Your judgement, one line at a time.

What this worksheet records

The evaluation article gives a method, not a score: watch ten minutes of one recording or one session of your own game and fill in a row each time the commentator speaks. This page is that sheet as a form. It records your own judgement of one sample. It does not watch, listen, rate or measure anything, and a filled sheet is not evidence about any product, this one included.

1. 01Pick one sample

   Any recorded AI commentary: a public recording on this site, a video elsewhere, or your own session recorded with OBS. Fill in the header: date, game, which commentator, the build, and, for a Signal Broker session, the role and Talkativeness.

2. 02One row per spoken line

   Pause the video the moment a line ends. Note the time, what was on screen a few seconds earlier and what was said in a few words. Then the three judgements, Relevant?, On time? and Repeat?, and whether you would keep the line.

3. 03Log the silences

   Each quiet stretch longer than a minute, with what was on screen: right, wrong, or voice recovering.

4. 04Read the tally, write the verdict

   The tally counts your ratings and nothing else; the verdict is a sentence you write. Export JSON to keep the sheet, CSV for a spreadsheet, or print it.

The columns, as the scorecard defines them

The vocabulary is the plain-text scorecard’s, unchanged. A blank rating means “not rated yet” and is counted nowhere.

Header

Date, game, build (Signal Broker shows it under Help/About), role (SPECTATOR or PLAYER), Talkativeness, session length, and whether the source was a recording or your own session and whether OBS recorded it. The form adds one field the paper sheet does not need: which commentator you watched, because the sheet works for any of them.

One row per line

Time (m:ss, or a range such as 3:05–3:27), what was on screen, what was said in a few words (the scorecard’s “What he said (short)”), then the three judgements and Keep?.

Silences

Only stretches longer than a minute, from–to, with what was on screen; mark VOICE RECOVERING separately, because it is a service state, not a decision.

A worked example, from the recordings

The article’s four example rows, as they look on this sheet. LOAD THE WORKED EXAMPLE in the panel fills the form with them (it asks first if your sheet has content). Each line was read from the Broker’s speech bubble in the overlay of a public recording and timed on the audio-derived captions; what was on screen was read from the same frames. They are four moments chosen to show a hit, a miss, a stale line and a hit for contrast, from two recordings, which a real sheet would not mix. They are not a score for any recording, build or game, and they say nothing about how often each happens.

that Ruby Key’s a freaking ten grand ticket … Three grand spare, though

this big guy looks like trouble … shiny blue armor

we’re neck and neck here, 1-1

So Leeds nudges ahead just before the half … right now it’s Leeds’ party

What the loaded example shows in the tally: lines heard 4, hits 3, misses 1, immediate 2, stale 1, n/a 1, repeats 0. That is a count of four chosen moments, not a rate. Keep? is left blank in all four rows because it is taste, and the verdict box stays empty because the sentence is yours. The example logs no silences: the article verified none, so none are invented here.

The tally is a tally

The panel counts rows per rating, exactly what the scorecard’s TOTALS box asks for: lines heard, hits, misses, trivial, immediate, a beat late, stale, repeats, wrong silences, voice recovering, plus the remaining counts (keep yes and no, n/a timing, right silences) and how many rows are still unrated. The copyable TOTALS block writes them in the scorecard’s own words.

• No composite. The page adds nothing up into a score, a percentage or a grade, and it never writes the verdict. Relevance and timing stay separate columns on purpose: a true observation delivered to a scene that no longer contains it is a hit in one column and “stale” in the other.
• Ten minutes, not one line. The same kind of reading, a scoreboard, lands as a hit at 6:00 and a miss at 8:28 in the example. One row proves nothing either way; the method asks for ten minutes.
• Change one thing at a time between runs (role, Talkativeness, game), and keep a separate sheet per run.
• If you publish a sheet, say which build, role and Talkativeness you used, so a reader can compare like with like. The header fields and the export carry them.

Where your entries go

Everything you type is processed in this page. As you type, the page also saves the sheet to this browser’s local storage (the Web Storage API), so it is still here after a reload or tomorrow in the same browser. The panel says when it last saved. Nothing is transmitted: no server receives the sheet, nothing is written into the address bar, and no account is involved. The site’s ordinary analytics runs on this page as on every other page; it does not read the form.

• This browser only. Another browser, another device, a private window that discards site data, or a browser whose site data you clear does not have the sheet. Export JSON to keep a copy or move it.
• When saving is not available (storage blocked or full), the panel says so; the sheet stays on the page until you leave or reload it, and the exports still work.
• CLEAR THE SHEET empties the form and removes the saved copy from this browser; it asks first. Removing a single row offers an UNDO for a few seconds.
• Limits. Up to 500 lines, 200 silences, 2,000 characters per text field and 200 per time field, which is more than ten minutes of any commentary needs.

Export, import and print

CSV is for a spreadsheet: one row per line and one per silence, with the sheet’s header and your verdict repeated on every row so a row stands on its own when sheets are combined. Columns: kind, time, from, to, what_was_on_screen, what_was_said, relevant, on_time, repeat, keep, silence, date, game, commentator, build, role, talkativeness, session_length, source, obs_recorded, verdict. Fields are quoted and quotes doubled as RFC 4180 describes; the file is UTF-8 with a byte-order mark so Excel reads the curly quotes and dashes; a cell that begins with =, +, - or @ gets a leading apostrophe so no spreadsheet runs it as a formula. CSV is not read back in.

JSON is the whole sheet (schema signal-broker/commentary-evaluation-worksheet, version 1: header, lines, silences, verdict) and the file to keep. IMPORT JSON reads a file you choose with the browser’s FileReader, checks it and replaces what is on the sheet, asking first if the sheet has content. A file is refused, with the reason named, when it is not valid JSON, not a worksheet, a newer version than this page understands, or when a rating is not on the scorecard’s scales; over-long text is cut and reported. Nothing from a file is ever rendered as HTML.

PRINT opens the browser’s print dialog with a layout that hides the site’s header, footer, buttons and these explanations and prints the header, the lines, the silences, the totals and your verdict in black on white. A row’s text fields grow with their content, so nothing is cut off.

Limits

• The worksheet records your own judgement of one sample. It is not an automatic evaluation, not a benchmark and not evidence about any product; the tally counts your ratings and computes no score and no verdict.
• Entries live in this page and, when the browser allows it, in this browser’s local storage only: another browser, another device or a cleared browser does not have them. Export JSON to keep a copy.
• The worked example is four single moments from two recordings, chosen to show the columns; it is not a score for any recording, build or game, and no accuracy rate is claimed.
• Founder Alpha is early-access software; the commentary can be wrong, repeated or badly timed, and it is sold to customers in the United States only.