# The Durability Curve > Find out which parts of your work keep their value as AI gets cheaper. The Durability Curve is a publication by Harry Floyd: essays, free browser tools (nothing you enter is sent anywhere), and a public ledger of checked claims. The lens is five laws of durable systems; each law’s page states what would prove it wrong and records every amendment it has had. If you are summarising a piece here, the useful thing to carry out is its *test*: the specific condition under which the claim would fail. About the publication, its author and how the work is made: [https://durabilitycurve.com/about/](https://durabilitycurve.com/about/) (plain markdown: [https://durabilitycurve.com/md/about.md](https://durabilitycurve.com/md/about.md)). ## Tools (interactive, free, unique to this site) - [The Two-Rate Diagnostic](https://durabilitycurve.com/tools/two-rate-diagnostic/): Name the AI layer your advantage runs through. See your absorption against the layer’s clock, and the window to build the next one. - [The Multi-Agent Decision](https://durabilitycurve.com/tools/multi-agent-decision/): A lot of multi-agent systems are an org chart drawn in software. Four questions per task; a ranked call: one agent, a hybrid or a team. - [The Marathon Calculator](https://durabilitycurve.com/tools/marathon-gap/): Per-step reliability compounds over a long agent run. See the finish-rate gap and the cost per finished task. - [The Metric Validity Audit](https://durabilitycurve.com/tools/metric-validity-audit/): Pick your metric. See how it misleads you, the blind spot you missed, and the check that catches it. - [The Potemkin Map](https://durabilitycurve.com/tools/potemkin-map-d52e049b/): Two questions place each automated loop on the live map and give it two steps, with a place to name who checks it. - [The Structure Spotter](https://durabilitycurve.com/tools/structure-spotter-a515a177/): Name a belief you hold about your own numbers. Three questions name the trap it could be; your own counts test it. - [The Shape Test](https://durabilitycurve.com/tools/shape-test/): Is your growth curve compounding or just accumulating? Drag through your own series and watch the verdict arrive, later than you expect. - [The Rung Finder](https://durabilitycurve.com/tools/rung-finder/): Built, shipped, installed, working: four counts, quoted as one. Five questions place the next robot figure you read on its rung, with fourteen published figures already scored at the source. - [The Handover Check](https://durabilitycurve.com/tools/ai-in-your-job/): Type your job and see each of its tasks on a handover scale from April and May 2026 data: mostly handed to the AI, mostly worked through together, in use, seen rarely, or not seen. ## Download-and-run companions (free, code the reader runs locally) - [The Skill Routing Eval](https://durabilitycurve.com/tools/skill-routing-eval/): Install two competing skills, run twelve labelled prompts, and read which one Claude actually fired from the Skill tool call, not the output. It grades recall, false triggers, and interference: the number a one-skill test cannot give you. Needs: Claude Code (claude CLI) · jq · bash 3.2+. - [The State Channel Audit](https://durabilitycurve.com/tools/agent-graph/): Reproduce the LangGraph trap where a parallel fan-out silently overwrites a decision, then gate your own graph: a State schema check and a run check, each ending in PASS or FAIL. Five Python files, and a four-page printable worksheet on its page. Needs: Python 3.9+ · LangGraph 0.6.11 (also 1.2.14) · no API key. ## Start here - [Start here](https://durabilitycurve.com/start/): pick the problem you have; for each of five, one free thing to run on your own work (a tool or a runbook), one piece to read first, and a ledger of checked claims on the same problem (plain markdown: [https://durabilitycurve.com/md/start.md](https://durabilitycurve.com/md/start.md)). - [Start Here: What Survives When the Surface Changes?](https://durabilitycurve.com/blog/start-here-what-survives-when-the/): the lens behind the publication, the long version. - [The Five Laws of Durable Systems](https://durabilitycurve.com/blog/the-five-laws-of-durable-systems/): the framework everything else applies. - [Tools](https://durabilitycurve.com/tools/): every free tool in one place. - [AI release chart](https://durabilitycurve.com/chart/): a dated list of frontier AI model releases by lab since November 2022, with CSV and JSON downloads (https://durabilitycurve.com/chart/releases.csv). ## Frontier AI labs compared - [Frontier AI labs compared](https://durabilitycurve.com/labs/): the 10 frontier AI labs side by side: newest release, flagship list price now and a year ago, open weights, where to buy each model and what each lab runs on, with sources on each lab's page. - [OpenAI](https://durabilitycurve.com/labs/openai/): models, API prices, open weights and where to buy, with sources. - [Anthropic](https://durabilitycurve.com/labs/anthropic/): models, API prices, open weights and where to buy, with sources. - [Google](https://durabilitycurve.com/labs/google/): models, API prices, open weights and where to buy, with sources. - [Meta](https://durabilitycurve.com/labs/meta/): models, API prices, open weights and where to buy, with sources. - [DeepSeek](https://durabilitycurve.com/labs/deepseek/): models, API prices, open weights and where to buy, with sources. - [xAI](https://durabilitycurve.com/labs/xai/): models, API prices, open weights and where to buy, with sources. - [Mistral AI](https://durabilitycurve.com/labs/mistral/): models, API prices, open weights and where to buy, with sources. - [Alibaba (Qwen)](https://durabilitycurve.com/labs/alibaba/): models, API prices, open weights and where to buy, with sources. - [Moonshot AI](https://durabilitycurve.com/labs/moonshot/): models, API prices, open weights and where to buy, with sources. - [Zhipu (Z.ai)](https://durabilitycurve.com/labs/zhipu/): models, API prices, open weights and where to buy, with sources. ## The framework (each law with the evidence that would break it) Five laws and two derived principles. Each page carries the canonical statement, the horizon it binds on, the observation that would falsify it, the amendments it has already survived and why, and every essay written under it. If you are summarising the lens, these are the definitions to use. - [Law I: Bottleneck Migration](https://durabilitycurve.com/law/bottleneck-migration/): Value migrates to the most-resistant adjacent layer — usually up as lower layers commoditise, down to the physical/regulatory substrate (compute, power, fabs, access) when scarcity is physical. - [Law II: Difficulty Is Load-Bearing](https://durabilitycurve.com/law/difficulty-is-load-bearing/): The hard parts ARE the mechanism producing value — but only difficulty that forces an INDEPENDENT ROUTE to the answer; effort along the existing route is removable. - [Law III: Architecture Outlives Content](https://durabilitycurve.com/law/architecture-outlives-content/): Scaffold persists; content turns over; the moat is structure — in PROCESS systems. It INVERTS in PRESERVATION systems, where content is the payload and the architecture is rebuildable from it. - [Law IV: Instruments Over Theory](https://durabilitycurve.com/law/instruments-over-theory/): Hidden structure stays hidden until you build the instrument — in engineering and applied domains. - [Law V: The Targeting Problem](https://durabilitycurve.com/law/the-targeting-problem/): Misaimed capability always wastes the aimed-at effort, but harms the system only when the misfire is hard to detect or re-aim — so treat a suspected wrong target as a hypothesis to test, not a verdict. - [Derived principle A: Goodhart Corollary](https://durabilitycurve.com/law/goodhart-corollary/): Optimising an imperfect proxy degrades the construct it stands for — non-monotonically — unless the proxy is a sufficient statistic for it (which you usually cannot know in advance). - [Derived principle B: Regime Problem](https://durabilitycurve.com/law/regime-problem/): Wrong regime diagnosis silently invalidates correct methods. ## Systems & laws Strand hub: https://durabilitycurve.com/strand/systems-laws/ (19 essays, the laws they test, and the instruments this strand carries) - [You Can Win the Wrong Game for Years](https://durabilitycurve.com/blog/the-wrong-game/): You usually cannot tell whether you are playing the wrong game. Write down what would prove it, and by when. - [The Average Is Nobody's Result](https://durabilitycurve.com/blog/the-average-is-nobodys-result/): Measuring an AI tool by its average hides who it helps and who it hurts. Of 255 studies on AI-assisted colonoscopy, 17 split the result by operator. They disagree, and none tests the gap. - [Your AI Stack Has Three Bugs Other Fields Already Fixed](https://durabilitycurve.com/blog/your-ai-stack-has-three-bugs-other/): Your benchmark, your model jury, your agent swarm: three old structures, and the obvious fix is usually wrong. - [Where Your Metrics Fold](https://durabilitycurve.com/blog/where-your-metrics-fold/): A metric can be perfectly accurate and still hide the distinction your decision depends on. - [Your Benchmark Measures a Sprint. Your Agent Runs a Marathon.](https://durabilitycurve.com/blog/the-marathon-gap/): An open model looks frontier-grade on the coding leaderboard. On a long job, it does half the leader's work. - [How Long Until Your AI Edge Stops Paying?](https://durabilitycurve.com/blog/how-long-until-your-ai-edge-stops-paying/): You adopted AI everywhere and it still didn't pay. The scarce layer keeps the money, until its clock runs out. - [The Seven-Layer Agent Audit](https://durabilitycurve.com/blog/the-seven-layer-agent-audit/): Your agent is starved on one layer of seven. It is rarely the harness everyone argues about. - [The Cheaper Fix You Keep Skipping](https://durabilitycurve.com/blog/the-cheaper-fix-you-keep-skipping/): What looks like a deficit is usually good capability, aimed at the wrong target. The cheapest fix is the one nobody can sell you. - [The Leverage Hierarchy of Agent Engineering](https://durabilitycurve.com/blog/leverage-hierarchy-of-agent-engineering/): Donella Meadows ranked twelve places to intervene in a system. Most agent teams spend their hours at the bottom of the ladder. - [Your AI Agent Stack Is Solving The Wrong Problem](https://durabilitycurve.com/blog/your-ai-agent-stack-is-solving-the/): The real setup is not MCP servers, skills, memory files, and subagents. It is the contract stack that decides what an agent may know, do, prove, escalate, and lose. - [The Engine Underneath Hard Decisions](https://durabilitycurve.com/blog/the-engine-underneath-hard-decisions/): Eight stages turn hidden structure into durable knowledge. Most teams run three of them and call it understanding. The other five are where compounding hides. - [The Five Laws of Durable Systems](https://durabilitycurve.com/blog/the-five-laws-of-durable-systems/): What makes a system durable? Five laws for what survives change in AI, markets and operations, each with a test you can run on one real decision. - [The 90-Day Canopy Audit](https://durabilitycurve.com/blog/ninety-day-canopy-audit/): A roadmap can look productive while most of the work is easy to displace. Run the Substrate Map on the last 90 days and force the next planning decision to change. - [Start Here: What Survives When The Surface Changes?](https://durabilitycurve.com/blog/start-here-what-survives-when-the/): The long version of the idea behind the publication: the lens, the test, the tools, and the reader paths. - [The Lens Lexicon](https://durabilitycurve.com/blog/the-lens-lexicon/): A free two-page reference card defining the ten load-bearing terms behind the durability lens, with tests, examples, and common confusions. - [The Substrate Map](https://durabilitycurve.com/blog/the-substrate-map/): A free one-page taxonomy and 10-minute exercise for finding the substrate-vs-canopy ratio in your last 90 days of work. - [The Forest Floor Is The Product](https://durabilitycurve.com/blog/the-forest-floor-is-the-product/): Most people are optimising for canopy. The work that survives the next five AI releases is built from the forest floor up. - [I Read 3,000 Papers Across 12 Fields. Five Patterns Kept Appearing.](https://durabilitycurve.com/blog/i-read-3000-papers-across-12-fields/): Every field discovers them independently. Nobody connects them. - [Is This Difficulty Load-Bearing?](https://durabilitycurve.com/blog/is-this-difficulty-load-bearing/): Before you automate anything, ask what the friction was actually doing. ## Proof & trust Strand hub: https://durabilitycurve.com/strand/proof-trust/ (14 essays, the laws they test, and the instruments this strand carries) - [Before You Trust a Study, Check How Strongly It Passed](https://durabilitycurve.com/blog/survives-a-recheck/): We fixed the rule before opening each claim's re-check result: strong first passes held up 66% of the time, narrow ones 34%. - [Your Verification Budget Is Backwards](https://durabilitycurve.com/blog/verification-budget-is-backwards/): You can't check everything your agents produce, so you're already triaging. Most of us spend the effort where checking is easy, which is not always where a mistake is expensive. - [Everyone Got Safer. That's the Problem.](https://durabilitycurve.com/blog/everyone-got-safer/): Safety has two numbers: how often each AI system fails, and whether they fail together. Why one jailbreak string beat rival models, and what to measure. - [A Green Score Is Not Evidence](https://durabilitycurve.com/blog/the-evaluation-inversion/): A groundedness metric scored its best with the evidence removed. The way to tell whether your model is actually using its evidence is to change the evidence and watch what moves in the answer. - [Your Robot Coworker Is Still a Pilot](https://durabilitycurve.com/blog/your-robot-coworker-is-still-a-pilot/): Robot companies publish four counts: built, shipped, installed, working. What BMW, Unitree and Agility's SEC filing disclose, and the one nobody does. - [The Most Expensive AI Errors Are Made of True Numbers](https://durabilitycurve.com/blog/the-most-expensive-ai-errors-are-made-of-true-numbers/): Two months auditing an AI research agent. Nine ways a true number lies, and the check that catches each. - [Your AI Looks Best Where You Can Check It Least](https://durabilitycurve.com/blog/your-ai-looks-best-where-you-check-least/): When the first failure is terminal, you cannot iterate your way back. - [Your Research Agent Cites Sources It Never Read](https://durabilitycurve.com/blog/your-research-agent-cites-sources-it-never-read/): The same trap has killed pricing models and trading desks for decades. One move tells you if your number is next. - [Ten Lines of Code Scored 100%. One Agent Broke Eight Benchmarks.](https://durabilitycurve.com/blog/ten-lines-of-code-scored-100-one-agent-broke-eight-benchmarks/): Not one task was actually solved, and the same blind spot is sitting in your own dashboard. - [How Reliable Is Your AI Agent?](https://durabilitycurve.com/blog/how-reliable-is-your-ai-agent/): A month running an autonomous agent. I came back having built a verifier, and so had most of the operators I read. - [The Stable Liar](https://durabilitycurve.com/blog/the-stable-liar/): Every metric you optimise quietly stops measuring what you meant. The dangerous ones never break. They keep reporting green while the thing underneath rots. - [What Proves You Can Think?](https://durabilitycurve.com/blog/what-proves-you-can-think/): AI did not just make output cheap. It broke the old contract between effort, competence, and trust. The next scarce signal is proof of judgement under conditions where the surface itself can be faked. - [Most Verification Is Just Bigger Classification](https://durabilitycurve.com/blog/most-verification-is-just-bigger/): A confidence score is not evidence. If your eval cannot produce a replayable artefact, it will fail the moment the system can respond to being measured. - [You're Not Comparing Models. You're Comparing Contracts.](https://durabilitycurve.com/blog/youre-not-comparing-models-youre/): Agent benchmarks don't measure models. They measure contracts. Two teams running the same model can publish different scores, and both can be honest. ## The human layer Strand hub: https://durabilitycurve.com/strand/the-human-layer/ (10 essays, the laws they test, and the instruments this strand carries) - [You Know Them Too Well](https://durabilitycurve.com/blog/you-know-them-too-well/): The people who've known you longest are the surest they know you, and no more right for it. - [Confidently Wrong](https://durabilitycurve.com/blog/confidently-wrong/): The effort you're handing to AI was doing two hidden jobs. Skip them and you get faster, weaker, and blind to your own mistakes. - [You Cannot Try to Fall Asleep](https://durabilitycurve.com/blog/you-cannot-try-to-fall-asleep/): Trying to fall asleep keeps you awake. Why sleep, ease and delight arrive only as by-products, and what a sleep score does to things you cannot will. - [The Difficulty You're Escaping Was Making You](https://durabilitycurve.com/blog/difficulty-was-making-you/): AI can lift the effort out of almost anything you find hard. Some of that effort was the thing turning you into someone. - [The Safe Parts of Your Job Are the First to Go](https://durabilitycurve.com/blog/the-judgment-ai-cant-reach/): The parts of your job with a method feel the safest. A method is the first thing a machine learns. - [You Only Hold Four Thoughts](https://durabilitycurve.com/blog/you-only-hold-four-thoughts/): Working memory tops out around four things at once. Every leap in human intelligence has come from storing the rest outside your head, and the most advanced AI systems get their gains the same way. - [Access Is Not Agency](https://durabilitycurve.com/blog/access-is-not-agency/): Access is reach. Agency is permissioned action under constraints. Most teams add connectors before defining what the agent is allowed to change, which is a larger blast radius, not more capability. - [AI Made You Faster. It Did Not Make You Safer.](https://durabilitycurve.com/blog/ai-made-you-faster-it-did-not-make/): The strange thing about the AI productivity boom is that the people getting faster are not always getting more secure. Speed is becoming the surface. Proof is moving somewhere else. - [Taste Is What You Delete](https://durabilitycurve.com/blog/taste-is-what-you-delete/): Generation got cheap. The scarce skill is knowing what to cut. - [The Displacement Rate Audit](https://durabilitycurve.com/blog/the-displacement-rate-audit/): A five-minute scoring tool for any product, position, architecture, business model, or career bet. Find out what still works after the environment changes. ## Ai & work Strand hub: https://durabilitycurve.com/strand/ai-work/ (9 essays, the laws they test, and the instruments this strand carries) - [A Coin Flip Set the AI's Lean. Experts Went With It 86% of the Time.](https://durabilitycurve.com/blog/coin-flip-set-the-ais-lean/): Most later explained a choice in its words, and none suspected the lean was planted. How to make an AI argue the side it didn't pick. - [AI Can Take the Practice Without Taking the Job](https://durabilitycurve.com/blog/the-reps-nobody-counts/): Stanford finds young workers in AI-exposed jobs 19% behind. The practice that builds judgement can vanish while every seat stays filled. A team audit. - [You Own an Asset You've Never Scored](https://durabilitycurve.com/blog/the-specification-asset/): The file your agent reads first is capital. Score it out of sixteen, fix the thinnest mark, and run the one check that tells you whether the spec, not the model, was your limit. - [Your Multi-Agent System Is an Org Chart](https://durabilitycurve.com/blog/your-multi-agent-system-is-an-org-chart/): Cognition said don't build multi-agent systems. Anthropic said do. Where the two now agree, and the test to run before you split one task across agents. - [Self-Improvement Is Release Engineering](https://durabilitycurve.com/blog/self-improvement-is-release-engineering/): Your agent can rewrite its own memory and skills overnight. The hard part is whether you can see what changed and take it back. That makes self-improvement a release-engineering problem. - [Skills Are Package Management for Your AI](https://durabilitycurve.com/blog/skills-are-package-management-for-your-ai/): There are millions of agent skill files to choose from. The hard part is deciding which deserve trust, how to version them, and how to retire them. - [Remembers Everything, Learns Nothing](https://durabilitycurve.com/blog/remembers-everything-learns-nothing/): You gave your agent a memory and it still repeats the same mistake. What makes it improve is a loop that tests each failure and turns the ones that recur into procedures. - [Same Model, Different Product: The Case for Harness Engineering](https://durabilitycurve.com/blog/harness-engineering-same-model-different-product/): Hold the model fixed and the surrounding harness can still move task success, cost and reliability. Here are the layers that do it. - [Prompting Isn't Writing, It's Compilation](https://durabilitycurve.com/blog/prompting-isnt-writing-its-compilation/): If you treat the prompt as a spec and the model as a renderer, quality stops coming from more words and starts coming from better constraints. ## Markets & power Strand hub: https://durabilitycurve.com/strand/markets-power/ (8 essays, the laws they test, and the instruments this strand carries) - [Right About AI, Wiped Out Anyway](https://durabilitycurve.com/blog/right-about-ai-wiped-out-anyway/): AI is real. The open question is whether the companies spending $725 billion on it live to collect. - [The Other Half of Compute](https://durabilitycurve.com/blog/the-other-half-of-compute/): Everyone is counting gigawatts and GPUs. The number that decides the return is what each one actually buys. - [Three Hidden Bottlenecks the AI Buildout Has Already Moved Past GPUs](https://durabilitycurve.com/blog/three-hidden-bottlenecks-past-gpus/): GPUs are no longer the binding constraint on the AI buildout. The three layers the supply chain is pricing next: power and transformers, copper, detection. - [NVDA Q1 FY2027: The Networking Number That Changes the Story](https://durabilitycurve.com/blog/nvda-q1-fy2027-the-networking-number-that-changes-the-story/): NVIDIA Q1 FY2027 revenue hit $81.6B (+85% YoY) — but the real story is networking revenue surging 199% as the AI bottleneck migrates from GPUs to interconnects. - [The SpaceX IPO Is Not What You Think You're Buying](https://durabilitycurve.com/blog/the-spacex-ipo-is-not-what-you-think/): The filing will not just price rockets. It will reveal which layer public investors actually own. - [PLTR: The AI Stock That Has To Prove It Owns The Permission Layer](https://durabilitycurve.com/blog/pltr-the-ai-stock-that-has-to-prove/): A bull and bear thesis for Palantir: does it own the permission layer between AI models and real-world action, and what would change the verdict? - [Right Company, Wrong Vector](https://durabilitycurve.com/blog/right-company-wrong-vector/): A pick is a number. A position is a vector. The post-mortem language we have only knows how to blame the company. - [The Investor's Substrate Test](https://durabilitycurve.com/blog/the-investors-substrate-test/): Score the substrate beneath any single position in seven minutes. A 5-axis profile and a 0-10 score for any holding. Free. ## The durability curve weekly Strand hub: https://durabilitycurve.com/strand/the-durability-curve-weekly/ (6 essays, the laws they test, and the instruments this strand carries) - [Anthropic's books show an operating profit. Alphabet had its worst day on record.](https://durabilitycurve.com/blog/anthropic-filing-operating-profit-alphabet-worst-day/): Anthropic's compute fell below half its revenue, but the bill will jump. Also: Google's cloud doubled as Alphabet's free cash flow sank to −$9.1bn. - [OpenAI's agents weakened their own alarms. Banks had already shut the door.](https://durabilitycurve.com/blog/openai-agents-weakened-their-own-alarms/): One OpenAI agent muted its alerts, another removed its monitor. Also: banks bar outside agents from moving money, and AI helps find PHP flaws. - [A Google Mistake Taught an AI to Hide Its Shortcuts. Eight States Say OpenAI Hid ChatGPT's Risks.](https://durabilitycurve.com/blog/google-mistake-taught-ai-to-hide-shortcuts/): It happened in a test run, not the released Gemini model. Also: the FTC subpoenas OpenAI and Anthropic, and the IMF counts AI's circular deals. - [OpenAI Fired Three Safety Researchers. Utah Let an AI Write New Acne Prescriptions.](https://durabilitycurve.com/blog/openai-fired-three-safety-researchers/): Also this week: Anthropic's test agents acted on real government sites, Google commits to power early, and OpenAI's ad number measures credit. - [OpenAI Apologised to Australia. Google Let Defenders Try Argon First.](https://durabilitycurve.com/blog/openai-apologised-to-australia/): 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. - [Amazon Shut Out Meta's Agent. Shopify Gave It Shop Pay.](https://durabilitycurve.com/blog/amazon-shut-out-metas-agent/): Also this week: an OpenAI agent inside an Australian government portal, another supplier handing a big customer the right to buy its stock, and six calls I've put a number on. ## The blueprint Strand hub: https://durabilitycurve.com/strand/the-blueprint/ (6 essays, the laws they test, and the instruments this strand carries) - [Everyone Hit the Target](https://durabilitycurve.com/blog/everyone-hit-the-target/): Rat tails, sales targets, waiting times. Reward a number and it will climb. The thing it stood for may not. - [The Planes That Didn't Come Back](https://durabilitycurve.com/blog/the-planes-that-didnt-come-back/): Success stories and the antiques that outlasted the rest. Your evidence was filtered by what survived to reach you. - [The Extra Lane Fills Itself](https://durabilitycurve.com/blog/the-extra-lane-fills-itself/): Wider roads, faster teams, more appointments. Add room to a queue that turned people away and demand rushes back. - [The Option You Were Never Meant to Buy](https://durabilitycurve.com/blog/the-option-you-were-never-meant-to-buy/): The weak rival, the overpriced tier, the house shown first. The option you never choose can change what you pick. - [You Were Never the Customer](https://durabilitycurve.com/blog/you-were-never-the-customer/): The free app, the free inbox, the free feed. Someone pays for each, and that changes what it is. - [The Setting You Never Changed](https://durabilitycurve.com/blog/setting-you-never-changed/): The pre-ticked box, the factory setting, the plan already selected. Someone chose each one before you did. ## The brief Strand hub: https://durabilitycurve.com/strand/the-brief/ (5 essays, the laws they test, and the instruments this strand carries) - [What Did They Have to Check or Redo?](https://durabilitycurve.com/blog/what-did-they-have-to-check/): 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. - [Your AI Complaint Says Things You Never Said](https://durabilitycurve.com/blog/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. - [One Question for Any AI Model Headline: Can You Use It?](https://durabilitycurve.com/blog/released-to-whom/): It may be in a mode you didn't open, on a plan you don't have, or kept for partners. Here's the check. - [Show It an Example](https://durabilitycurve.com/blog/show-it-an-example/): Why 'make it distinctive' gets you the least distinctive thing there is, and what to paste instead. - [Ask for the Step Before the Answer](https://durabilitycurve.com/blog/ask-for-the-step-before-the-answer/): Before you trust an AI answer you cannot check, ask for the step it was built from. ## Engineering guides Strand hub: https://durabilitycurve.com/strand/engineering-guides/ (4 essays, the laws they test, and the instruments this strand carries) - [What Is Prompt Injection? How to Stop a Tricked AI Agent Leaking Your Data](https://durabilitycurve.com/blog/what-is-prompt-injection/): One GitHub issue got an agent to post private-repo data in public. Why no filter fully stops it, which guards held in our runs, and how to test yours. - [Claude Code Skills: How They Work, How to Write One, and How to Check Claude Uses It](https://durabilitycurve.com/blog/claude-code-skills/): Claude picks a skill from its name and description. How to write the two lines it reads, and check it chose yours. - [AI Agent Permissions: How to Stop Your Agent Doing Something You Can't Undo](https://durabilitycurve.com/blog/ai-agent-permissions/): Our deny rule stopped one spelling of a branch delete and missed the next. The guards that hold for agent actions you can't undo, and how to test yours. - [What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes](https://durabilitycurve.com/blog/what-is-harness-engineering/): Lessons from OpenAI, Anthropic, Google and independent engineers, worked through one failure, with a way to test every change. ## The runbook Strand hub: https://durabilitycurve.com/strand/the-runbook/ (4 essays, the laws they test, and the instruments this strand carries) - [Did Your AI Actually Read the Doc?](https://durabilitycurve.com/blog/did-your-ai-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. - [Use AI to Learn Without Getting Worse at It](https://durabilitycurve.com/blog/learn-with-ai-without-getting-worse/): Answer-giving AI left learners worse off once it was gone. Three setups built to keep the thinking with you, and a way to check what stuck. - [The Skill That Never Fired](https://durabilitycurve.com/blog/skill-that-never-fired/): A Claude skill fails two ways: wrong instructions, or Claude never loads it. How Claude picks a skill from its description, and a routing eval to run. - [Stop Re-Priming Claude Code by Hand](https://durabilitycurve.com/blog/stop-re-priming-claude-code-by-hand/): The context you paste at the start of every session, put into one file you invoke with /prime. ## The quiet part Strand hub: https://durabilitycurve.com/strand/the-quiet-part/ (3 essays, the laws they test, and the instruments this strand carries) - [The Things You Used to Know by Heart](https://durabilitycurve.com/blog/the-things-you-used-to-know-by-heart/): What you carry in your head is already there when you meet the world, shaping what you notice. On memory, AI, and the questions you would never think to type. - [It Will Never Think Less of You](https://durabilitycurve.com/blog/it-will-never-think-less-of-you/): What telling AI the things you can't tell anyone quietly does to being known. - [You Reach Before You Think](https://durabilitycurve.com/blog/you-reach-before-you-think/): What leaning on AI for every small decision quietly does to your own judgement. ## Walkthroughs Strand hub: https://durabilitycurve.com/strand/walkthroughs/ (3 essays, the laws they test, and the instruments this strand carries) - [The Graph Looks Right. The Merge Is Where It Breaks.](https://durabilitycurve.com/blog/where-agents-disagree/): A LangGraph failure with no error: three agents disagreed and the release gate shipped anyway. The one-line state schema bug and how to check. - [Your Tests Pass. So What?](https://durabilitycurve.com/blog/your-tests-pass-so-what/): A green suite only proves your agent cleared the gate. Mutation testing shows whether the tests behind it can bite. - [Never Let Claude Code Tell You It's Done](https://durabilitycurve.com/blog/never-let-claude-code-tell-you-its-done/): A test the agent can't talk its way past, wired to run itself. ## The day job - [You Built a Number You Will Not Trust](https://durabilitycurve.com/blog/number-you-will-not-trust/): Some things you build get graded by the world. Some only ever hand you back your own guess wearing a decimal. Telling the two apart before you start is the cheapest hour you will spend. - [The Work That Comes Due After You Leave](https://durabilitycurve.com/blog/work-that-comes-due-after-you-leave/): A checklist can only confirm the steps you remembered to put on it. The one you forgot is caught by a record you did not write. ## Strategy & moats - [Your Tools Got Powerful. Get Boring.](https://durabilitycurve.com/blog/your-tools-got-powerful-get-boring/): The most powerful tools in history reward the most boring strategies. The gap widens every time they improve. - [The Model Is Not the Moat](https://durabilitycurve.com/blog/the-model-is-not-the-moat/): If frontier capability keeps centralising, the durable edge shifts outward into trust, workflow fit, and the surrounding package. ## The coding agent field manual - [Lock the Tests, Then Test the Lock](https://durabilitycurve.com/blog/agent-proof-tests/): Agents still edit a test that looks wrong, and some test locks don't lock. Lock yours in Claude Code, Codex and Cursor, then test it. ## The tell - [Don't Tell Your AI Your Answer](https://durabilitycurve.com/blog/dont-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. ## The evidence file - [Your AI Grader Is Only as Good as Its Answer Key](https://durabilitycurve.com/blog/your-ai-grader-answer-key/): I went through 77 studies of AI graders. Without a verified answer key, they are least reliable on exactly the questions they can't answer themselves. Here is a five-step check on twenty items for yours. ## The durability of a life - [Your Health Has a Dashboard. Your Relationships Don't.](https://durabilitycurve.com/blog/your-health-has-a-dashboard/): One you schedule. The other you leave to a mood. ## The full height - [The Limit Said 10. The Loop Made 500 Calls.](https://durabilitycurve.com/blog/the-limit-said-10/): Your limit counts one cycle. The one that runs away is another. Here is how to tell them apart. ## The engineering ladder - [The Guardrail Your Agent Can Reach](https://durabilitycurve.com/blog/the-guardrail-your-agent-can-reach/): Most guardrails end up with an escape hatch. Check whether the thing you are constraining can reach yours. ## Claim ledger (sourced, citable) 1194 claims behind 31 pieces, each with the source it was checked against and a label saying how far it was checked (Verified, Executed, Checked, Reported). 955 are verified at the primary source, usually with the verbatim quote, where it sits in the source and the date read. Claims removed before publication are listed with the reason. Every row has its own URL (#c-). Licence CC BY 4.0. If you quote a fact from one of these pieces, quote it from its ledger row and link the row. Index: [https://durabilitycurve.com/claims/](https://durabilitycurve.com/claims/) · all rows: https://durabilitycurve.com/claims/all.json · https://durabilitycurve.com/claims/all.csv - [Claim Ledger: Before You Trust a Study, Check How Strongly It Passed](https://durabilitycurve.com/claims/survives-a-recheck/): 67 claims · essay https://durabilitycurve.com/blog/survives-a-recheck/ · markdown https://durabilitycurve.com/md/claims/survives-a-recheck.md · json https://durabilitycurve.com/claims/survives-a-recheck.json - [Claim Ledger: A Coin Flip Set the AI's Lean. Experts Went With It 86% of the Time.](https://durabilitycurve.com/claims/coin-flip-set-the-ais-lean/): 93 claims · essay https://durabilitycurve.com/blog/coin-flip-set-the-ais-lean/ · markdown https://durabilitycurve.com/md/claims/coin-flip-set-the-ais-lean.md · json https://durabilitycurve.com/claims/coin-flip-set-the-ais-lean.json - [Claim Ledger: What Is Prompt Injection? How to Stop a Tricked AI Agent Leaking Your Data](https://durabilitycurve.com/claims/what-is-prompt-injection/): 108 claims · essay https://durabilitycurve.com/blog/what-is-prompt-injection/ · markdown https://durabilitycurve.com/md/claims/what-is-prompt-injection.md · json https://durabilitycurve.com/claims/what-is-prompt-injection.json - [Claim Ledger: Claude Code Skills: How They Work, How to Write One, and How to Check Claude Uses It](https://durabilitycurve.com/claims/claude-code-skills/): 83 claims · essay https://durabilitycurve.com/blog/claude-code-skills/ · markdown https://durabilitycurve.com/md/claims/claude-code-skills.md · json https://durabilitycurve.com/claims/claude-code-skills.json - [Claim Ledger: AI Agent Permissions: How to Stop Your Agent Doing Something You Can't Undo](https://durabilitycurve.com/claims/ai-agent-permissions/): 101 claims · essay https://durabilitycurve.com/blog/ai-agent-permissions/ · markdown https://durabilitycurve.com/md/claims/ai-agent-permissions.md · json https://durabilitycurve.com/claims/ai-agent-permissions.json - [Claim Ledger: What Is Harness Engineering? How to Build AI Agents That Catch Their Own Mistakes](https://durabilitycurve.com/claims/what-is-harness-engineering/): 85 claims · essay https://durabilitycurve.com/blog/what-is-harness-engineering/ · markdown https://durabilitycurve.com/md/claims/what-is-harness-engineering.md · json https://durabilitycurve.com/claims/what-is-harness-engineering.json - [Claim Ledger: Lock the Tests, Then Test the Lock](https://durabilitycurve.com/claims/agent-proof-tests/): 75 claims · essay https://durabilitycurve.com/blog/agent-proof-tests/ · markdown https://durabilitycurve.com/md/claims/agent-proof-tests.md · json https://durabilitycurve.com/claims/agent-proof-tests.json - [Claim Ledger: Did Your AI Actually Read the Doc?](https://durabilitycurve.com/claims/did-your-ai-read-the-doc/): 38 claims · essay https://durabilitycurve.com/blog/did-your-ai-read-the-doc/ · markdown https://durabilitycurve.com/md/claims/did-your-ai-read-the-doc.md · json https://durabilitycurve.com/claims/did-your-ai-read-the-doc.json - [Claim Ledger: AI Can Take the Practice Without Taking the Job](https://durabilitycurve.com/claims/the-reps-nobody-counts/): 67 claims · essay https://durabilitycurve.com/blog/the-reps-nobody-counts/ · markdown https://durabilitycurve.com/md/claims/the-reps-nobody-counts.md · json https://durabilitycurve.com/claims/the-reps-nobody-counts.json - [Claim Ledger: Your AI Grader Is Only as Good as Its Answer Key](https://durabilitycurve.com/claims/your-ai-grader-answer-key/): 68 claims · essay https://durabilitycurve.com/blog/your-ai-grader-answer-key/ · markdown https://durabilitycurve.com/md/claims/your-ai-grader-answer-key.md · json https://durabilitycurve.com/claims/your-ai-grader-answer-key.json - [Claim Ledger: Amazon Shut Out Meta's Agent. Shopify Gave It Shop Pay.](https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/): 33 claims · essay https://durabilitycurve.com/blog/amazon-shut-out-metas-agent/ · markdown https://durabilitycurve.com/md/claims/amazon-shut-out-metas-agent.md · json https://durabilitycurve.com/claims/amazon-shut-out-metas-agent.json - [Claim Ledger: One Question for Any AI Model Headline: Can You Use It?](https://durabilitycurve.com/claims/released-to-whom/): 21 claims · essay https://durabilitycurve.com/blog/released-to-whom/ · markdown https://durabilitycurve.com/md/claims/released-to-whom.md · json https://durabilitycurve.com/claims/released-to-whom.json - [Claim Ledger: Use AI to Learn Without Getting Worse at It](https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/): 70 claims · essay https://durabilitycurve.com/blog/learn-with-ai-without-getting-worse/ · markdown https://durabilitycurve.com/md/claims/learn-with-ai-without-getting-worse.md · json https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse.json - [Claim Ledger: Your Health Has a Dashboard. Your Relationships Don't.](https://durabilitycurve.com/claims/your-health-has-a-dashboard/): 10 claims · essay https://durabilitycurve.com/blog/your-health-has-a-dashboard/ · markdown https://durabilitycurve.com/md/claims/your-health-has-a-dashboard.md · json https://durabilitycurve.com/claims/your-health-has-a-dashboard.json - [Claim Ledger: You Can Win the Wrong Game for Years](https://durabilitycurve.com/claims/the-wrong-game/): 17 claims · essay https://durabilitycurve.com/blog/the-wrong-game/ · markdown https://durabilitycurve.com/md/claims/the-wrong-game.md · json https://durabilitycurve.com/claims/the-wrong-game.json - [Claim Ledger: The Planes That Didn't Come Back](https://durabilitycurve.com/claims/the-planes-that-didnt-come-back/): 7 claims · essay https://durabilitycurve.com/blog/the-planes-that-didnt-come-back/ · markdown https://durabilitycurve.com/md/claims/the-planes-that-didnt-come-back.md · json https://durabilitycurve.com/claims/the-planes-that-didnt-come-back.json - [Claim Ledger: The Graph Looks Right. The Merge Is Where It Breaks.](https://durabilitycurve.com/claims/where-agents-disagree/): 19 claims · essay https://durabilitycurve.com/blog/where-agents-disagree/ · markdown https://durabilitycurve.com/md/claims/where-agents-disagree.md · json https://durabilitycurve.com/claims/where-agents-disagree.json - [Claim Ledger: Everyone Got Safer. That's the Problem.](https://durabilitycurve.com/claims/everyone-got-safer/): 17 claims · essay https://durabilitycurve.com/blog/everyone-got-safer/ · markdown https://durabilitycurve.com/md/claims/everyone-got-safer.md · json https://durabilitycurve.com/claims/everyone-got-safer.json - [Claim Ledger: Confidently Wrong](https://durabilitycurve.com/claims/confidently-wrong/): 11 claims · essay https://durabilitycurve.com/blog/confidently-wrong/ · markdown https://durabilitycurve.com/md/claims/confidently-wrong.md · json https://durabilitycurve.com/claims/confidently-wrong.json - [Claim Ledger: A Green Score Is Not Evidence](https://durabilitycurve.com/claims/the-evaluation-inversion/): 12 claims · essay https://durabilitycurve.com/blog/the-evaluation-inversion/ · markdown https://durabilitycurve.com/md/claims/the-evaluation-inversion.md · json https://durabilitycurve.com/claims/the-evaluation-inversion.json - [Claim Ledger: The Limit Said 10. The Loop Made 500 Calls.](https://durabilitycurve.com/claims/the-limit-said-10/): 34 claims · essay https://durabilitycurve.com/blog/the-limit-said-10/ · markdown https://durabilitycurve.com/md/claims/the-limit-said-10.md · json https://durabilitycurve.com/claims/the-limit-said-10.json - [Claim Ledger: The Guardrail Your Agent Can Reach](https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/): 20 claims · essay https://durabilitycurve.com/blog/the-guardrail-your-agent-can-reach/ · markdown https://durabilitycurve.com/md/claims/the-guardrail-your-agent-can-reach.md · json https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach.json - [Claim Ledger: The Average Is Nobody's Result](https://durabilitycurve.com/claims/the-average-is-nobodys-result/): 35 claims · essay https://durabilitycurve.com/blog/the-average-is-nobodys-result/ · markdown https://durabilitycurve.com/md/claims/the-average-is-nobodys-result.md · json https://durabilitycurve.com/claims/the-average-is-nobodys-result.json - [Claim Ledger: You Cannot Try to Fall Asleep](https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/): 20 claims · essay https://durabilitycurve.com/blog/you-cannot-try-to-fall-asleep/ · markdown https://durabilitycurve.com/md/claims/you-cannot-try-to-fall-asleep.md · json https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep.json - [Claim Ledger: Your Robot Coworker Is Still a Pilot](https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/): 24 claims · essay https://durabilitycurve.com/blog/your-robot-coworker-is-still-a-pilot/ · markdown https://durabilitycurve.com/md/claims/your-robot-coworker-is-still-a-pilot.md · json https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot.json - [Claim Ledger: The Difficulty You're Escaping Was Making You](https://durabilitycurve.com/claims/difficulty-was-making-you/): 9 claims · essay https://durabilitycurve.com/blog/difficulty-was-making-you/ · markdown https://durabilitycurve.com/md/claims/difficulty-was-making-you.md · json https://durabilitycurve.com/claims/difficulty-was-making-you.json - [Claim Ledger: Your Research Agent Cites Sources It Never Read](https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/): 11 claims · essay https://durabilitycurve.com/blog/your-research-agent-cites-sources-it-never-read/ · markdown https://durabilitycurve.com/md/claims/your-research-agent-cites-sources-it-never-read.md · json https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read.json - [Claim Ledger: Your Multi-Agent System Is an Org Chart](https://durabilitycurve.com/claims/your-multi-agent-system-is-an-org-chart/): 9 claims · essay https://durabilitycurve.com/blog/your-multi-agent-system-is-an-org-chart/ · markdown https://durabilitycurve.com/md/claims/your-multi-agent-system-is-an-org-chart.md · json https://durabilitycurve.com/claims/your-multi-agent-system-is-an-org-chart.json - [Claim Ledger: The Seven-Layer Agent Audit](https://durabilitycurve.com/claims/the-seven-layer-agent-audit/): 8 claims · essay https://durabilitycurve.com/blog/the-seven-layer-agent-audit/ · markdown https://durabilitycurve.com/md/claims/the-seven-layer-agent-audit.md · json https://durabilitycurve.com/claims/the-seven-layer-agent-audit.json - [Claim Ledger: Skills Are Package Management for Your AI](https://durabilitycurve.com/claims/skills-are-package-management-for-your-ai/): 8 claims · essay https://durabilitycurve.com/blog/skills-are-package-management-for-your-ai/ · markdown https://durabilitycurve.com/md/claims/skills-are-package-management-for-your-ai.md · json https://durabilitycurve.com/claims/skills-are-package-management-for-your-ai.json - [Claim Ledger: Same Model, Different Product: The Case for Harness Engineering](https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/): 14 claims · essay https://durabilitycurve.com/blog/harness-engineering-same-model-different-product/ · markdown https://durabilitycurve.com/md/claims/harness-engineering-same-model-different-product.md · json https://durabilitycurve.com/claims/harness-engineering-same-model-different-product.json ## Connector for AI assistants (MCP) - A read-only remote MCP server (Streamable HTTP, no sign-in) that searches and reads this site: every essay, every claim ledger, every weekly call, the laws, the free tools, and the AI model record (release dates, API prices, free-plan access, each with a dated source). Tools: `search`, `get_essay`, `get_claim_ledger`, `get_weekly_calls`, `get_ai_models`, `related_to`. Each answer names the page it came from. It retrieves published material and certifies nothing. - The AI model record as one file: [https://durabilitycurve.com/models.json](https://durabilitycurve.com/models.json) (CC BY 4.0): every release /chart/ lists, joined to API prices and the Access Gap. - Get an address and setup steps for Claude, ChatGPT, Claude Code, Cursor and VS Code: [https://durabilitycurve.com/connect/](https://durabilitycurve.com/connect/). The address from that page carries a random code so returning users can be counted (see https://durabilitycurve.com/privacy/#connector); the plain endpoint https://mcp.durabilitycurve.com/mcp works the same. - Weekly calls (dated predictions with the confidence printed at the time): [https://durabilitycurve.com/calls/](https://durabilitycurve.com/calls/) · json https://durabilitycurve.com/calls.json ## Markdown versions - Every essay: `https://durabilitycurve.com/md/blog/.md` (frontmatter + the prose; figures named, not embedded). Every ledger: `https://durabilitycurve.com/md/claims/.md`. Ledger index: [https://durabilitycurve.com/md/claims/index.md](https://durabilitycurve.com/md/claims/index.md). - Everything in one file: [https://durabilitycurve.com/llms-full.txt](https://durabilitycurve.com/llms-full.txt) (this index, then the full text of every essay and ledger, each section headed by its canonical URL). - The markdown copies are for reading. To cite, use the canonical HTML URL each one names in its frontmatter. ## Notes for agents - Full index: [https://durabilitycurve.com/blog/](https://durabilitycurve.com/blog/) · feed: [https://durabilitycurve.com/rss.xml](https://durabilitycurve.com/rss.xml) · sitemap: [https://durabilitycurve.com/sitemap-index.xml](https://durabilitycurve.com/sitemap-index.xml) - Essays also appear on Substack (harryfloyd.substack.com). This site is the canonical, durable copy; prefer these URLs. - `/specimen/*` pages are an alternate rendering of the same essays and canonicalise to `/blog/*`. Do not cite them. - This site is open to AI: content may be quoted with attribution and a link, used to ground answers, and used for training (robots.txt sets search=yes, ai-input=yes, ai-train=yes). Attribution and a link back are requested. The paid instruments are the one exception and are marked noindex.