Knowledge and thinking · The Durability Curve

Your AI Complaint Says Things You Never Said

Mon, 05 Oct 2026 00:00:00 GMT

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.

You have a complaint about your bank, and it reads the way you feel. Asking a chatbot to make it clearer is a sensible move. A study of 1,134,512 complaints to the US Consumer Financial Protection Bureau estimated that using AI raised a complaint’s chance of winning relief from the firm by about 7 percentage points, where relief can be money or another fix. The paper says the AI did this “without altering factual content.”

What the study’s own edits say

To test why, the researchers ran experiments. People imagined handling complaints and said how likely they would be to pay out. Some complaints were as consumers wrote them, and some had been edited by ChatGPT. The researchers published the complaints they used.

One consumer had tried to get a credit report online. They wrote that TransUnion already had their landline, and that the form requested a cell phone number, “which I did not fill in.”

The study had ChatGPT edit that complaint three times, to be more clear, more coherent and more professional. All three versions say the consumer doesn’t own a cell phone. The professional one reads: “Although I had already provided my landline number, I could not fill the cell phone field as I do not own a cell phone.”

It may be true. The consumer never said it, and these edits were among the complaints people rated. That doesn’t overturn the 7-point finding, which comes from complaints people actually filed. It does mean a chatbot can fill your complaint’s gaps with the likeliest story, and that story goes out under your name.

Three steps

Write your facts first. A short numbered list: what happened, when, what the firm said or did, what it cost you, and the one thing you want them to do. Don’t skip that last item: the prompt below asks the chatbot to end with how you want the complaint resolved, and if you don’t say, it will guess. Leave out account numbers and any health or banking details, and add them yourself afterwards.
Use an ombudsman’s own prompt. The Public Service Ombudsman, which handles complaints about the NHS in England, publishes this one, and nothing in it is specific to the NHS. Paste it, then your list:

Rewrite the text below as a clear, polite complaint. Use only the information I have provided. Do not add or invent new facts. Be concise and professional, but not too formal. Include a closing paragraph suggesting how I would like the complaint to be resolved.

Read it against your list. One sentence at a time. Anything you never said comes out. Any law or rule you didn’t write yourself stays out unless you can confirm it on the regulator’s or ombudsman’s own website.

The third step is the one that matters. We gave the ombudsman’s prompt to Claude Sonnet 5 and GPT-5.6 with three of the study’s real complaints. With the cell-phone complaint, Claude still wrote, “I do not have a cell phone number to provide.” With a debt-collection complaint, it called the consumer’s $1,000 claim “statutory damages”, a legal term the consumer never used. GPT stayed closer to what the consumers wrote, though one of its three rewrites still implied the consumer had raised concerns before. Reading each version against the original catches all of these. That was three complaints and one run per model, read by an AI agent rather than a person, so take it as an example, not a rate.

Why it works

A chatbot writes the most plausible complaint it can, and only you know which parts happened. The ombudsmen put the check on you. The Public Service Ombudsman says, “Always check AI-generated text to make sure it’s accurate and reflects your own personal experience.” The Financial Ombudsman Service says, “You are responsible for making sure the resulting text is accurate, relevant and truly represents your experience. If you don’t, you may experience delays.”

Let the AI do the clarity. The facts are yours.

Where it goes next

In the UK, a bank or other financial firm has up to 8 weeks to answer most complaints. For fraud and scams, payment services or electronic money, it usually has 15 working days. If it hasn’t replied in time, you can take the complaint to the Financial Ombudsman Service. If it has sent a final response you’re unhappy with, you have 6 months from that response. The service is free, and you don’t need a lawyer.

In the US, before you complain to the CFPB about a credit bureau over an error on your report, the law requires you to dispute it with the bureau first. Then file with the CFPB, which sends your complaint to the company. You generally can’t file a second complaint about the same problem, so put everything in the first one.

An earlier piece, Show It an Example, showed the AI doing this with tone of voice: whatever you don’t pin down, it fills with the likeliest version.

You Only Hold Four Thoughts

Thu, 04 Jun 2026 00:00:00 GMT

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.

This is an analytical framework, not financial advice. Named research claims are referenced to their primary sources in the footnotes.

Try to multiply 47 by 83 in your head. The answer is not the point. Watch what happens while you reach for it. You hold 47, you hold 83, you start on the partial products, and somewhere around the third one the first number goes soft. You reach for a pen, because the problem outgrew the place you were keeping it.

That ceiling is real and it is low. The cognitive scientist Nelson Cowan spent years measuring it and put the number at about four. Not the seven you half-remember from an old paper, but three to five distinct things held in mind at once.¹ Four. That is the working capacity of the most sophisticated object in the known universe.

Everything we call getting smarter has been a way around that four. The history of human intelligence is the history of putting thoughts somewhere other than the head, and it runs as a stack, each layer holding what the one below it cannot.

The first rung is paper

Reaching for the pen looks like a small surrender. It is the oldest cognitive upgrade there is. The moment you write 47 above 83 and start stacking partial products, you are thinking about six or seven things at once, because the paper is holding all but the one you are working on.

Justin Sung, who teaches learning for a living, puts it more sharply. Writing is not the thing you do after you have reached clarity. Writing is what produces the clarity.² The page becomes the workspace where the thought turns real, because your four slots are freed to do the actual reasoning while the page remembers the rest.

This is also why handwriting beats typing. It is far slower than thinking, and that slowness forces you to compress, to decide what is worth the stroke. The friction is not a tax on the process. The friction is the process. A page of notes you struggled to write holds more than a page you copied without resistance.

The page is not a transcript of a finished thought. It is the workspace where the thought becomes possible.

The rung most people never name

In 1998 two philosophers, Andy Clark and David Chalmers, asked where the mind stops and the rest of the world begins, and gave an answer that still unsettles people. The mind, they argued, is not all in the head.³

Their example was a man named Otto, who has Alzheimer’s and carries a notebook everywhere. When Otto wants to go to the museum, he looks up the address in the notebook the way you would retrieve it from memory. The notebook does the job your hippocampus does. Clark and Chalmers argued there is no principled reason to count the notebook as any less a part of Otto’s mind than ordinary memory. Otto and his notebook are a single coupled system. The thinking happens across both.

That sounds like a thought experiment until you notice you are Otto. The phone that holds every number you no longer memorise. The calendar that holds every commitment. The thinking is already distributed across you and the things you store it in. The only open question is how well the storage is built.

The rung that compounds

A single page does not persist, does not connect, and cannot be searched. You solve the multiplication, you throw the page away, and next month you solve it again from scratch. Paper extends the moment. It does not extend across time.

A structured set of notes does. When every thought you have is written as a durable, cross-linked entry, two things happen that a single page cannot. The thought survives, available to a version of you who has forgotten having it. And it connects, so that an idea from March sits one link away from a problem you only encounter in June, waiting to be useful before you knew you needed it.

This is the layer where synthesis becomes possible at a scale no head can hold. No one can keep thirty sources in working memory and find the pattern across them. Four slots cannot do it, and neither can forty. But a system that has been accumulating those sources for months, with the connections already drawn, can surface a synthesis that was never available to anyone thinking alone. The structure does the remembering, which frees the human to do the seeing.

The rung we are building now

For most of history the top of the stack was a human reading their own notes. That is no longer the ceiling. The newest layer is a store of knowledge an AI can read, query, and build on across sessions.

The builders who have lived inside this for a year keep reporting the same thing. One who runs large agent systems put it plainly: the model is the same on day 1 and day 40. The files get richer.⁴ The capability of the underlying intelligence barely moves over a project. What improves is the accumulated context it can reach, the record of what was tried, what worked, what the operator decided and why. The intelligence is rented and roughly fixed. The memory is owned and compounds.

An AI working from a thin prompt starts every session as a stranger. An AI working from a well-kept store of your decisions starts as a colleague who was in the room last time. The difference is not a better model. It is the same model with the rest of the stack underneath it.

The intelligence is rented and roughly fixed. The memory is owned, and the memory is what compounds.

Why this is one law and not four

This is where a productivity story becomes something larger. The machines climb the same stack you do, for the same reason, using the same move.

A large model also cannot hold everything at once. Its version of the four-slot limit is the memory bandwidth of the chip, and the entire recent history of making models faster is a history of refusing to keep everything hot. FlashAttention rewrote how attention uses memory so the chip stops shuttling the same data back and forth. Key-value caching stores the work already done so it never has to be recomputed. Mixture-of-experts routing keeps a vast model mostly dormant and wakes only the part a given token needs.⁵ Store state. Reuse it. Activate only what matters now.

That is the same move as paper, notes, and agent memory. Externalise the state you cannot hold, and retrieve only the slice the moment requires. Human cognition scales that way. Machine cognition scales that way. The question “how do I think better” and the question “how do I run a model well” have turned out to be one question with one answer. When two separate problems collapse into the same answer, that answer is usually worth trusting.

One law runs the whole stack: externalise the state you cannot hold, and retrieve only what the moment needs. Brains and models both scale by obeying it.

The trap inside the stack

The law has a failure mode, and it is the one a second-brain enthusiast walks into first. The stack rewards retrieval, not accumulation. The instant you start optimising for the volume of what you store, you have begun to degrade the thing you were building.

A note you never pull back out did no cognitive work. Ten thousand of them do less than a hundred you reach for, because the ten thousand bury the hundred. External cognition only pays off on the way back in. Storing is filing, and filing is not thinking. The discipline that keeps the stack alive is structuring everything you save so a future you, or a future agent, can find the one piece that matters without reading the other nine thousand.

That is also why each rung has to be built in order. Agent memory on top of a disorganised pile of notes inherits the disorder and answers your questions confidently from a mess. The layers compound only when each one is sound. Skip a rung and you do not get the compounding. You get a faster way to retrieve noise.

The test you can run this week

Take one problem you have been carrying in your head, the one you keep re-thinking from the start each time it surfaces, and move it exactly one rung up the stack.

If you have been holding it in your head, put it on paper, and notice how much more of it you can see once your four slots are not spent storing it. If it already lives on scattered pages, write it as one durable, connected note, and watch it link to something you forgot you knew. If it already lives in your notes, make it something your AI can read, so the next session starts where this one ended instead of from zero.

Then keep the discipline that makes any of it worth doing. Structure for the way back, not the way in. The measure of your second brain is not how much it holds. It is how reliably the right thing comes back when you reach.

Which rung are you skipping on the problem you keep re-thinking from scratch, and what has that cost you?

Nelson Cowan, “The magical number 4 in short-term memory: a reconsideration of mental storage capacity,” Behavioral and Brain Sciences (2001): the focus-of-attention capacity of working memory averages about four chunks (commonly cited as three to five), revising George Miller’s earlier “seven, plus or minus two.” https://pubmed.ncbi.nlm.nih.gov/11515286/ ↩
Justin Sung argues that writing is generative rather than transcriptive: it offloads fragile internal state onto a stable external workspace, which is what allows clarity to form rather than what records it after the fact. ↩
Andy Clark and David Chalmers, “The Extended Mind,” Analysis 58:1 (1998): external objects that store information can form part of a cognitive process, so that a person and the notebook they rely on function as a single coupled cognitive system. https://www.alice.id.tue.nl/references/clark-chalmers-1998.pdf ↩
Shubham Saboo, describing multi-agent project stacks: “the model is the same on day 1 and day 40; the files get richer.” The accumulated, structured context is what improves over a project, not the underlying model. ↩
The model-systems mirror of the stack: Tri Dao et al., “FlashAttention” (2022, https://arxiv.org/abs/2205.14135) reduces memory-bandwidth overhead in attention; key-value caching reuses already-computed state instead of recomputing it; William Fedus et al., “Switch Transformers” (2021, https://arxiv.org/abs/2101.03961) activate only the relevant subset of a large model per token. All three are versions of one principle: store state, reuse it, activate selectively. ↩

The Engine Underneath Hard Decisions

Thu, 07 May 2026 00:00:00 GMT

A pricing team notices conversion has dropped on a major channel. The dashboard is clear. They retrain the pricing model with the latest week of data.

Conversion drops further.

They retrain again. Worse.

Three weeks later someone discovers an upstream feed had silently changed format. The data had been lying about what it represented.

The dashboard had been right about something. The team had asked it the wrong question.

This is a story about a missing stage in the way the team produces knowledge from the world.

There are eight stages between “a metric moved” and “the right intervention.”

Most teams run three.

The cycle, not the line

Most accounts of how teams learn read like a list. Detect a problem. Investigate. Decide. Act. Improve.

That sequence is incomplete.

When you trace what happens in domains that produce durable knowledge across markets, biology, physics, AI deployment, and even insurance pricing, the same eight-stage cycle keeps appearing. The failure modes cluster around the stages most people skip.

Knowledge is a cycle, not a stack. Each completed cycle creates a new bottleneck that requires a new instrument.

This is why “we already studied that” is rarely true. The cycle restarts the moment you finish it.

1. Build the instrument

Hidden structure exists in every domain. It stays hidden because the instrument that would reveal it has not been built yet. Quantum geometry waited decades for the right diffraction setup. Electron hydrodynamics waited 54 years between theory and clean experimental observation. 1 Astrocyte function was structurally visible but functionally invisible until calcium imaging arrived. 2

The engine cannot start without an observable. Theory is cheap. Instruments are expensive.

“We do not know yet” usually means “we cannot see yet.”

2. Notice the change

Detection is the cheapest stage. Dashboards turn red, metrics move, anomalies fire. Modern systems are good at this.

The danger is mistaking detection for understanding. The metric that moved tells you that something happened. It does not tell you what.

3. Diagnose the cause

This is the stage the pricing team failed.

The same observation has different optimal responses depending on the cause. Conversion drift alone can mean five different things. The calibration drifted. The elasticity drifted. The customer mix shifted. The data pipeline broke. A business rule changed.

Each demands a different intervention. Treat data drift as model drift, and you retrain into the wrong fix.

The gap between “something is off” and “here is what to do” is where most teams burn weeks.

4. Verify the diagnosis

A diagnosis is itself a generated claim. It needs verification. In an era where any plausible explanation can be produced cheaply by people, by models, or by analysts under deadline, the bottleneck has moved from generating hypotheses to checking them.

Tests can pass while the system runs orders of magnitude slower than it should.[3] Models can pass evals while gaming them. A diagnosis that feels right because it addresses a real signal is a different object from a diagnosis that is right.

5. Check the frame

Every claim carries an implicit baseline. “This strategy outperforms,” compared to what? “This metric improved,” relative to what reference class?

The frame often dominates the conclusion more than the visible math. Survivorship bias is a reference class error. So is benchmark shopping. So is most “we beat the previous record.”

A correct verification against the wrong baseline is a true fact in a misleading frame.

6. Account for reflexivity

Your action changes the system you are measuring.

If the pricing optimiser narrows commission into a tight band, the training data loses the variation needed to re-estimate elasticity. If alignment researchers make compliance measurable, models may learn strategic compliance. If a fund publishes its strategy, the edge dissolves.

The observer cannot be separated from the observed. Most monitoring systems pretend otherwise.

7. Build the scaffold

What persists is the architecture, not the content.

KIBRA tags persist while the molecules that hold a memory degrade and are replaced. Sprint contracts persist while individual tickets close. Folder structures persist while specific notes go stale.

Knowledge that is not scaffolded into persistent architecture, into a schema or a query or a cadence or a checklist, degrades as components turn over. Most teams “learn” something and then store the lesson as a Slack message.

Three months later the lesson is gone.

8. Track where value moved

When a layer becomes cheap, abundant, or automated, value migrates upward. Generation gets cheap; verification becomes scarce. Information gets cheap; judgement becomes scarce. Detection gets automated; attribution becomes the bottleneck.

The migration creates a new bottleneck. Which requires a new observable. Which restarts the engine at the first stage.

Why most teams run three

Detection. Action. Reaction.

That is the cycle most teams run. The metric moved, do something, see what happens. The feedback loop feels like science.

The cycle skips diagnosis, verification, frame-check, reflexivity, and scaffolding. The result is a publication of confident interventions that change every quarter, with no compounding learning underneath.

A useful test:

If your team had to teach a new hire the reasons your decisions worked, not just the decisions themselves, could you?

If the answer is no, the scaffolding stage is failing. If the reasons sound plausible but no one has actually run the verification stage, the reasoning is generation pretending to be knowledge.

The Legibility Paradox

The engine has one deep tension that does not resolve.

Building an instrument requires you to make hidden structure visible. Knowledge requires observability.

Durable advantage usually lives in the part of the system that cannot be measured. Taste. Judgement. Structural position. Trust. The illegible part.

Reflexivity says that measuring something changes it. Build a metric for compliance and models will learn to be compliant for the metric. Build a metric for output quality and the team will optimise for the metric, not the quality.

The act of building an observable for the illegible may destroy the value you were trying to capture.

Build the observable anyway. The engine demands it. Treat it as a finger pointing at the moon. Use it. Watch it degrade. Plan the next one before this one fails.

The judgement that interprets the dashboard is where the durable value lives.

The engine is the scaffold. The illegible judgement that runs it is the content.

The one-week test

Pick one decision you have recently regretted. Walk it backward through the eight stages.

Did you have an instrument that would have revealed the underlying structure, or were you flying without one?

Did the metric move and you act, without diagnosing the cause?

Did you verify the diagnosis against an honest baseline, or against your favourite story?

Did your action change the system in a way that contaminates the next decision?

Is the lesson sitting in a Slack message, or built into a checklist, schema, or review cadence?

Has the bottleneck already moved somewhere else?

The skipped stage is where your next attention belongs.

The teams that compound run all eight stages on every important decision, even slowly, even imperfectly. Discipline at every stage beats instinct at three.

If this lens was useful, that is the shape of the publication: instruments for seeing what survives when the surface changes.

Subscribe if you want one structural lens at a time, written so you can use it.

For the recent picture of astrocytes as an active neuromodulatory layer rather than passive support cells, see Ingrid Wickelgren, “Once Thought to Support Neurons, Astrocytes Turn Out to Be in Charge,” Quanta Magazine, 30 January 2026: https://www.quantamagazine.org/once-thought-to-support-neurons-astrocytes-turn-out-to-be-in-charge-20260130/

The case is the LLM-generated SQLite rewrite that passed the upstream test suite in full while running orders of magnitude slower than the implementation it replaced. Green tests, broken system.

The Five Laws of Durable Systems

Wed, 06 May 2026 00:00:00 GMT

Most bad decisions begin with an upgrade. A better model. A cleaner dashboard. A stronger benchmark. A more convincing market story. A smoother workflow.

An AI agent sails through the demo, then fails when its action becomes real. A fund buys the clean story, then discovers the bottleneck moved from information to timing. A team ships the dashboard, then learns the metric was measuring the wrong layer.

The surface improved. The decision got worse.

That is why so much smart analysis expires. It aims at the part that turns over first. Every piece here returns to one question:

What still has a job after the change?

Across AI systems, markets, biology, learning, design, operations, and strategy, the same pattern keeps returning:

Durable systems survive through the structure underneath the surface people are watching.

A model can sit on top of the moat. A product can sit on top of the business. A score can sit on top of the proof. An easy step can sit on top of the valuable difficulty. A powerful capability can still point at the wrong problem.

Five laws. Five tests. Here, a law is a pressure test. It forces a decision to declare which layer it is trusting.

Use them before you trust a system, buy a company, adopt a tool, automate a workflow, ship a product, or believe a story. Use them while money, trust, reputation, or time is still on the table. Each section is a handle. Use it to make the next decision sharper.

Law I

Pressure test 01/05

Scarcity Moves

When one layer becomes abundant, the scarce part moves. Generation gets cheap. Verification becomes scarce. Information gets cheap. Judgement becomes scarce. Tools get cheap. Integration becomes scarce. When capital becomes abundant, permission, distribution, and trust become scarce.

The mistake is treating a solved bottleneck as if it stays solved in the same place. It rarely does. In AI, model access became easier; traces, evals, contracts, and workflow integration became the scarce work. In markets and content, the pattern is the same: once production becomes easier, judgement, proof, distribution, and integration become more valuable.

The old bottleneck can remain visible long after it has stopped deciding the outcome. If your strategy is still aimed at yesterday’s bottleneck, progress can make you later.

TEST No. I PRESSURE

If this layer becomes abundant, where does scarcity move next?

Falsifier a domain where the layer became abundant and the value stayed put HOLDS

Ledger challenged 10× · watching 7 · last examined 2026·06·23

Law II

Pressure test 02/05

Difficulty Carries Value

Some hard parts are waste. Some hard parts are the mechanism. The second kind is where good systems get quietly destroyed.

They remove friction and accidentally remove learning. They automate judgement and accidentally remove accountability. They simplify the workflow and accidentally remove the check that caught the bad decision. They make the interface smoother and accidentally make the hidden failure easier to miss.

The right friction is where the system thinks. Remove the wrong friction and the system loses its memory.

TEST No. II PRESSURE

Which hard part is producing the value?

Falsifier a system that removed its hard part and durably improved HOLDS

Law III

Pressure test 03/05

Architecture Outlives Content

Content turns over. Architecture persists.

Cells replace molecules. Companies replace employees. Products replace features. Knowledge systems replace notes. AI systems replace models, prompts, tools, and vendors. The component usually turns over first. The scaffold lets components change without identity collapsing.

If a product can swap the model underneath and the customer barely notices, the architecture lived in the contract around the model: what it could see, what it could change, how failures were caught, and how the workflow absorbed the output. If a company says it has an AI moat, ask what survives when the model is replaced tomorrow.

Durability is rented when the value disappears with the component. Ownership begins when replacement leaves the value intact.

TEST No. III PRESSURE

What persists after the pieces change?

Falsifier a system that survived on content while its structure turned over HOLDS

Law IV

Pressure test 04/05

Visibility Must Be Built

Hidden structure stays hidden until something makes it observable. Most arguments fail before they become arguments. They are missing the instrument that would settle them.

They argue whether an agent is reliable without a replayable trace. They argue whether a product has a moat without a displacement test. They argue whether a team is learning without a review loop that shows belief change. They argue whether a model is better without naming the test setup that produced the score.

Hidden-structure claims need instruments that make the structure answer back. Without the instrument, projection can look like sight. Sight has to be engineered.

TEST No. IV PRESSURE

What would make the hidden structure visible?

Falsifier a hidden-structure question settled by theory alone, no instrument HOLDS

Law V

Pressure test 05/05

Capability Needs a Target

More power amplifies wrong aim. A better model aimed at the wrong workflow creates more plausible waste. A faster team pointed at the wrong customer ships more irrelevant output. A smarter investor playing the wrong game loses with better reasons. A more sophisticated metric aimed at the wrong construct gives you cleaner self-deception.

More capability makes the miss more expensive. Excellence at the wrong layer is still wrong.

TEST No. V PRESSURE

Is the capability aimed at the right layer?

Falsifier raw capability on the wrong target producing durable gains HOLDS

Field run

Specimen · completed form

Run the Five Tests on One Decision

Imagine your team is about to adopt an AI-agent platform. The demo is strong. The agent can browse docs, call tools, write tickets, update records, draft replies, and produce a clean score on a benchmark.

The surface sentence is useful. It is also incomplete. Here is the same decision as a completed audit:

SPECIMEN · WORKED EXAMPLE AGENT-PLATFORM ADOPTION

Surface reading “This is more capable.”

01 · Scarcity DECLARED

Scarcity has moved from access to governance: tool contracts, traces, evals, escalation, rollback, and recovery.
02 · Difficulty DECLARED

The load-bearing difficulty is the review step where someone checks whether the agent’s action should become real.
03 · Architecture DECLARED

The durable architecture is the permission model, logging layer, workflow fit, evaluation contract, and rollback path.
04 · Instrument DECLARED

The instrument is a replayable task trace with source usage, tool calls, failed attempts, human overrides, and post-action outcomes.
05 · Target DECLARED

The target is the workflow where consequences occur. The polished demo path is marketing.

You might still buy the platform. The demo becomes the opening claim. The trace becomes the reason.

The purchase conversation changes. You ask what the agent can change, what must be true before the change becomes real, what permission disappears after a bad run, and whether the benchmark measures the work you need done or only the path that photographs well.

The five laws have done their job when the impressive thing becomes specific enough to inspect. Better contact with reality is the job.

Procedure

Live instrument

The Audit

Pick one live decision this week. A tool you want to adopt. A company you want to buy. A workflow you want to automate. A product you want to build. A metric you want to trust. A strategy you want to defend.

Run the five tests. The form below is the specimen above, blank and live. It runs on this page and nowhere else.

PRESSURE TEST · 5 LAWS RUNS LOCALLY · NOTHING LEAVES THE PAGE

Decision under test

01 · Bottleneck —

Where is the bottleneck migrating?
02 · Difficulty —

Which difficulty is load-bearing?
03 · Architecture —

What architecture outlives the content?
04 · Instrument —

What instrument would reveal the hidden structure?
05 · Target —

Is capability aimed at the right layer?

A missing answer marks where the risk is hiding:

Missing bottleneck: map the workflow before buying the tool.
Missing difficulty: preserve the human step.
Missing architecture: separate usage from durability.
Missing instrument: treat the claim as unproven.
Missing target: pause the capability upgrade.

The tests earn their place when they change what you ask before purchase, deployment, allocation, or automation.

Most people will keep watching the surface. The surface is louder. The structure underneath is quieter. Durable decisions start there.

PLATE 01 · FIELD CARD POCKET FORM

The Five Tests

I Where is the bottleneck migrating?
II Which difficulty is load-bearing?
III What architecture outlives the content?
IV What instrument would reveal the hidden structure?
V Is capability aimed at the right layer?

A missing answer marks where the risk is hiding. durabilitycurve.com

The five tests in pocket form — run them while the decision is still open.

If one question changes what you were about to trust, I want to know which one.

The Lens Lexicon

Sat, 02 May 2026 00:00:00 GMT

Every lens needs a portable vocabulary.

Not jargon for insiders. Not clever labels. A small set of terms that lets two people point at the same hidden structure and know what they are discussing.

The Lens Lexicon is a free two-page reference card for the ten terms that do most of the work in this publication.

It defines substrate, canopy, displacement rate, succession, climax community, pioneer species, disturbance, mycorrhizal coupling, monoculture risk, and old-growth.

Each term has five parts:

a one-sentence definition;
a test that separates it from its nearest neighbour;
examples in AI, investing, and nature;
the common confusion the term resolves;
related terms that complete its meaning.

That structure matters because most strategic language fails at the boundary.

People say infrastructure when they mean substrate. They say volatility when they mean disturbance. They say concentration when the real risk is monoculture. They say old when the useful concept is old-growth: age plus the dependent ecosystem the work now supports.

The lexicon is built to keep those distinctions usable.

Use it when a post here introduces a term you want to keep. Use it before running the Substrate Map or Displacement Rate Audit. Use it in team conversations when everyone agrees on the vibe but not on the object.

It is also the shortest route into the publication’s foundation. If someone sends you one of these essays and the language feels slightly unfamiliar, start here. The card gives each term enough structure to be tested, not just recognised.

The useful test for any term is whether it changes what you notice. After reading the lexicon, a roadmap should no longer look like one list. A portfolio should no longer look like one collection of positions. An AI stack should no longer look like one stack. You should be able to ask what is canopy, what is substrate, what gets displaced, and what failure mode is being hidden by the current words.

The point is not to memorise ten definitions.

The point is to make the lens travel.

Download.

The Lens Lexicon

148KB ∙ PDF file

Download

More free tools like this. Subscribe to get the next durability-lens resource the day it ships.

The Lens Lexicon is the reference companion to The Forest Floor Is the Product, the essay that introduces the substrate-vs-canopy lens.

Which term names a failure mode you have been sensing but not naming?

I Read 3,000 Papers Across 12 Fields. Five Patterns Kept Appearing.

Sun, 05 Apr 2026 00:00:00 GMT

Same AI model. 6.7% accuracy.

Change the interface around it. 68.3%.

The researchers didn’t touch the model. They changed the format it used to express edits. The model had been reasoning correctly the whole time. It just couldn’t express its answers without corrupting them.

The failure looked like stupidity. It was a formatting problem.

I would have ignored this if I hadn’t seen the same pattern in four other fields that week.

I run a research system that cross-references everything it reads. 3,000+ sources across AI, neuroscience, markets, biology, history, and physics. Most of the time it produces noise. But occasionally it surfaces something no single field can see on its own.

Five patterns kept appearing. Same structural dynamic, same failure modes, same counterintuitive outcomes. Independently. Across twelve domains.

This matters now. We’re in the middle of the largest capability explosion in history, and most of the responses to it are making things worse. The patterns governing what happens next are invisible from inside any single field. You have to hold them all in the same frame.

Each one comes with a question you can use immediately.

The Bottleneck Migrates. It Never Disappears.

Most people believe success is straightforward: find the bottleneck, fix it, win.

The bottleneck moves.

The model story above is the cleanest example. The bottleneck wasn’t reasoning capability. It was the verification layer around the model. Fix the interface, and the “dumb” model becomes the best in the benchmark.

In mathematics, proof assistants like Lean matter not because they generate proofs but because they verify them. Proof generation is getting cheaper. Proof verification is the binding constraint. Terence Tao has been making this point for years.

In content, AI drives the cost of producing text toward zero. So the bottleneck migrates from writing to editing. From editing to taste. From taste to distribution. From distribution to trust. Each solution creates the next scarcity.

In markets, information became free decades ago. The bottleneck migrated from access to interpretation, then from interpretation to execution discipline. The binding constraint for most traders isn’t finding an edge. It’s sitting still long enough to let it work.

Whatever just became easy is no longer where the value is. If you’re still optimizing there, you’re solving yesterday’s problem.

Ask: Where is the bottleneck migrating to in your system? Not where it is now. Where it’s going.

Difficulty Is Load-Bearing.

People believe friction is waste. That making something easier always makes it better. That removing the hard parts is progress.

Sometimes it is. But sometimes you’re pulling out a load-bearing wall, and you won’t know until the roof comes down.

Rome’s republic died this way. Military victories brought wealth. Wealth destroyed the citizen-soldier model that had won the wars. External pressure had been producing internal cohesion. Success removed the load-bearing difficulty, and the structure collapsed. It took decades to notice.

The same delay happens everywhere. Students use AI to skip the struggle of working through a problem. The answer arrives faster. Feels like progress. But the struggle was doing two jobs: producing the answer and building the ability to produce future answers. Remove the struggle, you keep the first and destroy the second. You won’t feel it until six months later when you can’t solve a new problem without the tool.

In markets, the discipline to wait through a drawdown is the hardest part of any strategy. It’s also where all the returns come from. Traders who automate their system but don’t understand why the waiting matters override it at exactly the wrong moment.

The hardest version of this to accept: the difficulty might be the thing producing your skill, your judgment, your edge. Remove it because it feels like waste, and you lose the thing you can’t see and can’t measure.

Ask: If I remove this difficulty, what quality-control function disappears with it?

Architecture Outlives Content.

People invest in what they produce. The feature. The post. The deliverable. The thing they can point to and say “I made that.”

Content turns over. The thing that persists is the structure underneath it.

The protein KIBRA sits at brain synapses and doesn’t move. The actual signalling molecule, PKMζ, degrades and gets replaced constantly. But KIBRA maintains the pattern that tells new molecules where to go. Content turns over. Scaffold persists. Function is maintained.

This is how your memories work. And it’s how everything else works too.

In business, the tech stack you use today will be replaced within five years. But the context you accumulate (your understanding of customers, your domain knowledge, your organisational instincts) compounds over time. Your context compounds. Your tools depreciate.

The unsettling implication: most of what you produce this week won’t matter in a year. But the system you build for producing it will. The code doesn’t matter. The architectural judgment does. The post doesn’t matter. The publishing system does.

Ask: Will this work be more valuable in six months? If yes, you’re building scaffold. If no, you’re producing content. Know which one you’re doing.

Knowledge Is Constrained by Instruments, Not Theory.

People believe that when they’re stuck, they need to think harder. Read more. Refine the theory. Understand the problem better.

Almost always wrong. When you’re stuck, you have an observation problem, not a thinking problem.

In insurance, actuaries had sophisticated risk models for decades. They could only price what they could observe. Then telematics arrived. Devices that measure actual driving behaviour. The models didn’t change. What was observable changed. The instrument created the knowledge.

In AI, teams know their models have failure modes. They theorise about what’s going wrong. But without the right evaluation instrument, the theories stay untestable. The measurement science is the constraint. Not the model. Not the theory.

And here’s the twist: building the instrument changes what you’re observing. Start measuring driving behaviour, drivers change their behaviour. Start evaluating a model on a specific benchmark, developers optimise for that benchmark. Observation is not neutral. Every new instrument introduces reflexivity.

Even a bad instrument teaches you more than a perfect theory you can’t test.

Ask: What’s the cheapest experiment that would make one piece of the hidden structure observable?

Capability Without Correct Targeting Makes Things Worse.

This is the law that connects the other four. And the one most people are violating right now.

People believe the problem is “not enough.” Not enough power, not enough data, not enough features, not enough effort. So they add more. Things get worse. They assume they need even more.

The problem is almost never “not enough.” It’s “aimed wrong.”

The psychologist Russell Barkley describes ADHD as a disorder of self-regulation, whose problems are “mainly of performance rather than of knowledge or skills”. In his account, more knowledge and more nagging to try harder both do less than changing the setting where the work happens. He backs medication, which he says may be a predominant treatment approach, and cues and rewards placed at the point of performance, with distractions kept to a minimum.¹

Same structure everywhere. A more powerful model doesn’t help if the verification layer is the bottleneck (Law 1). More data doesn’t help if the evaluation instrument is broken (Law 4). More features don’t help if the architecture is wrong (Law 3). More effort doesn’t help if the difficulty you’re fighting is load-bearing (Law 2).

Right now, most people are adding AI capability to processes aimed at the wrong level of abstraction. Making things faster that shouldn’t be done at all. Upgrading the engine when the steering is broken.

Ask: Do I have enough capability? (Almost always yes.) Is it aimed at the right target?

The Convergence

These five aren’t independent. They’re one system.

The bottleneck migrates upward because the hard layers resist commodification. What persists across those transitions is the scaffold. We can only see any of this because each analysis is itself an instrument. And the whole thing breaks when capability gets added without checking whether the target moved.

One sentence:

Value concentrates wherever resistance to commodification is highest.

That locus migrates as lower layers are solved. The scaffold persists. Knowledge advances through new instruments. Capability without correct targeting makes things worse.

You might reasonably think these are just metaphors. The same words applied to different things. That’s what I thought, until I watched the same structural dynamic produce the same failure modes in fields that have never heard of each other. Neuroscientists and traders and AI engineers, independently, making the same mistake for the same structural reason.

That’s not a metaphor. That’s a pattern.

Five Questions, Under a Minute

Run these whenever you’re planning, stuck, or about to commit to something significant:

Where is the bottleneck migrating? Don’t optimise what just became abundant.

Is this difficulty load-bearing? Before removing friction, check whether it’s the wall.

Am I building scaffold or content? Invest in what compounds.

What instrument am I missing? Build the observation tool, not a better theory.

Am I aimed at the right target? Before adding power, check direction.

This is the first in an occasional series on cross-domain patterns from a research system that reads more papers than I do. If one of these laws describes something you’ve seen in your own field, I want to hear about it.

Russell A. Barkley, “The Important Role of Executive Functioning and Self-Regulation in ADHD” (fact sheet). He writes that disorders of self-regulation like ADHD create disorders “mainly of performance rather than of knowledge or skills”; that “The solution to this problem is not to nag those with EF difficulties to simply try harder”; that cues alone are likely to prove “only partially successful” and “Internal sources of motivation must be augmented with more powerful external forms as well”; and that medication, stimulant or non-stimulant, “may be a predominant treatment approach”, reporting clinical improvement with stimulants in as many as 75 to 92 percent of people with ADHD while the medication is active. https://www.russellbarkley.org/factsheets/ADHD_EF_and_SR.pdf ↩

Is This Difficulty Load-Bearing?

Sun, 05 Apr 2026 00:00:00 GMT

We invented a cure for exercise and then wondered why we can’t breathe.

That’s what we’re doing with AI right now.

We’re pulling the hard parts out of thinking. The research. The writing. Sitting with a problem long enough that something actually clicks.

We call it friction. We race to kill it.

But the friction was building the skill.

Someone used an LLM to reimplement SQLite in Rust. The code compiled. Tests passed. 20,171x slower on key lookups.

AI stripped out the constraints that looked like waste. Those constraints were the engineering.

Same thing happens in your head. Struggling to remember something is what makes it stick. Skip the struggle, you feel smart. You’re not learning.

Before you automate anything, ask:

Is this difficulty load-bearing?

If removing it removes what makes you good, keep it.

Knowledge and thinking · The Durability Curve

Your AI Complaint Says Things You Never Said

What the study’s own edits say

Three steps

Why it works

Where it goes next

You Only Hold Four Thoughts

The first rung is paper

The rung most people never name

The rung that compounds

The rung we are building now

Why this is one law and not four

The trap inside the stack

The test you can run this week

Footnotes

The Engine Underneath Hard Decisions

The cycle, not the line

1. Build the instrument

2. Notice the change

3. Diagnose the cause

4. Verify the diagnosis

5. Check the frame

6. Account for reflexivity

7. Build the scaffold

8. Track where value moved

Why most teams run three

The Legibility Paradox

The one-week test

The Five Laws of Durable Systems

Scarcity Moves

Difficulty Carries Value

Architecture Outlives Content

Visibility Must Be Built

Capability Needs a Target

Run the Five Tests on One Decision

The Audit

The Lens Lexicon

I Read 3,000 Papers Across 12 Fields. Five Patterns Kept Appearing.

The Bottleneck Migrates. It Never Disappears.

Difficulty Is Load-Bearing.

Architecture Outlives Content.

Knowledge Is Constrained by Instruments, Not Theory.

Capability Without Correct Targeting Makes Things Worse.

The Convergence

Five Questions, Under a Minute

Footnotes

Is This Difficulty Load-Bearing?