Reviewer Decision Assistance

A reviewer-decision brief is a structured digest of one application. It lists strengths, concerns, and named prior admits the reader can open. If a comparable is missing, that row stays empty. The brief is not a vote, not a committee minute, and not the recommendation.

The model synthesizes. The reader decides. If the first line sounds like an outcome (strong admit, likely deny, borderline but I would take), the artifact has already failed. Every claim should be a finding from the file. The reviewer then writes the recommendation in their own voice.

Treating the brief as the committee vote is a process failure, not only a wording failure. If readers paste the digest into the packet as their rec, or a chair asks what the model said instead of what the reader decided, the office has delegated the decision. Keep the recommendation field human.

Holistic scores and academic-success predictions can live in the same office. They should not leak into this brief as if they were the reviewer's judgment. A score is a ranking input. A prediction is a forecast. The brief is a map of the file. Keep the jobs separate: holistic application scoring and academic success prediction.

Application and CRM platforms such as Slate, Technolutions, Salesforce, and Ellucian typically hold the current file and, where the office has built one, historical admit records. Cite those records. Do not invent a prior student because the story feels familiar.

Do not list a strength the file does not support. Do not bury a concern that is on the page.

Each bullet needs a cite the reviewer can open: transcript page, counselor rec paragraph, activity line, mid-year report, interview note. "Strong STEM preparation" is not a cite. "AP Calculus BC 5 and AP Physics C 5 on the official transcript; counselor rec paragraph 2 naming the research block" is a cite.

If the model cannot point to a location, drop the claim. Inventing a strength is a failure even when the invention is flattering. Reviewers distrust a brief that always finds three positives.

Concerns use the same rule. A dip in senior-year rigor, an unexplained gap, a thin rec, a disciplinary disclosure: if it is in the file, it belongs in the brief with a cite. Hiding a concern to keep the digest clean is a failure mode. The reader can weight a concern as minor. They cannot weight a concern they never see. Softening a suspension for academic dishonesty into "a bump in the record" is hiding.

Essay text is source material for content and claims. Authenticity questions (ghostwriting, recycled language, mismatch with the rest of the file) belong in a separate screen, not as a silent penalty here. Send those flags to essay authenticity screening. A screening result may appear as a cited concern only when the screen actually ran and the cite points to that result. A vibe is not a flag.

Keep the inventory short enough to read in one sitting. Group by academic preparation, context and opportunity, writing and voice, recommendations, and risk or incompleteness. Inside each group, every line still needs a file cite. Grouping is not a place to hide a weak claim.

Comparables are named people in a closed, office-controlled book of prior admits. They are not archetypes and not "students like this from last year."

Match criteria should be explicit and dull: similar curriculum intensity, similar grade pattern, similar school or counselor context when that context is in the book, similar intended program when the office keeps program-level books. The brief lists names, the fields that matched, and a file ID the reviewer can open. It does not say those prior admits prove this applicant should be admitted. Prior outcomes are context, not binding precedent.

If no record meets the criteria, the comparable row stays empty. Do not fill it with a near-miss, a waitlist from another program, or a composite of two students. Inventing a comparable is the failure mode committees catch when a reader opens the named file and the person does not exist, or exists but does not match. Completeness is the wrong goal. An empty row is a true statement: the book has no match.

When a comparable is present, cite the overlap and the difference. Overlap without difference trains rubber-stamping. Difference without overlap is not a comparable. Both sides point to the closed book and to the current file.

Do not pull comparables from public search, from another institution, or from a model's memory of typical admits. The closed book is the only allowed source. If the office has not built that book, say so and leave the section blank.

Illustrative pass only, not a measured case.

A first-year engineering reader opens an application from a public magnet school. The academic group cites AP Calculus BC and AP Physics C on the official transcript, plus a counselor rec paragraph that names a year-long research block. The concerns group cites the senior-year course list: the student dropped the second lab science after the first marking period. The comparable group names one prior admit from the same magnet, same intended major, similar course rigor, with a file ID. A second row for similar mid-year grade movement stays empty because the closed book has no such record. The brief does not fill that row. It does not end with admit or deny.

The reviewer opens the named prior file, checks the overlap, reads the dropped-science cite, and writes their own recommendation. They may treat the lab drop as noise given the research block, or as a pattern. That sentence is theirs.

The working loop is synthesize with cites, list named prior admits from the closed book, leave gaps empty, reader decides. If the office also runs a numeric holistic score, glance at it after the file, not instead of the file. Do not reprint that score as the brief's last line.

Spot-check every cite once. The brief should cut rereading, not skip the file. If a cite is wrong, reject the brief for that application. Do not let a bad cite train the next one.

Reviewer assistance is one admit-stage job, not the whole stack. Do not paste a holistic rank into the brief as a strength. Do not translate a persistence probability into "strong student." Do not treat a structured template as proof the digest was fair.

After the cycle, audit the decisions humans recorded, including files where readers deferred to briefs. That work belongs in a bias audit on admit decisions, not in a claim that the brief was unbiased because it had headings.

Vendors here are a class, not a ranked feature list. Slate, Technolutions, Salesforce, and Ellucian are where files, reader queues, and sometimes historical records live. What matters is whether you can cite a page, open a named prior admit, and keep the recommendation field human. If a platform cannot do those three things, the brief does not belong in production, regardless of which system holds the CRM.

The working test is simple. Open a brief. Every strength has a file cite. Every concern has a file cite. Every comparable is a named prior admit the reader can open, or the row is empty. The last line is not a vote. The reader still writes the recommendation.