University Extractor

Prose asks the model; a program-rules block is applied in code after every AI pass, so it holds even when the model ignores it. A block with any error is never registered — half a rule is worse than none, because half a rule still looks like it worked. Every save keeps the previous version.

Tick the clusters to merge and, inside one, the row to keep — the others are folded into it. Two steps, and only the second one removes anything. Ask the model settles the to ask groups and records the answer on those rows — they become certain, and no group is ever paid for twice. Merge them then collapses every certain group into one row, filling any cell the survivor was missing from the rows it absorbs. Whatever the model answers is re-checked against the same arithmetic that ruled pairs out beforehand — a group whose rows disagree on a duration, a credit value, a fee or a declared variant such as (Honours) is split apart before anything is merged. The survivor of a cluster is the row that says the most, and it inherits every cell the others had and records what it absorbed.

Every entry below is one ROW. The panel above groups by LINK, which is why a link shared by three programs shows one verdict there and three here — at most one of those rows owns it. This panel only ever reads. It writes no flag and replaces no link: repairing is what Find better ones in the Links panel does, and keeping the two apart means this can be run over everything, as often as you like, without being a decision. A page nobody could read is unreadable and never wrong — a site that serves a challenge page to every automated client has told us nothing about whether the link is right.

A correction is stored beside the row, not in it, so a re-crawl cannot erase it — and it appears immediately, without re-extracting anything. Your value always wins. When a later crawl extracts something different the correction is listed as drifted rather than overwritten, because a fee corrected last year should not quietly override one the university has since changed. To correct one cell by hand, open the row with ⤢ and use the pencil beside it — against the value you are looking at. Fix selected rows is the only control here that costs anything: it hands the ticked rows to the agent together with the rules the format check says they break, and anything the page and the row do not support is dropped before it is written.

Fix the top rule and every row under it goes with it. One rule broken three hundred times is one instruction to write in Instructions, not three hundred rows to edit. An empty cell is not a failure here — a blank has no shape, and that is what Missing values asks. A cell that says N/A or TBD is, because it reads as an answer in every spreadsheet downstream and is not one. Nothing in this panel writes, fetches or spends.

Every entry below is one ROW. Show rows on a column filters the table to exactly the rows whose cell is empty, and the chips above filter it by verdict. Free, and nothing is written — the verdict is recomputed from the row each time it is asked for, so it can never be stale or describe a row you have since edited. Unusable means the row has no name, no level or no university and cannot be identified at all; incomplete means some other required cell is empty. A placeholder — “N/A”, “TBD”, “unknown” — counts as missing, because in a spreadsheet it reads as an answer and filters as an answer. An empty cell is usually the university's doing rather than the crawler's: many sites never publish a tuition fee. Fill selected is the one control here that costs credits and the one that writes. It climbs four rungs in order — the program's own page (read from what the crawl already stored, so a site that blocks robots is not a dead end), then site: searches on the university's domain, then wherever that page points, then the open web — and only asks the next rung about what the one before could not answer. The third rung is what makes partner-taught degrees fillable: when a university writes “contact the ERC Institute for fee information” and links to them, that link is the university telling you where the fee lives, so what it says counts as official. Fees, deadlines and living costs are dropped unless the university itself stated them, whichever rung found them, and every filled cell records how it was found so the evidence filter stays meaningful.

Every entry below is one LINK, however many rows carry it. Show rows filters the table to exactly the rows holding that link, and the Wrong / Suspect / Sound links filter in the toolbar selects them by verdict across the whole sheet — the link cells there are tinted to match. A link that fails is flagged and kept — never blanked. A wrong link is a lead and an empty cell is nothing, and Program URL has a standing never-empty rule. Nothing is ever swapped in that was not itself visited and judged by the same rules. Replacements are looked for on the source's own pages first, then through a site-scoped web search — which for a row that came from an uploaded PDF is the only place to look. Course brokers are refused wherever they appear. A page that is not in English is read in translation before it is judged, so a Korean, Japanese or Thai source is checked like any other rather than skipped.