit duration, drawn from eleven of the fourteen interviews. It splits cleanly once a person reads it. Operators talk about the wait; regulators talk about incomplete applications. Those are two different claims wearing one theme label.

A smaller cluster matters more. Three equipment engineers, independently, describe a chloride buildup problem at high substitution rates. None of them framed it as a barrier. It appeared as an aside in three separate conversations.

That is the finding the draft did not have. Ines pulls the three quotes, checks them against the engineering literature, and rewrites the claim. The rewritten claim now says permitting and kiln chemistry both bind, and names which experts said what.

Clustering counts statements, not people. One expert repeating a point six times produces a cluster of six statements. Six experts agreeing once produces the same cluster. Always show distinct speaker count beside statement count.

Selection bias is heavier here than in any other deep-assess method. Experts who agree to talk are the ones with a view and the time to share it. Record who declined and who never replied, in the same file. A theme drawn from four suppliers when eleven were approached should say so.

The model will smooth hedges out of quotes. A hedged sentence becomes a clean assertion in a summary. Require verbatim quotes in theme panels and never paraphrase inside quotation marks. The hedge is often the most accurate part of the sentence.

Leading questions contaminate the whole set. If every engineer was asked whether permitting was the main barrier, permitting will be the largest theme. Store the question that produced each answer alongside the answer. Then a reviewer can see which themes the interviewer planted.

Interview evidence is also not adversarial review. Experts are inside the field and share its assumptions. A structured attack from skeptical and base-rate personas tests things fourteen practitioners will agree about without noticing.

Cluster names are proposals. The person who sat in the conversation confirms, splits, or rejects each one. They heard the tone, the pause, and the thing said after the recording stopped. That context does not exist in the transcript.

Attribution decisions stay human too. Whether a named expert can be quoted, paraphrased, or only counted is a promise somebody made. The pipeline enforces the flag; it does not set it.

Disagreement between experts should be preserved, not resolved by the synthesis step. Two credible people contradicting each other is information about the state of the field. Carry that forward intact, because extracting which assumption each side rests on needs the raw positions, not an average.

File the transcripts where the next round can find them. An interview from this quarter often answers a question asked two rounds later. That only works if the files sit in a searchable store of decision artifacts rather than a personal drive.

One habit is worth enforcing. Before a theme reaches a recommendation, open two of its quotes and listen to them. The label and the actual sentence disagree more often than anyone expects.