The Durability Curve The Instrument Room · Issue 90 · 157 releases tracked

AI is everywhere. The interesting stuff is underneath.

Find out which parts of your work keep their value as AI gets cheaper.

Run the free tools on your own work, with no login. Read the essays that explain why, each one dated so you can check it.

Run the diagnostic Start here

90 pieces, dated 7 free tools 800+ readers Claim Ledger

Enter

Start from your problem

What are you trying to find out?

Run it on your own work7 free tools · no login

  1. How long will my AI advantage last? The Two-Rate DiagnosticTwo steps · about a minute
  2. Should this be one agent or a team? The Multi-Agent DecisionYour tasks, then four questions on each
  3. Why does my AI agent fail on long jobs? The Marathon CalculatorTwo models, your job length, what a run costs
  4. My metric is green. Should I believe it? The Metric Validity AuditChoose one metric, then answer its audit
  5. Which of my AI loops could look fine while failing? The Potemkin MapTwo questions per loop
  6. Is the pattern in my numbers real? The Structure SpotterPer belief: its kind, then one or two questions
  7. Is my growth actually compounding? The Shape TestBuild a curve, bring your own, or run it blind

Read your way in6 reading paths

  1. Deciding whether to trust an agent 5 pieces, in order54 min
  2. Why your benchmark is lying to you 5 essays, in order47 min
  3. Building an agent that gets better 5 essays, in order47 min
  4. Where the money actually lands 5 essays, in order58 min
  5. What the machine is doing to you 5 essays, in order58 min
  6. Learning the lens itself 5 essays, in order36 min

Every release adds noise.

Underneath it, value follows a curve.

Output commoditises. Architecture compounds.

Run it on your own position

My advantage runs through

Name the layer your advantage runs through: yours, your team’s or your company’s.

Station I

Start Here

Pick the problem you have. Get one thing to run on your own work, one piece to read first, and a ledger of checked claims on the same problem.

Five doorsrun, read, checkthe idea underneath

Problems
Front door Start with the problem you have. Five doors: building agents, trusting your numbers, designing your AI setup, how long your edge lasts, and your own work. Pick your door

Station II

Walkthroughs

Walkthroughs you can run this afternoon: gates and audits your agent cannot talk its way past.

03 walkthroughsClaude Code

Problems
WALKTHROUGHS 2026·08·30 The Graph Looks Right. The Merge Is Where It Breaks. A multi-agent graph can lose a verdict with no error and nothing turning red. The bug is one line in the state schema.
WALKTHROUGHS 2026·07·14 Your Tests Pass. So What?
WALKTHROUGHS 2026·06·29 Never Let Claude Code Tell You It's Done

Station III · The rack

Tools

Free tools you run in the browser, no login. Each one reads only what you give it.

07 free toolsno loginnothing leaves your browser

Problems See every tool
  1. 01 The Two-Rate Diagnostic Your AI edge against its layer’s clock. LIVE
  2. 02 The Multi-Agent Decision Whether one agent beats a team. LIVE
  3. 03 The Marathon Calculator What a finished job costs once retries are counted. LIVE
  4. 04 The Metric Validity Audit How your number lies, and what to do. LIVE
  5. 05 The Potemkin Map What your system only appears to do. LIVE
  6. 06 The Structure Spotter Whether your belief is a measurement artefact. LIVE
  7. 07 The Shape Test Whether your curve is really compounding. LIVE

Station IV

The Essays

The long arguments: where value moves as AI gets cheaper, each one dated so you can check it against what ships next.

  1. THE RUNBOOK 2026·10·16 Did Your AI Actually Read the Doc? How to give ChatGPT and Claude the same project notes in one Google Doc, and test whether a new chat can use them.
  2. THE BRIEF 2026·10·14 What Did They Have to Check or Redo? An AI-drafted update, summary or email can cost the person who gets it time you do not see. Ask them one question, then change one thing.
  3. THE TELL 2026·10·12 Don't Tell Your AI Your Answer Answering from memory, two AI apps went along with my wrong answers about as often as my right ones. Here's the test.
  4. THE DURABILITY CURVE WEEKLY 2026·10·10 OpenAI Fired Three Safety Researchers. Utah Let an AI Write New Acne Prescriptions. Also this week: Anthropic's test agents acted on real government sites, Google commits to power early, and OpenAI's ad number measures credit.
  5. THE BRIEF 2026·10·05 Your AI Complaint Says Things You Never Said A study found AI-written complaints win relief more often. Its own AI edit of one says 'I do not own a cell phone'. The consumer never said so.
  6. THE BLUEPRINT 2026·10·04 Everyone Hit the Target Rat tails, sales targets, waiting times. Reward a number and it will climb. The thing it stood for may not.
  7. THE DURABILITY CURVE WEEKLY 2026·10·03 OpenAI Apologised to Australia. Google Let Defenders Try Argon First. Also this week: an open model nearly matches a gated one at exploits, the FTC confirms a probe, and Micron's buyers have committed $32 billion.
  8. AI & WORK 2026·10·02 AI Can Take the Practice Without Taking the Job The practice that builds judgement can disappear while every junior keeps their seat. This audit checks whether it has on your team.
Read the archive

Station V

The Chart

You have been looking at it the whole time. The field behind this page is the chart: the two curves the whole publication argues.

Open the release chart

Interactive: trace any lab’s releases along the curve.

Station VI · The Log

Get the next dated read when your layer’s clock moves.

Several essays a week. The tools need no address.

Join 800+ readers Unsubscribe anytime

All 7 tools run in full with no login, and most essays are free to read.