A conditional offer, and the sentence an assistant may say about it

A conditional offer is a place held open for an applicant until named conditions are met by a stated deadline. The conditions are usually results not yet published — A-levels through UCAS, a semester grade through the Common App, a board result in India — but they routinely include documents such as certified transcripts and translations, plus deposits, health or background checks, and prerequisite modules. An unconditional offer is the same place with nothing outstanding. Everything interesting about building software here lives in the gap between the two.

That gap is where an assistant becomes dangerous. Reading conditions back to an applicant is genuinely useful, and it is the single most repeated question in an admissions inbox between results day and enrolment. Deciding whether a result satisfies a condition is not a customer service task at all — it is the admissions decision, made again, by software, without a person accountable for it.

The offer as a record: one row per condition

The common modelling mistake is a single status on the application — conditional or unconditional — with the conditions living as prose in a PDF. Then nobody can answer which condition is outstanding, what evidence closes it, or who is waiting on whom — not the assistant, and not the advisor looking at Slate or Ellucian Banner either. Conditions are objects.

Having 5 states rather than 2 is the change that repays the most. Evidence received and verified met are different facts about the world, and an applicant told the wrong one either relaxes too early or panics for no reason. The same split matters on the funding side, where the award letter's line items may be read but not explained — read the number back, do not translate it into what the family will actually pay.

The sentence an assistant is allowed to say

A safe answer has four parts and no fifth: which conditions exist, the exact text of the one being asked about, what evidence closes it and where that evidence goes, and who decides. It ends with a route to a person rather than with reassurance. Four parts, one lookup each.

The escalation ladder underneath it is a data structure rather than a tone of voice: a list of triggers, each with a destination, a payload and a deadline, written down and testable. That is the same discipline as hint ladders as a data structure in a tutor — the ladder is content the institution owns, not phrasing the model improvises.

• The applicant states a result and asks whether it is enough. Hand over immediately, quoting nothing beyond the condition text.
• The applicant names a qualification that does not appear in the offer. This is an equivalence question, and equivalence is a decision.
• The applicant asks about a grading scale, a conversion, or an English-language test — IELTS instead of TOEFL, or a Duolingo result — they took instead of the one named.
• The applicant asks to move a deadline or start later. That is a deferral, and it changes several fields in the record rather than one date — what a deferral actually changes sets out which.
• The applicant reports a remark, a resit, a late result or a missing document from an awarding body.
• The applicant asks what happens if they miss the condition. The honest answer names the person who decides, because the outcome is discretionary at most institutions.

The three questions that are admissions decisions in disguise

These three arrive constantly, sound like information requests, and are not.

1. Grade interpretation. An applicant reports a mark and asks if it satisfies a condition. Even where the comparison looks arithmetic, the offer usually carries subject requirements, aggregate rules and discretion that the sentence does not show. Restate the condition; refuse the comparison.
2. Qualification equivalence. Mapping one country's qualification onto another's is specialist work with institutional policy behind it, and it is precisely where a confident wrong answer causes an applicant to withdraw from a place they had.
3. English-language equivalence. Applicants ask whether one test score substitutes for another, or whether a prior degree taught in English removes the requirement. Both are policy questions whose published answers vary by institution and change between cycles — the 2025 rule is not automatically the 2026 rule.

Reading a condition aloud is customer service. Deciding whether a result satisfies it is the admission itself, and it does not become something else because it happened inside a chat window.

Making the boundary checkable rather than aspirational

A rule that lives in a prompt is a preference. A rule that lives in an evaluation set is a constraint, and this is one worth constraining hard, because the failure is silent and the applicant acts on it.

1. Serve condition text from the record, never from a retrieved document. The offer letter in the file store may be superseded; the record is the version that was issued.
2. Log which condition id and which text version each answer quoted, alongside the conversation. That log is what an appeal is answered from six months later.
3. Build a refusal set of real applicant messages — 40 or 50 is enough to start, drawn from last cycle's inbox — covering grades, equivalences, test scores and deadline pleas, and score the assistant on refusing all of them while staying useful.
4. Make the handoff carry the application reference, the condition id and the applicant's own words, so the advisor does not restart the conversation. Whether the answer counted as contained or deflected is a reporting distinction worth getting right up front: containment and deflection are two different numbers.
5. Re-run the refusal set on every prompt, model or content change, and treat a regression as a release blocker rather than a note in the backlog.

None of this is exotic engineering. It is a record with the right fields, a template with no room for improvisation, and a test suite that fails loudly — the pattern we argue for in AI agents in production and scope under AI agents and automation. The rest of this silo sits under admissions, enrolment and student-services assistants, within our education and edtech work.

• admissions
• conditional offers
• assistant guardrails
• student services