---
title: "What GPT-6.1 Sol chose to build"
date: 2026-10-01
description: "Three autonomous GPT-6.1 Sol experiments: a fictional record, a sheet-cutting planner and a congestion model, with their review fixes and limits."
tags:
  - "ai-coding"
  - "ai-experiments"
  - "llm-agents"
---

GPT-6.1 Sol chose to build a record that loses its meaning, a cutting plan for leftover sheet material, and a road that makes everyone's journey longer. Each makes a sentence, a cut or a route choice inspectable.

I wanted to repeat the [Choice exercise](/posts/llm-choice-as-an-idea-engine/): let a new model choose what to make, then look at the outputs and how their style differs. This batch used GPT-6.1 Sol at Extra High reasoning. The coordinating agent set three levels of ambition, from a compact piece to a flagship. One fresh builder chose all three subjects and implemented the first two; a second fresh Sol builder took the flagship before its public code existed.

That configuration matters when comparing it with [September's Astra experiments](/posts/gpt-6-astra-choice/), which used three separately assigned builders at Max. These are different subjects, briefs and working conditions, rather than matched tasks.

## The Minutes: what a sentence leaves out

[The Minutes](/choice/experiments/2026-10-01-sol61-xhigh-jewel/outputs/minutes) follows an invented exhibition incident through six authored records. Select a sentence and its earlier wording appears beside it. You can compare drafts and recover facts that have disappeared.

The final statement says, “The issue was managed promptly.” Its trace leads back to a request to isolate the power, refused because the display lighting had to stay on. The reassuring sentence has replaced an action and a named person with an assurance that the first account does not establish.

![The earlier isolation request and refusal are struck through above the introduced reassurance, ‘The issue was managed promptly.’](media/minutes-trace-092380efbae2-b89917041143-349.png)

*The final record compared with the first account in The Minutes. This ancestry belongs to an explicitly authored fictional incident.*

It also includes an edit that improves accuracy: an uncertain recollection of when the pump stopped becomes a time supported by the controller log. That keeps the argument more precise than simply treating every revision as concealment.

Review found that advancing the records began below the visible page. Navigation moved above the document, and selecting a sentence on mobile now brings its provenance into view with a route back. This remains a fixed fictional argument, not a tool for detecting dishonesty in arbitrary documents.

## Offcut: a layout you can actually cut

[Offcut](/choice/experiments/2026-10-01-sol61-xhigh-substantial/outputs/offcut) accepts sheet sizes, rectangular parts, quantities, grain constraints and blade width. It returns a layout and an ordered sequence of full cuts through the remaining rectangles. Stepping through that sequence shows pieces becoming available as they are released.

The supplied example fits nine parts onto a 1,220 × 610 mm sheet. Blade loss is part of the calculation: two 50 × 100 mm pieces fit a 100 × 100 mm sheet with a zero-width blade, but only one fits when the blade removes 3 mm.

![A 1,220 by 610 mm sheet contains three labelled shelves, two sides, smaller pieces and hatched remainder areas, with grain running horizontally.](media/offcut-plan-e64d4cd26f34-ae6f1fda1bf7-768.png)

*Offcut's supplied nine-part layout. The diagram accounts for blade loss and grain; no physical cutting was tested.*

You can keep a dimensioned SVG, a CSV cut order, editable JSON and remaining-stock inventory. Review caught an ambiguity in the CSV cut order: two copies of identically named stock were indistinguishable. Physical sheet IDs now carry that distinction. Mobile calculation feedback also needed work because the result appeared offscreen.

Its search is bounded. A complete plan is checked against the cutting model, but an unplaced part does not prove that no better arrangement exists. Damaged edges, trimming allowance, clamping and tool clearance are outside the model.

## The Shortcut Tax: the road and the incentives

[The Shortcut Tax](/choice/experiments/2026-10-01-sol61-xhigh-flagship/outputs/shortcut) opens on a small directed road network. At the default west demand of six flow units, closing the shortcut gives an average journey of 80 minutes. Open it under individual route choice and that rises to 100. No driver can improve their own trip by switching routes.

![The open shortcut carries four flow units, and average travel is 100 minutes beside an 80-minute outer-roads comparison. Individual route choice is selected.](media/shortcut-model-d8d6ad4323db-e27a43a08298-388.png)

*The default six-unit demand with the shortcut open. These are results of the teaching network, not measured traffic times.*

Coordinated routing brings the average back to 80. Pricing the delay each driver imposes on others can reproduce that result too; the tolls are expressed in equivalent minutes, not money. Change demand and the shortcut can help, hurt or cease to be used.

The early prototype exposed its calculations as raw JSON. Review pushed it towards a visible network, road and route costs, a demand sweep, and a ledger explaining the tolls. A second origin lets you inspect who gains or loses when another neighbourhood shares the bottleneck. The network still assumes fixed demand, instantaneous equilibrium and simple congestion costs, with no queues or real-road calibration.

## A familiar visual language

These pieces share quite a lot with the Astra batch: warm paper, serif headings, muted green or blue ink, and rust accents. Their layouts follow their tasks. The Minutes pairs a document with margin notes; Offcut centres a dimensioned sheet; The Shortcut Tax pairs a network with its current outcome, then exposes deeper controls below.

The subjects have changed more visibly than that palette. Astra's string puzzle, supply-chain model and folded booklet have become close reading, material planning and congestion. Three freely chosen pieces cannot establish a model's taste, especially when review shaped the navigation, exports and presentation. Those interventions are part of what the reader sees, alongside the choices Sol made before anyone asked it to polish them.