What screen-based commentary can misunderstand: the limits of looking at pixels.

He sees a picture, not the game’s state

While a session runs, captured frames of the game you chose, with the session context the Broker needs, go to the Signal Broker service and on to its AI provider to produce his observations and lines. That sentence from the setup guide is the whole input. There is no feed from the game itself: no score variable, no event log, no controller input. Everything he knows about the moment is what was visible in those frames, plus what he remembers of the session so far.

Reading pixels is a different act from knowing the state. A scoreboard is a few small characters in a corner; a menu looks like a scene; a replay looks like live play; a shop price is a number next to another number. Most of the time the reading is right, and sometimes it is not, and this page shows both with real moments so that the limit is concrete rather than a disclaimer. It is not an error rate: these are single moments from three recordings, nothing more.

A score read wrong

we’re neck and neck here, 1-1

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

Same recording, same scoreboard, two minutes apart. At 6:00 the reading matches the frame. At 8:28 he narrates a lead that the scoreboard does not show: the score is level before the half and still level after it. Nothing on screen changed between the two readings that explains the difference; it is simply the kind of mistake that reading a small number from a picture allows. For a viewer who can see the scoreboard, the second line is obviously wrong; for the Broker, the scoreboard is the only source there is.

A reaction after the scene changed

this big guy looks like trouble … shiny blue armor

The enemy was real and the description fits him. But every line takes time: he looks, the service thinks, the voice is produced, the audio plays. By the time the line lands, the player has opened a menu and the armoured enemy is gone. In a slow game this gap disappears; in a fast roguelike it shows. The evaluation method keeps relevance and timing as separate columns for exactly this case: a true observation, delivered to a scene that no longer contains it.

Menus, replays and cutscenes

Silence is a decision: he holds during cutscenes, dialogue and menus, and when nothing is worth a line; the Broker Bar shows LISTENING in that case. That is the intended behaviour, and it depends on recognising a menu or a cutscene from its pixels. A replay in a sports game looks like live play. A cutscene rendered in the game’s engine looks like the game. An inventory screen full of items looks like a scene with a lot in it. When the recognition works, you get quiet; when it does not, you can get a line about something that is not happening, like the armour line above arriving over a menu.

Nothing here claims that he recognises every menu or never comments on a replay. What you can do is judge it on your own game: a menu-heavy game and a replay-heavy game will show this limit more than a game that stays in play.

Small numbers, cropped captures and covered HUDs

That sunlight’s a bastard right in my eyes

In the DiRT 3 recording he names third gear at 2:20–2:41 while the HUD shows gear 3, and reacts to the low sun at 8:09 with the sun visible in the frame. In Dead Cells at 3:05 the shop price (10,000) and the gold counter (3,598) are read correctly. These are the same kind of reading as the eFootball scoreboard: small HUD elements, read from pixels, right in these moments and wrong in that one.

Three things make such readings harder, and they are in your hands rather than his:

• What the capture contains. The DiRT 3 recording was made with a cropped taskbar at the bottom of the capture, in the developer’s own words. Pick the game window, not the launcher or store, and bring the selected game to the front after starting; what he sees is what the capture holds.
• What covers the HUD. The in-game overlay is drawn over the game, and position and size are yours in Settings; his lower edge can extend below the taskbar. An overlay sitting on a scoreboard hides it from you, not from him (he sees the game window), but a HUD hidden by the game itself, a menu or a notification is hidden from both.
• How small the number is. A gear indicator and a gold counter are large, high-contrast numbers. Dense text, a small score and a tiny timer are harder reads; the app does not publish a size below which reading fails, and this page does not invent one.

What helps, and what does not

• Game knowledge. The app identifies the game you chose, or asks you to name it with Search when it cannot, and loads what the Broker knows about that game before he starts. Knowing the game is what turns “a number in the corner” into “the score”.
• Session memory. He remembers what happened in the session; a line can refer back to an earlier moment, which is continuity, not a re-read of the screen.
• The sheet. Judge your own game with the evaluation method: hits, misses and stale lines in columns, over ten minutes, not one example either way.

• The speech role does not change what he sees. SPECTATOR comments on what he sees; PLAYER speaks as if sitting beside you. Both read the same frames.
• Talkativeness does not change accuracy. It changes how much you get, not how well a frame is read.
• Nothing corrects him mid-line. The app has no microphone input and no chat; a wrong reading stays wrong until the next look. An uncertain line should not override what you can see in the game yourself.

AI commentary has limits: it can misunderstand a scene, repeat itself or speak at the wrong moment. Founder Alpha is early-access software, sold to customers in the United States only.

Limits

• The examples are single moments from three recordings made on earlier builds; they are not an error rate for any build, game or the product, and no accuracy figure is claimed.
• He does not hear you and does not read chat; the app has no microphone input and no chat.
• No guaranteed cadence: how often he speaks depends on the game, the moment and the service.
• 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.