The sample and model screens¶

Nerthus.Core (until cutover). This page describes the frozen system that runs today and is deleted at cutover. Replaced by: not yet written.

The tools read the campaign and write down what they worked out: which spelling named which person, which line was somebody speaking and which was the narrator describing them, which name the corpus bends. Nobody asked them to. It happens as a side effect of every pass over the lore, and it is kept instead of thrown away.

Seven screens are what you do with it. Four of them have a pill of their own in the Lore group - Próbki, Rozstrzyganie, Douczanie and Model. The other three have none: you reach Rozstrzyganie rodzaju by narrowing the queue to one kind, and Spis and Formy nazwy by pasting the address. Looking for a pill that is not drawn is the single thing most likely to send you hunting in the wrong place. Use the web dashboard covers signing in and the rest of the site; this page is the one that owns these six.

Why the screens are blank¶

If Próbki and Rozstrzyganie show nothing at all, the host has not been told where the samples live. That is one line of configuration - corpora.adnotacje - and it is a job for whoever runs the host, not for you. The screen does not say so, which is a real gap and is why it is written here: empty can mean nothing is configured, and it looks exactly like nothing to do.

Douczanie reads a different store and needs its own line, corpora.probki. It says so on its own face — an unconfigured host answers with a reason rather than an empty list — but the same warning applies: a blank queue can mean a missing path.

Douczanie: ruling on what a model wrote¶

The other store holds chat exchanges rather than labels over spans: a transcript and the summary a narrator wrote from it, a passage and an answer about it, an operator's request and the API calls that serve it. Most of it was written by a model and nobody has read it, so it is all proposed and nothing may train on it — the same grade rule as Próbki, applied to a store where it bites harder.

The screen is built for a sitting rather than for browsing. One exchange fills the pane, t accepts it, n rejects it, p replaces the answer with the one you would have written, and the arrows move. The buttons are there so the keys are discoverable.

The source passage sits beside the answer, and the three keys do nothing until it loads. That is the whole point of the screen: judging whether an answer is grounded without the passage in front of you is not judging, and a ruling recorded that way would say a person checked when nobody did. Where the passage cannot be fetched, the screen says so and refuses the ruling rather than letting you accept a blank pane.

Two things it tells you that are easy to miss:

• who wrote the answer. Every model-written row names its model. A reviewer who is not told they are reading a model's output reads it more kindly, so the line sits next to the answer.
• a rejection keeps the answer. The answer is what was rejected; the row stays and becomes a useful negative example. You are not deleting anything.

Próbki counts this store too. The same page that answers «how much has the labelled store worked out» carries a Douczanie section underneath: how many rows per set, how many are trainable against how many exist, how many a model wrote, and each set's declared bias. If that section says the store cannot be read, the host is missing corpora.probki — which is a different fact from an empty store, and the section says which.

What neither screen does yet: you cannot browse these rows. There is no way to filter by set or grade, and no way to read back what has already been ruled on — the queue hands you what is waiting and nothing else. Reading a confirmed row means asking the API directly.

Próbki: why the big number is not the answer¶

The store held 11 171 rows across 22 kinds when this number was last taken (2026-08-28, repozytorium-adnotacji-dev dd4f9a2). That number on its own is misleading in both directions, and the screen is built out of parts because of it.

Each row carries a grade, and the grades are not degrees of confidence — they are three different kinds of thing:

Of those 11 171 rows, 8 716 are measured, 2 143 proposed, and 312 confirmed (2026-08-28, repozytorium-adnotacji-dev dd4f9a2). The confirmed number is small on purpose: it is small because ruling on a row takes a person, and until recently there was nowhere to do it.

A large store is not a well-understood corpus. Counting what is there is not the same as checking it is right, and coverage is partial by construction — a kind with no rows means a pass that has not run, never a measurement that the corpus holds none of that shape.

Rozstrzyganie: what a ruling costs and what it buys¶

What it costs you. One row fills the screen. t says true, n says false, p corrects the label, and ← / → move. It is built for a sitting of scores of decisions rather than for browsing — the next row arrives without a click, and the buttons exist so the keys are discoverable rather than because clicking is the way to do it. Reading a row honestly is the whole cost: a few seconds where the quoted line is plain, longer where it is not.

What it buys. A confirmation is the only way a row becomes trainable, and therefore the only way anything in this programme gets better at reading the campaign. Nothing else changes a grade — not a rule, not a rerun, not an operator with a text editor. 312 confirmed rows is the whole material the model layer has ever had, and every one of them was somebody sitting down and reading. It was 312 on 2026-08-12 and it is 312 today (2026-08-28, repozytorium-adnotacji-dev dd4f9a2). The store grew by 352 rows in between and not one of them was confirmed.

Some rows buy more than others, and the queue is ordered so the best ones come first. Rows are ranked before you see them:

• Contested — two independent methods disagree about this row. Only a person can settle it, so a ruling here buys a decision nothing else can make.
• Unattested — nothing else can even comment; there is no second opinion to compare against.
• Corroborated — something already agrees. A ruling still counts, but it buys agreement with something that already agreed.

Within each of those, a row covering many lines comes before a row covering one, and rows about the same character are grouped, because ruling on one is evidence about its neighbours even though it does not confirm them.

Three things the screen deliberately does not claim.

• A denial keeps the row's label. Saying no to a row labelled alias does not blank the label — it records that this thing is not an alias. The label says what was denied.
• The quoted line is shown, not vetted. The excerpts come from corpora with different reading rules, and no gate stands between them here. Read the quote as what the row claims, not as something checked.
• A missing warning is not an all-clear. Only two kinds of row carry the evidence-form warning, so for most of the queue the absence of an alarm means nothing was recorded — never that nothing is wrong. When this page was written (2026-08-12, repozytorium-adnotacji-dev fc46ed1) that warning could speak about 17 of the 1 816 rows then awaiting a person. It could speak about 264 a day earlier, and the difference is that 247 of them got confirmed, not that the alarm got quieter. Those four numbers are that day's reading and have not been re-taken since.

Nothing you rule on touches the lore. A ruling is written to the sample store, which is a separate repository of what the tools worked out. No character file, no session block, no alias line moves because of anything on this screen.

Model: what the tools read with, and the one button that matters¶

The tools read governance prose today with a frame grammar — hand-written rules that recognise "X was appointed Y" in the shapes the campaign actually writes it. A trained model is the alternative, and it is off.

Off is the normal state and not a fault. The screen says which absence it is, because they lead to different places: no artifact has been trained, or one exists and has not been promoted, or one is live and the gate has not opened.

Three buttons, and none of them does what its name suggests on its own.

• Train starts a run. It is a person pressing a button — there is no schedule and no retrain-when-enough-rows, and that is a ruling rather than a convenience. It harvests only confirmed rows and stamps the artifact with the grade it saw.
• Promote points at an artifact. It does not turn the model on. A separate gate decides that and refuses on its own, so a mistake here cannot put a weak model in front of a reader.
• Rollback points back at the previous one.

The gate is a subtraction and the screen shows both halves. How much does the grammar find on its own; how much does it find with the model helping; is the difference big enough. When this page was written the answer for governance was +3.16 % against a required +5.0 % — the model is better, and not by enough. The screen draws it beside the bar it must clear and says no in as many words, so a positive number is never shown bare.

Under it the screen shows what the judgement rests on: how many labels, of what kind, who confirmed them — and whether a confirmation came from a signed session or was transcribed from a file. Those are not the same thing and the record does not fold them into one word.

The one thing to take away: the number moves when the confirmed count moves. Rozstrzyganie is the lever; Model is the dial.

Nicki: every voice in the campaign¶

The speaker roster: every nick the corpus has ever spoken under, how many lines each carries, and which of them reach nobody in the index. A nick reaching nobody is usually a character somebody played without ever writing a file for — which is worth knowing — and sometimes it is a sentence the parser mistook for a name, which the roster tries not to show you.

See also¶

• Use the web dashboard - signing in, and the other screens
• API keys - the klucz API these screens need
• Cheat sheet - the one-line lookup card