There is a number in the Sitcom Royalty ledger that has been telling the truth for seventeen consecutive snapshots, in plain sight, in a file rebuilt on every commit, and it went unread the entire time.
The project maps seventy years of American situation comedy as a network. Performers are nodes, castings are edges, and the question is who actually held the thing together. To weigh an edge properly you need to know how many episodes a person was credited in — a lead across seven seasons is not a guest who came through once — so two of the four centrality models the project computes require a sourced episode count on each credit.
Counts are expensive. Somebody opens the performer's own page and reads it. That is the whole method, and it is deliberately slow.
At snapshot v48 the corpus held 1,075 credits and 971 of them carried a read count. Ninety percent. The two quantified models were nearly ready to be read for the first time.
At snapshot v65 the corpus holds 2,734 credits and 979 of them carry a read count. Thirty-six percent.
Read those two figures again, in the order that matters
The share collapsed. The count did not.
Across seventeen recorded states of the corpus the numerator moved by eight. Eight credits acquired a sourced episode count in a year of work. In the same span the denominator grew by 1,659.
Coverage did not fall because counting became harder, or because a standard tightened, or because somebody found an error in the earlier figures. Coverage fell because counting stopped and acquiring did not. Every wave that admitted a kingdom admitted its cast, and every cast member arrived carrying a credit with no number on it. The percentage was not measuring the research. It was measuring the ratio between two activities, one of which had quietly ceased.
This is not a failure of the instrument. It is the instrument working, at some length, into a room with nobody in it.
The honest thing and the useless thing are the same thing here
The project's governing rule is that an unknown stays unknown. No episode count is ever inferred from a series total; no credit is floored at an imagined single appearance. An unmeasured credit is simply absent from the two models that need measurement, and those models refuse to be read at all below fifty percent coverage, where they label themselves provisional and are excluded from every claim the project makes.
That rule is correct and it has a failure mode, which the ledger also names: a blank with nothing scheduled against it is not discipline, it is absence.
A corpus can be scrupulously honest about what it does not know and completely silent about what it could know by Tuesday. The difference between those two states is invisible in the percentage. Both of them print 36%.
So the audit was rebuilt to carry the queue as well as the gap — every knowable-but-unrecorded fact, counted, ordered by what a row of work unblocks rather than by how many rows are in it. 387 more counted credits clears the gate. The worksheet that supplies them is ordered so that forty-six series articles produce most of them, because one page read that yields twenty-two counts is worth twenty-two page reads that yield one each.
That is a week of somebody's evenings. It is not a research programme, and it was never the reason the number stalled.
What actually stalled
Admission is exciting and counting is not.
Ruling a hundred and twenty shows into a corpus is an act with a visible result: the map gets denser, new corridors open, the components fall from sixty-three to thirty-seven, and the census note describing the wave is a paragraph worth writing. Opening a series article to establish that a supporting player appeared in nine episodes produces one cell.
Both are the same project. Only one of them feels like it.
The ledger has no opinion about which is more interesting, and that is precisely its use: it reported the consequence of that preference, quarter after quarter, in a line that reads 979/2,734 and does not editorialise. The instrument was built to catch exactly this — a claim resting on a measurement that had not been made — and it caught it. What it cannot do is make anybody look.
The general case
Every measured project eventually acquires an activity that improves its numbers and an activity that improves its coverage, and they are rarely the same activity. The first is legible, publishable and pleasant. The second is arithmetic in a spreadsheet.
The instruments that survive contact with this are the ones that report coverage beside every figure, refuse to be quoted below a threshold, and record their own history so that a stall is visible as a flat line rather than as a feeling that things have been slow lately.
The flat line is the deliverable. A corpus that could not see its own numerator stop would have gone on publishing a percentage that fell for reasons it would have had to invent — the research got harder, the standard rose, the field is simply large. None of those is what happened. Counting stopped, and the record says so in the only way a record can, which is by having been kept.
The figures above are read from the project's own census at the fingerprint it carries today, and they will be different when you check. They are meant to be checked: the ledger is public, the audit that produced them is published unedited, and the queue of what remains unknown is longer than the list of what is settled.
That ratio is not an embarrassment. It is the only honest shape for a thing that is still being made.