Agent mode
This is the site as an AI agent reads it.
In short
A lot of reading now starts with a model. So every page here has a plain-text twin, no JavaScript needed, and nothing in it is hidden from the human version.
The “For AI readers” link at the top right of every page brings you here. This page is a list of plain-text files, not a different view of the page you were on; “Human” takes you back to the normal pages.
Start here
Each file is made from the finished pages every time the site is built, so the numbers match what people see.
/llms.txtA short map of the site with a link to each page’s Markdown, in the llms.txt format./llms-full.txtThe whole site as one Markdown document: every page’s text, every table, and every chart’s numbers as a table.
Every page, as Markdown
Charts become tables of the numbers they show. Text behind “show the table” toggles is written out in full.
/index.mdHome: what PlannerBench is, the results in four lines, why I built it, and what it means for you and for firms/results.mdResults: the numbers behind each finding, with a link to the full detail/alone.mdAI on its own: simple questions vs real-life ones, in both tests, and GPT-6 Luna through Codex (the model on my ChatGPT subscription)/calculator.mdWith a calculator: the same models with the planning engine to call/wrong-facts.mdWrong facts: who notices when a client’s facts can’t be right/built.mdHow I built it: the story and the parts, in plain English/method.mdTechnical details: every test and its settings, the GPT-6 Luna run, how the answer key and grading are checked, the statistics, the words used and the limits/labs.mdFor AI labs: what the test set is, what it shows, and why it’s useful for training/firms.mdFor financial firms: what I found, what fixes it, and what a firm could do with it/chat.mdTry it in Claude or ChatGPT: the planning engine as a connector (not public yet)/report.mdReport: the full write-up, with corrections and limits/replay.mdReplays: one calculator replay, in full