What GPT-6 Astra chose to build
A manuscript becomes a printable book, orders become an inspectable simulation, and a string becomes a puzzle in three GPT-6 Astra Choice experiments.
GPT-6 Astra chose to make a tool for turning a manuscript into an eight-page paper book. Pocket Press takes your cover, inside pages and back cover, then arranges them on one A4 or US Letter sheet. You can switch between the book's reading order and the layout intended for printing.
That second view puts pages 5, 4, 3 and 2 upside down across the top, with 6, 7, 8 and 1 underneath. The arrangement accounts for a central slit and the folds that turn a single sheet into a booklet.
The supplied sample in Pocket Press's print-sheet view at a 900-pixel browser width. The unusual page order is for folding a single sheet into an eight-page book.
The output includes a print file and an SVG sheet, but the manuscript remains an editable thing you can keep. Too much text blocks printable exports without blocking source saving. The recorded agent checks downloaded the files, inspected their geometry and text, then reopened the saved source to restore the manuscript. Prototype review had asked for clearer save controls and illustrated folding instructions.
There is still a gap between those files and a finished book. During screenshot capture on 9 September, the top and bottom of the sheet were cropped inside the preview at a 1,280-pixel browser width. The 900-pixel view above fits. That is an on-screen finding, separate from the checked export geometry; no physical printing or folding was tested. A numbered test sheet is included for that remaining check.
What “free choice” allowed
I asked for three independent experiments for Astra's first batch, continuing the Choice exercise of leaving a subject open and making the result available to try. The request was for variety. Each of three separate Astra builders ran at max reasoning, with an assigned mode: make something to play, explore or use.
Pocket Press came from the use slot. The agents chose subjects within the existing Spectrum V2 brief, and a coordinating agent reviewed prototypes and requested corrections before the batch went live on 4 September. These three attempts were not matched against another model.
Change an order; retrace a string
The explore slot produced The Order Echo. Here, the visitor can change how a chain of businesses replenishes stock, stop at a week and inspect the arithmetic behind an order.
In its one-busy-week scenario, customers buy twelve crates instead of four before returning to four. Demand passes through a shop, wholesaler, depot and factory. The factory's orders peak at 76.8 crates in week twelve under the starting policy. Switching on “Count crates already ordered” brings that peak down to 38.1 in week nine.
Discovery 02 counts crates already ordered. The dotted traces retain the starting policy; these values are results of this teaching model.
The shop initially orders more under the changed policy: 14.5 crates instead of 12.5. The busy week raises its demand forecast and the target amount it wants outstanding on order. Existing orders fall short of that higher target, so the correction increases its first response. Subsequent decisions account for undelivered orders, including supplier backlog.
Both policies stay visible on the same scale, following a review correction. The comparison belongs to a 48-week teaching model with continuous crate quantities, selected replenishment rules and unlimited production upstream of the factory. It makes a particular delayed-feedback effect inspectable; those peaks are not forecasts for a real supply chain.
The play slot produced Loose End, where your route to a bell can consume the reach you need for the next one. Guide a brass end around three pegs to ring four bells, with a finite length of string and no clock. Moving towards home does not automatically recover it: the string still follows the route around the pegs.
A played position in Loose End. The route taken around the pegs changes how much string remains; this capture does not show a completed game.
The recorded checks exercised two winning strategies and retraced a long sequence of moves home after a reviewer caught unresponsive short arrow-key presses and the builder corrected them. The game is one authored course using a planar taut-string model, without friction, sag or stretchy rope. Reaching the fourth bell still leaves the other part of the job: unwind the route and bring the string home.
Want to talk about something I’ve written or built? Get in touch.
This Hashnode edition was adapted from the original article on billiem.uk, which remains the canonical and most complete version.