Structured one-pager auto-draft

Analysts reach the shallow-assess stage holding a stack of candidates in a dozen different shapes. Some arrived as PDFs, some as web pages, some as a paragraph from a colleague. The output of this job is one page per candidate, in the same field order every time. Every filled field points back to the document it came from.

The one-pager does not rank, score, or recommend. It makes candidates comparable so that a person can rank them later. A drafter that quietly adds a judgment about promise or fit has changed jobs. That judgment belongs downstream, with a named reviewer and a written rubric.

Two inputs feed the draft: public material about the candidate and whatever internal parsing already produced. Both are text somebody else wrote. Neither becomes a fact about the candidate simply because it landed in a tidy box. Upstream deduplication matters here, because near-duplicate candidates merged before scoring keeps the stack from carrying the same organization twice under two names.

The template is the whole product. Write the fields once, name them precisely, and change them only between rounds. A stack where candidate four has a field that candidate nine lacks is not comparable. Comparability is the reason the one-pager exists.

Each field needs three things written down before any parsing runs. A definition a stranger could apply. An allowed source class, such as a filed report, a registry record, or an internal memo. And an explicit instruction for what to write when nothing in scope answers it.

The answer to the third one is almost always the same. Leave it blank and name the field as unsourced. A blank field is a research task. A field filled from general knowledge about that sector is a claim nobody can trace.

Thresholds matter for numeric fields. Decide in advance whether a range counts as an answer, and whether a figure from a third party may fill a field about the candidate. Write the rule down. Ambiguity there produces a stack where two analysts would have filled the same cell differently.

Set the corpus boundary alongside the template. Name which public sources are in scope and which internal documents this round may read. Work retrieved through prior internal work with its decision attached can populate a history field, and it should be labeled as internal. Mixing a public claim and an internal note into one sentence destroys the audit trail.

Drafting runs field by field, not document by document. For each field, the model searches the candidate's in-scope material for a passage that answers it. It then writes the field in the template's wording and attaches the locator. One field, one passage, one pointer.

Mark every field with its type so a reader knows what they are looking at. Quoted means the source says it in those words. Extracted means a number or date lifted from a table. Summarized means the model compressed a longer passage, and the passage is still linked. Blank means nothing in scope answered the question.

Keep units and dates in the form the source used. Converting a budget figure to a common currency looks helpful and hides an assumption. If conversion is needed, do it in a separate column and keep the original.

Order matters for readability. Keep the template order fixed even when a candidate's material presents things differently. A reader who scans forty pages learns where the funding field sits and stops looking for it.

Verification is a different job and should stay one. The drafter attaches a pointer; it does not confirm that the pointer supports the sentence. That check belongs with a claim-to-passage verifier that flags unsupported lines, run after the stack is compiled.

This walkthrough is hypothetical, with invented organizations and figures. A health research funder is screening thirty-eight candidate programs for a maternal care portfolio. The template has nine fields: what the program does, where it operates, who runs it, how it is funded, the stated outcome, the strongest evidence offered, the delivery model, known partners, and prior contact with this funder.

Parsing runs over three sources per candidate. Those are the public description and the most recent annual report. Internal notes from earlier rounds are included where they exist. Thirty-eight drafts come back with the same nine fields in the same order.

The stack is not uniformly full, and that is the point. Thirty-one candidates have a funded amount; seven leave it blank because no source stated one. Twelve have a filled evidence field that quotes a named study. Nineteen have an evidence field that quotes only the program's own claim about itself, marked as such. Seven are blank.

The blanks change what the analyst does next. Seven missing evidence fields become seven specific questions, not a vague request for more information. Those questions are cheaper to ask because they are precise. Ranking which of them to chase first is the job of scoping by which input most reduces uncertainty.

The drafts that cause damage are the ones that read well. A model asked to fill nine fields will fill nine fields. When the source is silent, it writes the sentence that organizations like this one usually deserve. That sentence is plausible, generic, and untraceable, and it will be quoted in a meeting.

Sector priors are the specific mechanism. Ask for a delivery model on a community health program. The model knows what those usually look like. It will describe that pattern with no source behind it.

Stale sources are the second problem. An annual report from three years ago describes a staff size that has since halved. Put the source date in the field, not only in a footnote. Two candidates with the same figure and different source years are not comparable.

The third failure is template drift. Someone adds a field mid-round because one candidate had interesting information. Now half the stack has it and half does not, and the absence reads as a negative. Add fields between rounds, and re-run the whole stack when you do.

A fourth failure is quiet reconciliation. Two sources disagree about a founding year or a headcount. The draft picks one and moves on. Record both, name the sources, and let the analyst see the conflict.

A compiled one-pager is a draft until a named analyst reads it against its sources. That reading is short when the fields are cited and the blanks are honest. It is long when the drafter has been generous. The time saved is in compilation, not in judgment.

The analyst decides three things per candidate. Whether the filled fields are supported by what they point at. Whether a blank is a real gap or a search that was not run. And whether this candidate is now ready to be compared with the others.

Do not let the one-pager travel without its blanks. Stripping empty fields for a cleaner slide is how a gap becomes an absence of concern. The blank is the most useful thing on the page.

Downstream work assumes this discipline held. Cells filled across candidates by a comparison matrix built from defined columns inherit whatever the one-pagers got wrong. So does anything a reviewer carries into deep assessment. Sign the stack, or send it back.