The assistant is answering with last year's fee schedule

The assistant is quoting a superseded year because both years are in the index and nothing ever told retrieval which one is current. That is a corpus fault, not a model fault, and the distinction matters because the repairs are entirely different: no amount of prompt work makes a system prefer a document it has no reason to believe is newer.

It is also the failure a parent reports rather than the one you catch internally, because the answer is fluent and specific. A wrong-year figure looks like an error only once somebody compares it with the published page, and by then the family has budgeted against it.

Trace 10 answers back to the file they came from

Before anyone theorises about ranking, establish what the system is actually reading. The whole check is a sample, a lookup and a spreadsheet.

1. Take 10 recent answers that quoted a figure, a deadline or an instalment date. Draw them from real conversations, not from a test set, because the test set was written against the corpus you already believe in.
2. For each, resolve every retrieved passage to a source: file name or URL, the document's stated academic year, its publication or last-modified date, and where it physically lives — the CMS, a file store, a departmental subdomain.
3. Put the answer's implied year beside the document's stated year in 2 columns. Disagreement between them is the finding; agreement moves you to the next check.
4. Look at what else was retrieved and not used. If the current year's document was in the candidate set and ranked below the old one, this is a ranking problem. If it was never a candidate, it is an indexing or a publication problem, and those have different owners.
5. Note whether the answer carried a citation a human could open. An untraceable answer is a separate defect, and it blocks every diagnosis of this kind.

Why last year's PDF beats this year's page

Operators assume the newer document loses on a technicality. Usually it loses on merit, by the only measure retrieval has.

• The old artefact is prose and the new one is a table. A previous year's PDF often explains the schedule in sentences, while the current year is published as an HTML table whose cells chunk into fragments with no surrounding context.
• The old file repeats the query language. A document titled around tuition, fees and instalments matches a student's phrasing more times than a page that says the same thing once in a heading.
• Recency is not a signal unless you made it one. Vector similarity has no opinion about dates, and a last-modified timestamp only affects ranking if something in the pipeline reads it and applies it.
• The old file is more linked and more downloaded, so any popularity signal in the stack points the wrong way.
• The current page may sit behind a login, a cookie wall or a JavaScript render the crawler silently failed on, leaving exactly 1 candidate in the index.

Nothing in the system is broken. Two documents were indexed, one was better written, and no component was ever given the job of knowing which year we are in.

Five ways a superseded year stays reachable

The last row gets misdiagnosed hardest, because the system looks broken and nothing is: the institution has not decided the number yet. Tuning retrieval against that is wasted effort, and the correct response is a refusal with a named contact.

A fee question carries 2 years, not 1

Year-scoping fails when it is built as a single filter, because the year in the document and the year that applies to the student are different facts. A continuing student is frequently charged under the schedule of the year they entered, while an applicant is charged under the year they will start — and both are asking the same question in the same words.

1. Resolve the academic year of the answer from the conversation: an explicit year, the intake term on the record, or the institution's current cycle as the fallback. Log which of the 3 was used.
2. Resolve the cohort dimension separately for anyone already enrolled, from the record rather than from the question. This is a lookup, not an inference, and it is a good reason for the assistant to have a record-lookup path at all.
3. Filter the candidate set on the resolved year before ranking. A boost is not a filter: a boost still lets a superseded document win when it is 3 times as similar.
4. State the year in the answer, every time, in the same sentence as the figure. Undated numbers are how the same defect returns 12 months later and nobody can tell whether it was ever fixed.
5. Handle the cycle boundary explicitly, where 2 years are legitimately live at once and the right answer depends on which intake is meant.

Naming the year in the answer also protects the translated versions, where a figure can survive a translation intact while the year label around it does not — a divergence covered in the translated answer that contradicts the English policy page.

Supersession is an ingest decision, not an answer-time one

If currency is decided when a question arrives, every question re-litigates it, and the rule lives in a prompt where nobody can audit it. Decide once, at the point a document enters the corpus. The full ingest method for institutional documents — year stamping, programme-level chunking, an authority order between sources — is set out in preparing a catalogue and handbook set for retrieval; the part that matters here is narrow.

• Stamp the academic year at ingest from the document itself, not from the upload date. A file loaded in August can be next year's schedule or a scan of a 3-year-old one.
• Make supersession an explicit state with a pointer. Superseded-by is a link to the replacing document, so a stale citation can be answered with the current one instead of nothing.
• Withdraw superseded documents from the retrievable set rather than deleting them. Appeals need the text a student was shown at the time.
• Watch every location that can publish, including the departmental sites and shared drives nobody listed in the original scope. The prospectus deserves its own treatment, argued in the prospectus PDF as a source of truth.
• Fail the ingest run loudly when a document arrives with no resolvable year. Silent acceptance is what put the old file in the index in the first place.

A publication register — 1 row per document, with its year, its owner, its approval date and its supersession state — is the piece of software that makes all of this enforceable, and it is small: ordinary internal tools and operations work sitting beside the CMS rather than replacing it.

What the assistant says when the current year does not exist yet

Between the end of one cycle and the approval of the next, the honest answer is that the figure is not set. An assistant that answers anyway will reach for the only document it has, which is the one that is wrong. The refusal has to be specific enough to be useful: the year is not yet published, the last published year was this one and is not a guide to the next, here is the office that owns it, and here is the expected publication window if one is known.

Route the question rather than dropping it. That escalation is the right outcome and carries its own risk, because an escalation is a write into the CRM and that is exactly where a second enquirer record gets created. Peak weeks make both worse at once, which is why the publication state of every fee and deadline document belongs on the pre-peak list in readiness checks before results day and add-drop week.

What correct year-scoping still will not fix

Year-scoping gets the right document in front of the model. It does nothing about what happens next. An assistant that adds instalments together, applies a discount or converts a currency is computing rather than quoting, and that is a different defect with a different repair — the one in the working is right and the arithmetic is wrong. The safe rule for money is to quote a figure exactly as printed, with its year and its citation, and to hand over anything requiring a calculation.

It also will not fix a schedule that is correct and unreadable, where the published table runs to 40 rows of programme-by-residency variants and the student cannot find their own line. Name that as a content design problem before another sprint goes into ranking. Both sit under admissions, enrolment and student-services assistants, within our education and edtech work.

• retrieval
• document versioning
• admissions
• diagnosis