The knowledge atlas LEN P. VAN DER HOF

Good ideas are connected.

Explore the connections between AI, search, ventures and the strategic self. Find an idea. Follow a thread. See what connects.

Explore the atlas Latest articles

The knowledge atlas

Grows with every published article

127 articles 33 concepts 1510 connections

/

214 results/1510 connections

Drag to explore · Scroll to zoom · Select a node

Index Sort results

Articles · 2 min read

AEO meaning: answer engine optimization, not customs

AEO, on this site, means answer engine optimization: writing one URL so an answer engine can lift a clean capsule. Two other AEOs share the letters. Authorized Economic Operator is a customs status. A vendor AEO dashboard is a product named after the craft. If you cannot name which object would change this week, a form, a URL, or a login, you do not yet have a meaning. What AEO means is the definition of the publishing job. This page is only the collision.

Articles · 6 min read

AEO vs GEO vs SEO: three jobs, not three synonyms

SEO, GEO, and AEO are three jobs. SEO is ranking in ten blue links: crawl, index, rank. GEO, generative engine optimization, is the system around an answer engine: retrieve, interpret, extract, verify. AEO, answer engine optimization, is the craft of one URL so that engine can lift a clean capsule. GEO without AEO is a crawl of pages nobody can quote. AEO without GEO is a beautiful capsule the crawler never fetched. They run in parallel, not as a ladder, and they produce three separate numbers: rankings in Search Console, crawler fetches in your own access logs, and named citations in a prompt set you logged yourself. Authorized Economic Operator and geography are collisions, not this page.

Articles · 2 min read

AEO vs SEO: two jobs, two numbers

AEO versus SEO is two jobs. SEO is ranking in ten blue links: crawl, index, rank. The artefact is a ranking. The number lives in Search Console. AEO, answer engine optimization, is writing one URL so an answer engine can lift a clean capsule. The artefact is a quote-shaped first screen. The number is a named citation in a prompt set you logged yourself. They run in parallel, not as a ladder. GEO, generative engine optimization, is a third job (the system around retrieve, interpret, extract, verify) and is not this comparison. The three-way lives next door. Authorized Economic Operator is customs. This page is the two-way people search as AEO vs SEO.

Articles · 8 min read

Agent audit trail: reconstruct the action, approval, and outcome

An agent audit trail is the set of records that lets someone who was not there reconstruct one action by an AI agent: what it did, with which inputs and instructions, who or what approved it, and what actually happened in the system it touched. The records are written by the software around the agent, not by the agent itself, share one ID, and sit where the agent cannot edit them. A chat transcript is not enough: it shows what the model said, not what the tools did.

Articles · 8 min read

Agent runtime budget: keep retries inside the same ceiling

An agent runtime budget is the most one task may consume before the agent must stop and hand back: a ceiling on tokens, money, wall-clock time and tool calls. The choice that decides whether it works is scope. The ceiling must cover the whole task, including every retry, fallback, sub-agent and rerun. Retries multiply cost because each call re-sends the conversation so far and retry layers stack. If each attempt starts with a fresh allowance, a task that fails five times costs five budgets.

Articles · 8 min read

Agent shadow mode: evaluate proposals before granting action rights

Agent shadow mode means an AI agent receives real work and records what it would do, but has no permission to do it. A person or the existing process still acts, and afterwards you compare the two. The idea comes from software shadow testing, where a new version gets a copy of live requests and its answers are discarded or logged. For agents, the write tools must be switched off, not merely ignored. Grant action rights one narrow class of cases at a time, under a written rule.

Articles · 4 min read

An agentic workflow is not an automation

An automation runs a fixed path. Ticket in, fields filled, Slack ping. The diagram does not invent a new diagram. An agentic workflow holds a goal, uses tools, and may change the next step when a tool returns something messy. That is useful only if someone wrote a charter, named the human checkpoints, and set an authority ceiling. Without those, you did not design a workflow. You put a junior hire with root on a looping path.

Articles · 6 min read

AI agent or chatbot: three tests before you mix the words

An AI agent is a system that takes a goal, uses tools, and acts across more than one step without a human prompting each one. A chatbot answers a turn. You type, it replies, and the work dies when the tab closes unless you saved it. An automation runs a fixed script. Mix agent and chatbot and the chatbot inherits an API key. Three tests: does it hold a goal, use tools, and continue across steps? If the answer is no, no, no, you have a chatbot. That is a respectable product. It is not an agent. An agent is only accountable with a charter: inputs, outputs, an authority ceiling, and a named reviewer who can fail the work.

Articles · 3 min read

AI Agents for Startup Strategy: Build the Strategic Intelligence Layer

AI Agents for Startup Strategy teaches founders and operators how to deploy agentic AI as a strategic intelligence layer. Using the SENSE framework, five named agent systems, versioned claims, kill criteria, and a 90-day plan, it turns tactical AI use into governed strategic sensing for startups of roughly five to two hundred people.

Articles · 16 min read

How Answer Engines Choose Sources

Search and answer engines can cite a page only when they can access it, understand its subject, extract a relevant answer and trust the supporting evidence. The strongest strategy is therefore excellent technical SEO plus original, well-structured, source-backed content.

Articles · 2 min read

Answer engine optimization: the craft, not the acronym

Answer engine optimization is the craft of writing one public URL so an answer engine can lift a clean capsule from it: a direct answer near the top of the HTML, a person or organisation attached to the claim, visible sources, and no invented FAQ schema. AEO is the acronym. Dutch searchers use antwoordmotor-optimalisatie. GEO is the system around the crawl, not this craft. SEO is ranking in links, not this craft. Google ended FAQ rich results on 7 May 2026. Google Search ignores llms.txt. This page owns the expanded phrase. What AEO means owns the letters. How to get cited by ChatGPT is the week of work.

Articles · 4 min read

If the answer is in paragraph nine, you did not write a page

Answer-first writing means the first thing a tired reader can use is the conclusion, with the conditions attached, before the origin story, the market size, or the suspense. It is not a robot voice. Google says you do not need to rewrite for generative AI, you do not need to chop a page into chunks, and there is no ideal length. AEO is this craft on one URL. GEO is the system around it. Write the sentence a skeptic can check. Then earn the right to be long.

Articles · 5 min read

What is BooksOS? Seven gates, not a writing CMS

BooksOS is a gated production system for serious nonfiction written with AI agents. The seven stages are Brief, Bible, Map, Draft, Edit, Package, Launch. Each stage owns one decision and ends at a gate. Agents do specialist work inside the gates. Judgment, voice, and accountability stay human. A fluent draft that cannot survive a fact-check is not a book. It is inventory you will throw away. The operating system is the gates, not the model that drafted the chapter. It is not a commercial CMS and not a generic write-faster-with-AI course. The book that owns the name is The Agentic Author, which is live.

Articles · 8 min read

What does calibrate mean? Compare, record, then decide: calibration is not adjustment

To calibrate means to compare what an instrument shows with a reference you trust, record the difference and how sure you are about it, and use that record to read future measurements correctly. Adjustment, changing the instrument so it reads right, is a separate step that needs a calibration first. The International Vocabulary of Metrology, the reference dictionary of measurement science, warns against confusing the two. The same logic applies to judgment: compare your stated confidence with what actually happened before you change how you speak.

Articles · 6 min read

Calibration techniques for founders: five practices that fit a week

Calibration techniques for founders are writing habits that force confidence to meet evidence. The meaning of calibration on this site is matching the strength of a judgment to the strength of the evidence, then updating when the facts move. This page is not the nine-skill CALIBRATE expansion. It is five practices you can run this week: write the bet, put the base rate before the story, write what would move you, journal irreversible calls before you act, and separate process from outcome when you review. A loud claim with a thin source is a calibration failure, not conviction. The Calibrated Mind is live in English and Dutch paperback and hardcover. The five practices are already usable.

Articles · 2 min read

ChatGPT mentions vs citations: a name is not a source

A ChatGPT mention is your name or brand appearing in the answer text. An unnamed idea is not a mention of you. A citation is a named, checkable source the reader can open. They are not the same event. A mention without a URL cannot be independently checked. A URL without your name on the claim is a link, not a citation of you. Why ChatGPT does not name you is the diagnosis of causes. How to get cited by ChatGPT is the work. This page owns the mix-up: treating a name-drop as a citation.

Articles · 7 min read

CHORUS and Chorus AIDLC: which one do you mean?

CHORUS, in capitals, is a coordination protocol: six practices (Charter, Handoff, Orchestrate, Review, Update, Sync) that help small teams share work with AI agents without losing it at the handoffs. It comes from the book The Multi-Agent Organization. Chorus from the GitHub organization Chorus-AIDLC is open-source software: an agent harness, a program that hands tasks to AI coding agents, tracks their state and lets people verify what the agents propose. One is a method, the other is a tool, and you can use either without the other.

Articles · 6 min read

Cialdini in the funnel: six levers, one honesty test

Robert Cialdini's Influence (1984) named six principles of persuasion: reciprocation, commitment and consistency, social proof, liking, authority, and scarcity. In Pre-Suasion (2016) and in the expanded Influence (2021) he added Unity: people are more ready to say yes to someone they treat as one of them. A marketing or sales funnel is a sequence of those asks. You benefit by knowing which lever you are pulling on each step, and by using only the ones you can keep when a human answers. A genuine sample is reciprocity. A counter that invents '12 people viewing' is a claim. The published honest-conversion rule still holds: automate the follow-up, not the claim. This page is the map of the levers. It is not a script for pressure.

Articles · 9 min read

How to measure ChatGPT citations: a weekly log, not a screenshot

To measure ChatGPT citations, keep a log. Ask the same fixed questions in ChatGPT with web search switched on, one new chat per question, and write one row per run: the date, the exact question, whether the answer named you, and whether it linked to your page as a source (a citation). Count only runs that completed. 'Linked in 5 of 14 runs' is a measurement. A screenshot of one good answer is not.

Articles · 3 min read

Citation-readiness audit: eligible is not cited

A citation-readiness audit is a written fail/pass on one public URL. Four checks, in order: can OAI-SearchBot fetch it, is one checkable claim in the first HTML screen with a person or organisation attached, did you test ChatGPT Search rather than a memory chat, and did you log named / linked / used-not-cited as three different outcomes. A page can be eligible for ChatGPT Search and still never be named. Robots.txt, schema, and llms.txt do not close that gap. Google's generative reports measure Google, not ChatGPT. This page is the card; the how-to and the diagnosis are separate pages.

Articles · 9 min read

The cobra story is unverified. Hanoi's rat bounty is documented.

The cobra effect is an incentive that makes the problem it pays to fix worse. The familiar Delhi version, in which people bred cobras to collect a bounty, lacks contemporary documentation in the sources reviewed here. Naturalists consulted in Bombay in 1887 judged breeding for rewards highly improbable. Hanoi's 1902 rat-tail bounty has stronger support: historian Michael G. Vann, working from French colonial archives, describes tailless rats, rat farms, and rats brought into the city. The practical distinction is between the measure you reward and the outcome you want.

Articles · 4 min read

Content architecture that answer engines can lift

Content architecture for answer engines is not a schema trick. An engine can lift a page only when the claim is stated once, on one URL, attached to a named person or organization it can verify, and backed by evidence it can see. GEO is the system. AEO is the page. The standard is a page that is discoverable, attributed, answer-shaped, evidenced, and actually used. Structured data can name the entity. It does not guarantee a citation. FAQPage is not a current rich-result tactic. If the same idea lives in six half-pages, the engine will quote someone who wrote it once.

Articles · 6 min read

Do 90% of startups fail? How to fact-check a viral statistic in ten minutes

No, not in the sense most posts mean. US government data show that 34.7% of private-sector business locations opened in the year to March 2015 were still operating in March 2025, so about two in three had closed after ten years, not nine in ten. In the Netherlands, 64.7% of businesses founded in 2015 still existed five years later. A 90 to 95% figure does exist, but Harvard's Shikhar Ghosh used it for start-ups that fall short of their own projections, not for companies that close.

Articles · 8 min read

Do you need a literary agent for nonfiction? What agents do, what they cost, and when to publish yourself

You need a literary agent for nonfiction mainly if you want a large traditional publisher and your book can command a sizable advance. An agent sells your book to publishers, negotiates the deal, and takes a commission: in the US generally 15% on domestic deals and 20% on foreign deals that use a sub-agent. Nonfiction is often sold on a proposal before it is written. For niche expert books you may sell the project yourself, and in the Netherlands agents are a younger profession. Self-publishing is the third route.

Articles · 8 min read

Downstream effects: why the real cost of a decision lands later, and somewhere else

Affect is the verb. Effect is the noun. A downstream effect is the cost that arrives after the decision, on a surface the room was not watching, with no owner in the meeting. Downstream consequences names the same cost. You do not need to predict every consequence. You need to name the next bend before you commit.

Articles · 5 min read

Downstream effects or affects? The one-line rule, and the two exceptions

Write downstream effects. Effect is usually a noun: the result, the thing that shows up later. Affect is usually a verb: to influence or change something. So a pricing change affects your partners, and the downstream effect is what they feel later. There are two exceptions. Effect can be a formal verb meaning to bring about (effect change), and affect is a psychology noun for the emotion a person visibly shows (flat affect).

Articles · 8 min read

Downstream impact, consequence, implication, outcome, effect: five words, five jobs

Downstream impact means a strong effect that shows up later in a chain, away from the decision that caused it. Downstream means further along a flow. The nearby words have different jobs: an outcome is the result you score at the end; an effect is any change a cause produces; an impact is a forceful effect; a consequence is a result someone has to live with, often an unwelcome one; an implication is a possible future effect, or what something means for a later decision.

Articles · 8 min read

Dutch costs 20% more GPT tokens than English. We measured it on 78 article pairs

Dutch needs more tokens than English for the same content. We counted all 78 English and Dutch article pairs on lenvanderhof.com with tiktoken, OpenAI's open-source token counter. With o200k_base, the encoding tiktoken assigns to GPT-4o through GPT-5, Dutch needed 20% more tokens (109,235 against 90,765). With cl100k_base, which OpenAI names for its text-embedding-3 models, Dutch needed 49% more. A token is the chunk of text an AI model reads, and the unit you pay for, so per-token prices hit Dutch harder.

Articles · 4 min read

Evaluate the reasoning, not the fluency

Fluency is not evidence. A language model can return a clean paragraph that invented a number, cited a blog as a study, or retrieved the wrong chunk with a straight face. Evaluating the reasoning means scoring the trace: which sources were gathered, whether you can name the passage, what assumption is load-bearing, and how strong the claim is allowed to be. CALIBRATE is that match of confidence to evidence. GRAIN is the retrieval loop that keeps an answer tied to a chunk you can point at. If you cannot inspect either, you do not have evaluation. You have a vibe.

Articles · 8 min read

Fake urgency: how to tell honest scarcity from a false deadline

Fake urgency invents a reason a buyer must decide now: a countdown that restarts, a 'limited time' offer that never ends, or an invented stock count. Honest scarcity describes a real limit. EU consumer law bans, in all circumstances, falsely claiming that a product is available only for a very limited time in order to elicit an immediate decision and deny buyers time for an informed choice. False stock claims can be misleading too. The Dutch regulator ACM says timers and stock messages must be true. The test: check what happens when the timer reaches zero.

Articles · 4 min read

Write the five lines before you act

A five-line decision journal is Artifact 3 in The Psychology of Founder Resilience. Write it before you act. The live lines are: Decision (what you choose and what it forecloses), What I believed, What I feared, What happened (fill at review), Next time (one process adjustment). An entry after the result is compromised, because hindsight rewrites what you knew. The book's own evidence note is honest: a journal does not guarantee better decisions. It separates process from outcome so later-you can score the call. CALIBRATE is the practice of matching the strength of a judgment to the strength of the evidence. The Calibrated Mind is live in English and Dutch. The five lines are already usable.

Articles · 5 min read

Flow is a condition of work, not a brand mood

Flow, as Mihaly Csikszentmihalyi described it, is a state of deep involvement in an activity: challenge and skill in balance, clear goals, immediate feedback, and the work done for its own sake as well as for the outcome. He wrote Flow in 1990 after earlier work in Beyond Boredom and Anxiety (1975). In business it is not a vibe for the brand film. It is a condition you can design for in how people work, how a product teaches the next step, how a sales conversation is a craft, and how a founder day is sequenced. If the task is far too hard, you get anxiety. If it is far too easy, you get boredom. Marketing that promises 'flow' and then interrupts every twelve seconds is selling the word and blocking the state.

Articles · 4 min read

A holiday is not a refill protocol

Founder capacity is infrastructure you train and budget, not a mood that returns after a holiday. The live RESILIENT capsule names the work: sleep, recovery, decision quality, and the social cover that lets a founder tell the truth. Treat those four with the same seriousness as the cap table. RESILIENT is an authorial dashboard, proposed and applied, not a validated clinical scale. The Psychology of Founder Resilience is live. This page is the spec, not the book introduction and not a second weekly-review article.

Articles · 3 min read

Founder systems you can run when willpower is gone

A founder system you can run when willpower is gone does not ask you to try harder on a Tuesday at 23:00. The Psychology of Founder Resilience treats willpower as a folk model, not the instrument. The live replacement is a dashboard you read and a weekly review you do not skip: which RESILIENT gauge is empty, and what decision should be deferred. RESILIENT is an authorial dashboard, not a validated clinical scale. Leadership Under Load treats prolonged pressure as a design condition; reducing the load is a valid answer. The Attention Architect treats the inbox as architecture, not a sort you do with leftover grit. The first book is live. The other two are in production and not for sale. The cadence is already usable.

Articles · 3 min read

Frontier, workhorse, or local: three classes, not a religious war

Frontier, workhorse, and local are three classes of language-model work from The Model Portfolio. Frontier is the scarce, expensive tier, rationed for tasks that fail on anything cheaper. Workhorse is the default-capable tier where volume is meant to live. Local is a placement decision: the model runs on infrastructure you control because privacy, latency, sovereignty, or volume says it must. Frontier and workhorse are capability tiers. Local is about where the model runs, and a local model can serve a workhorse route or even a frontier route if your own evaluation says so. The book also names a fourth archetype, specialist, for narrow domains. A private, high-volume, easy task is not a frontier job. The Model Portfolio is live. This page is the split, not the routing policy.

Articles · 7 min read

GEO examples: six before-and-after edits that make a page easier for AI to quote

GEO (generative engine optimization) is the work of making a page easy for AI answer tools such as ChatGPT search, Perplexity and Google's AI Overviews to find, quote correctly and credit. In practice it is mostly editing: put the answer first, replace vague claims with sourced numbers, name things instead of writing 'it', turn image tables into text, write headings as questions, and show a date and a named author. None of these edits guarantees a citation.

Articles · 3 min read

GEO meaning: generative engine optimization, not geography

GEO, generative engine optimization, is the work of becoming a source that answer engines can retrieve, interpret, extract, and verify. It is not geographic SEO. Ranking in ten blue links is a different job. Eligibility, an extractable answer, original value, and evidence that survives scrutiny are the levers. Google does not require schema, chunking, or an llms.txt file for AI Overviews. A page that cannot be quoted cleanly will not be cited, even if it ranks.

Articles · 7 min read

Governed agent memory: what may persist, and who may write it

Agent memory is what an AI agent keeps between runs, such as notes, preferences, facts and lessons, and reads back in later tasks. Governed agent memory means four written rules: what may be written, who may write it, when each entry expires, and how you see and undo any entry. Without them, stale facts, instructions planted by outsiders, and personal data pile up and quietly steer later work. The model does not enforce these rules; the software around it must.

Articles · 4 min read

GPTBot vs OAI-SearchBot vs ChatGPT-User: three bots, three jobs

GPTBot, OAI-SearchBot, and ChatGPT-User do different jobs for OpenAI. GPTBot crawls content that may be used to train OpenAI's foundation models; disallowing it signals that your content should not be used for that training. OAI-SearchBot surfaces websites in ChatGPT's search features; disallowing it keeps your content out of ChatGPT search answers, although a navigational link can still appear. ChatGPT-User fetches a page when a person asks ChatGPT or a Custom GPT to look at it. OpenAI says robots.txt rules may not apply to those user-initiated actions and that this agent is not used to decide Search inclusion. The GPTBot and OAI-SearchBot settings are independent, and OpenAI says its Search systems can take about 24 hours to reflect a robots.txt change.

Articles · 5 min read

A go-to-market system that survives the first ten customers

A go-to-market system that survives the first ten customers treats those customers as evidence, not as trophies. Ten is the working set The Founder Sales System uses for deals closed on instinct. It is not a statistical threshold. If you have four, audit four. Write one sourced claim per customer: who they are, why they bought, what they use, what would make them leave. SENSE is how that claim enters a standing layer: Scan, Evaluate, Navigate, Simulate, Execute. PREDICT is how you keep the prior visible so the tenth customer cannot rewrite what you believed before they paid. AI Agents for Startup Strategy is live. The other two books are in production and not for sale. None of this is a vendor CAC story.

Articles · 9 min read

Hiring an AEO agency? Seven questions to ask before you sign

No AEO agency can guarantee a citation; OpenAI itself says placement in ChatGPT search is not guaranteed. What you can buy is the work and a record of what changed. Before you sign, get seven answers in writing: which questions the agency will measure and against what baseline, what it checks before touching the writing, what it will change and why, what it will not promise, which advice comes from official documentation and which from its own tests, how it reports mentions, citations, and clicks separately, and who checks the facts. AEO, answer engine optimization, is the work of helping AI answer engines find, understand, and cite your pages.

Articles · 6 min read

HITL in AI workflows: put the named person on the irreversible step

Applying human in the loop to an AI workflow means a named person sits on the irreversible step with a written fail line. Charter the agent. Write the override doctrine. Put the reviewer on send, spend, delete, and speak-as-the-company. Show that reviewer the exact payload, target, reason for action, and reversal cost on one screen, with a deny that stops the call. For a new write tool, inspect the first ten attempted writes end to end. That startup review does not retire permanent gates: sends, spend above the ceiling, deletes, and company speech still require the named person. If the output can leave while that person is offline, the loop is broken. A dashboard after the send is archaeology.

Articles · 4 min read

A conversion system that stays honest when you automate it

A conversion system that stays honest when you automate it treats the sequence as a clerk, not a closer. Automate the follow-up: the reminder, the unanswered question, the date you already named. Do not automate the claim. Fake countdown timers, manufactured scarcity, shame in the subject line, and a testimonial the customer did not write are not 'lifecycle marketing.' They are a faster way to spend trust. TRUST is the pre-send test from Marketing Without Manipulation: Transparency, Respect, Utility, Selectivity, Time. PRICE is the honesty of the number and the frame. VOICE is whose name the from-line still carries when a model wrote the email. All three books are in production. None is for sale. The send rule is already usable.

Articles · 10 min read

How to get cited by ChatGPT

There is no published ChatGPT citation ranking. OpenAI's publisher FAQ says any public site can appear in ChatGPT Search, and that summaries and snippets require you not to block OAI-SearchBot. That is eligibility. To be the source named in the answer, put one checkable claim in the first HTML screen, attach a person or organisation the machine can identify, and keep the evidence on the same URL. GPTBot is a separate training switch. ChatGPT-User is a person-triggered fetch and does not decide Search inclusion. Measure appearances, chatgpt.com referrals, and conversions as three numbers.

Articles · 8 min read

How to run a pre-mortem in 20 minutes

A pre-mortem is a short meeting before you commit to a plan: the team assumes the plan has already failed and explains why. Gary Klein described the method in Harvard Business Review in 2007. This 20-minute version keeps his sequence of briefing, declared failure, silent writing and one reason per turn, then gives three priority risks an owner and a response. Use it to surface risks; the evidence reviewed here does not show that it makes decisions more accurate.

Articles · 6 min read

Human in the loop meaning: a named person who can fail the work

Human in the loop means a named person holds an authority ceiling and a review gate. That person charters the work, can stop a tool call, and owns the outcome. It is not a human glancing at a dashboard. If no one can fail the output against written criteria, the loop is decorative. A vendor slide that says HITL without a named reviewer is a slogan. The name, the ceiling, and the fail criteria have to live in the runbook, not in the deck.

Articles · 2 min read

Human on the loop: watching is not a fail gate

On this site, human on the loop is a named monitor who can stop a later run after observing the current one. Human in the loop is a named reviewer who can fail the current run before it ships. The difference is intervention timing, not job title. This split is house usage, not an IEEE or ISO standard. Human in the loop meaning owns the in-the-loop definition; this page owns the current-run versus next-run distinction.

Articles · 9 min read

Is there an ideal word count for AI citations?

No word-count target appears in the guidance reviewed here. Google says there is no ideal page length and no requirement to break content into tiny pieces, and the OpenAI and Perplexity documents checked set no word count. Edit the passage that answers the question instead: can it stand alone with its subject, main qualification, and evidence intact? The passage test checks that. It does not predict whether a page will be cited.

Articles · 4 min read

Introducing Undominated.ai: the independent AI inference price index

Undominated.ai is an independent AI inference price index. It ranks models on independently measured capability first. Price breaks ties. A model is named as dominated when another scores at least as high, costs no more, and is strictly better on one of the two. Unrated is not zero. On 25 August 2026 the live board held 409 models from 50 providers; of 108 rated, priced rows, 97 were dominated and 11 were not. Those eleven are the value frontier. The site takes no cut of inference, no affiliate, and no paid placement. Snapshot numbers move. The live board is the source.

Articles · 7 min read

Is SEO dead? No. Answer engines turned it into the entry ticket

No. SEO (search engine optimization) is not dead; it became the entry ticket for AI answers. Google says a page must be indexed and eligible for a snippet to appear as a link in AI Overviews or AI Mode. OpenAI says that for your content to appear in ChatGPT's search summaries, its search crawler, OAI-SearchBot, must not be blocked. What changed is the click: people often get the answer without visiting, so measure citations and visits separately, in your own data.

Articles · 5 min read

Kahneman for marketing: System 1 is not a permission slip

Daniel Kahneman's work, much of it with Amos Tversky, is a map of judgment under uncertainty. Prospect theory (Econometrica, 1979) codes outcomes as gains and losses from a reference point. The value function is generally steeper for losses than for gains. Thinking, Fast and Slow (2011) popularised two modes of thought: a fast, associative System 1 and a slower, more effortful System 2. Marketing lives in that map every time a price, a plan, or a headline sets a reference. You benefit by designing for how people actually compare, not for the rational agent on the whiteboard. You do not benefit by treating System 1 as a door you are allowed to sneak through. A frame that would collapse if the buyer had ten quiet minutes is not applied psychology. It is a rush.

Articles · 6 min read

Len or a fractional CAIO: which work do you need?

A fractional CAIO (chief AI officer) is a senior AI leader hired part-time to own AI across a business: priorities, rules for safe use, and reporting to leadership. Len P. van der Hof does not offer that seat; his site calls his work 'not a packaged fractional CAIO shop'. He advises on one stuck decision about AI agents, does venture work, or installs a named framework on a system you already run. Need an owner for AI? Hire a CAIO. Stuck on one decision? The narrower work fits.

Articles · 8 min read

LEVEL vs CALIBRATE: check the claim, or check your own judgment

LEVEL and CALIBRATE are two frameworks by Len P. van der Hof that answer different questions. LEVEL, from The Evidence-Aware Life, checks one claim before one decision in five steps: Locate, Estimate, Value, Expose, Lean. CALIBRATE, from The Calibrated Mind, is a nine-skill practice for checking, across many judgments, whether your stated confidence matches what actually happens. Use LEVEL when a claim is in front of you. Use CALIBRATE to learn how far your own confidence can be trusted.

Articles · 5 min read

Literary agent or AI agent: two different jobs

A literary agent is a human who represents an author's work to publishers, usually on commission, and is not this website. An AI agent is a system that takes a goal, uses tools, and acts across steps under a charter. If you searched 'literair agent non fictie' looking for someone to sell a Dutch nonfiction manuscript, this page is a collision warning, not a roster. If you searched because a slide mixed the words, three tests still decide whether the software is an agent: goal, tools, continuation. This site publishes operator nonfiction. It does not shop manuscripts.

Articles · 4 min read

LLM integration is an operating-design problem

LLM integration is not an API problem. The call is cheap. The scarce work is operating design: who owns the model and the prompt, what the tools may see and change, who can fail the output, and whose name it carries. A named reviewer and a written override doctrine are the integration. STACK versions the artifacts. MCP is the contract on one tool. Neither substitutes for those four seats being filled.

Articles · 6 min read

Marketing nudges are choice architecture, not a growth hack

A marketing nudge is choice architecture applied to a commercial decision. Richard Thaler and Cass Sunstein, in Nudge (2008), treated a nudge as a change in how options are arranged that predictably shifts behaviour without forbidding options or significantly changing economic incentives, and that stays cheap to avoid. Fruit at eye height is a nudge. Banning the pastry is not. In a funnel the default, the order of fields, the extra click, and the pre-ticked box are the architecture. You already have them even if you never named them. The work is to design them so a tired buyer can still see the choice and reverse it. A default that hides the price, invents a limit, or makes opting out a maze is not a nudge. It is a claim you would not defend on a call.

Articles · 5 min read

MCP explained for founders

MCP, the Model Context Protocol, is a contract for how a language model calls tools. The contract names what the model may see, what it may change, and how each call is logged. A founder should treat an MCP server like an API with production credentials: no anonymous tools, no undocumented write paths, no server that can spend or delete without a ceiling. Without that contract you do not have integration. You have a hole in the repository.

Articles · 7 min read

MCP is not USB-C: where the favourite analogy breaks, and why that matters

MCP (Model Context Protocol) is an open standard that lets AI applications connect to outside tools and data through one common interface. Its official introduction compares it to a USB-C port, which is right about the plug. It breaks in four places: the model reads each tool's description as text and can be steered by it, a server can act rather than just carry data, permissions live in the app rather than the protocol, and a registry listing proves who published a server, not that it is safe.

Articles · 4 min read

MCP meaning: Model Context Protocol, not a product

MCP means Model Context Protocol. It is a contract for how a language model calls tools: what it may see, what it may change, and how each call is logged. The official split is host, client, and server. A server exposes tools, resources, and prompts to a host: tools are model-controlled, resources application-controlled, prompts user-controlled. The official reference servers are filesystem, fetch, git, memory, time, sequential thinking, and everything. They are educational examples, not a production catalogue; the other former reference servers are archived. The MCP Registry is the catalogue. ReasonKit is first-party to this site, not an official reference. MCP is not a product you buy or the definition of an agent. A clean server with no ceiling is still a hole with nicer plumbing.

Articles · 8 min read

MCP OAuth is not tool authorization: a valid token is not a yes

On a remote MCP server, OAuth answers two questions: is this access token genuine and issued for this server, and which scopes (named permissions) does it carry? It does not know which scope each tool needs, or that a user may only approve expenses in their own team. You write that map and check it on every tool call, before the tool changes anything. Local MCP servers that run over stdio skip this OAuth flow and take credentials from their environment.

Articles · 7 min read

MCP vs function calling: two layers of the same tool call

Function calling (also called tool calling) is a model replying with a structured request to run a named tool with specific inputs; your application runs the tool and sends back the result. MCP, the Model Context Protocol, is an open standard for putting tools in a separate server that any compatible AI app can discover and call. They are layers, not rivals: under MCP the model still makes a function call, and the app forwards it to the server. Neither layer decides what a tool may do.

Articles · 9 min read

Why multi-agent LLM systems fail: the 14 MAST failure modes

Multi-agent LLM systems fail when an agent makes a mistake and the workflow around it lets the mistake pass: roles are unclear, context is lost or withheld between agents, or checks accept unfinished work. The MAST study (Cemri and colleagues, NeurIPS 2025) labelled 1,642 execution traces from seven open-source frameworks and found 14 failure modes in three groups: system design issues, inter-agent misalignment and task verification. Clearer role prompts and a review loop helped in the authors' tests but did not establish reliability. The figures describe research frameworks on benchmark tasks, not your company's failure rate.

Articles · 6 min read

Is one AI model enough? When a single model works, and when you need a portfolio

One AI model is enough when volume is low, there is one kind of task, and no regulated or confidential data is involved. Outside that case, one model for everything means paying top rates for easy work and having no plan when that model fails or is retired. A model portfolio is the fix: a written list of which model does which job, why, what takes over when it fails, and when the choice is reviewed. More models without those rules is just a collection.

Articles · 7 min read

The 90% test: why 'I'm sure' is right far less often than it feels

The 90% test asks you to give ranges you are 90% sure contain the true answer, then counts how many do. Most people's ranges are too narrow. In one experiment with 334 people, ranges meant to be 90% sure held the right answer only about 54% of the time. Researchers call this overprecision: being surer about your numbers than your knowledge allows. The fix: ask for the low end and the high end as two separate questions, and track how often your ranges hold.

Articles · 3 min read

What an agent may never do

Override doctrine is the written list of decisions an agent may never take. The live glossary names four: spending above a ceiling, speaking as the company, changing production data, and closing a judgment the founder still owns. It is the complement of a charter. The charter says what the agent does. The doctrine says what it must refuse. Write it before the agent gets a tool that can send, spend, or delete. A named human in the loop is not a substitute. If the irreversible act already happened, the loop is decorative. CHORUS Review is the gate that enforces the list. The list is not the six-skill protocol.

Articles · 5 min read

Person entity SEO: a name an engine can attach

Person entity SEO is the work of making a public identity consistent and unambiguous across pages and references. Build one visible author profile, give its Person entity a stable @id, point every Article's author markup at that same @id, and use sameAs only for URLs that unambiguously identify the same person. Schema.org defines sameAs that way, with a Wikipedia page, a Wikidata entry, or an official website as examples; owning the destination is not required. Google supports Person as the main entity of a ProfilePage and recommends sameAs for external profiles or home pages, but structured data does not guarantee a knowledge panel, a citation, or a ranking benefit. The markup must agree with what readers can see.

Articles · 9 min read

The planning fallacy: a plan is not a forecast

The planning fallacy is the tendency to expect your project to go to plan despite knowing that similar projects ran late. Daniel Kahneman and Amos Tversky used the term in a 1977 research report. An estimate built from the steps of a plan can give too little weight to how similar work has gone before. Before committing to a date, ask how long comparable projects actually took. Start from that record and adjust only where you have evidence that this project differs. The check can expose an optimistic estimate; it cannot guarantee an accurate one.

Articles · 3 min read

Positioning is a decision, not a tagline

Positioning is a decision about who you will not serve and what you will not claim. A tagline is a sentence that can still fit everyone. A position is a refusal you can keep at the table, on the homepage, and in the company you would still want to run. The walk-away is part of the position: if you will take any logo and any discount to keep the conversation, the homepage is lying. The Negotiation Operator, The Meaning Operator, and The Authentic Operator are in production. None is for sale. The two lines are already usable: who you refuse, and the claim you will not make.

Articles · 3 min read

Pricing psychology: a price is never a naked number

Pricing psychology is the study and design of how a buyer experiences a price, not the price as a naked number. On this site the named system is PRICE: Perceive, Reference, Irrational, Context, Endow. Perception is what the number is seen against. Reference is the comparison the buyer actually uses. Irrational names the systematic biases you refuse to pretend are noise. Context is the frame around the offer. Endow is anticipated ownership: how it feels to have it versus not. It is not a playbook for hiding fees or nudging a tired checkout. A conversion system that stays honest is the neighbouring operating note. The Pricing Mind owns the expansion and is in production, not for sale. This page is the meaning, including the Dutch query prijspsychologie.

Articles · 8 min read

Prove you can stop an agent: interrupt the run and revoke access

You can stop an AI agent only if a named person, with a deputy, can do four things within a written time and show evidence for each: interrupt the run in progress, cancel or hold the work queued behind it, suspend the keys and tokens it acts with, and leave half-finished work in a safe state. A line in the prompt such as "do not change anything" is not a stop. Prove it with a timed drill on a harmless run before you need it.

Articles · 3 min read

The Psychology of Founder Resilience: Capacity Is Infrastructure

The Psychology of Founder Resilience is an operator manual for high-stakes founders who need trainable psychological capacity under chronic uncertainty. It replaces grit slogans with the RESILIENT nine-capacity model, concrete practices, and a weekly protocol covering identity boundaries, reappraisal, attention, decisions, recovery, social support, and antifragility.

Articles · 6 min read

RAG freshness: last month's PDF is not a lookup problem

RAG freshness is the managed agreement between authoritative sources and the index used at answer time. An index is stale when a source changes, expires, is deleted, or changes permissions and that state has not reached retrieval within the declared freshness window. A stale embedding is a vector built from superseded text, or a vector left behind by a preprocessing or embedding-model change that the serving system no longer matches; an unchanged vector does not decay because it is old. Freshness therefore needs source versioning, change detection, deletion handling, re-ingestion, index versioning with atomic cutover, evaluation, a per-corpus TTL or lag budget, and a defined response when the window is breached. A TTL is the boundary, not the update. In the authorial GRAIN model, Navigate freshness owns this loop.

Articles · 6 min read

RAG vs fine-tuning vs agents: three jobs, not a ladder

RAG, fine-tuning, and agents are three jobs. RAG, retrieval-augmented generation, looks up a named passage in a corpus you gave it, then writes. Fine-tuning changes model weights so a stable skill fires without fetching a document. An AI agent holds a goal, uses tools, and continues across steps under a written charter. They are not a ladder of sophistication. Fine-tuning does not make the model know Tuesday's PDF. RAG does not give the model a ceiling. An agent that cannot open a source still needs retrieval or a refuse. What is RAG owns the method. What is an AI agent owns the worker. This page is the comparison. The RAG Engineer is live. Fine-tuning win-rates are not this page.

Articles · 9 min read

How to retire a production agent (stopping it is not enough)

To retire a production AI agent, close every way it can start work, sign in or cost money, then prove each one is closed by trying it. In order: freeze new work, finish or hand over accepted jobs, redirect its callers, revoke its keys and grants, cancel its subscriptions and reservations, and keep the records your policy requires. Close the ticket only when the old triggers and credentials fail and every remaining charge matches a service period.

Articles · 6 min read

What is a roster? From the Monday rota to a team of AI agents

A roster is a named list of who is available for a job and what each one does, sometimes with when they do it. A staff roster lists employees and their shifts, a sports roster lists the players a team may use, an agency roster lists the outside agencies a brand retains, and an AI agent roster lists the software agents a business runs, each with one job. English borrowed the word from Dutch rooster, originally a gridiron, in 1727.

Articles · 7 min read

ROSTER vs CHORUS: AI agents for one person, or for a team?

ROSTER and CHORUS are two frameworks by Len P. van der Hof for working with AI agents. ROSTER (Roles, Objectives, Skills and tools, Triggers, Evaluation, Rotation), from The Agent Dream Team, is for one founder who runs several agents and holds every decision. CHORUS (Charter, Handoff, Orchestrate, Review, Update, Sync), from The Multi-Agent Organization, is for teams of 2 to 15 people who share work with agents. ROSTER defines each specialist; CHORUS moves work between people and agents.

Articles · 5 min read

Second-order thinking audit: the first-order metric is green

A second-order thinking audit is a written fail/pass on one decision before you commit. Four checks: who pays later, on which surface, whether you can reverse it this quarter, and which incentive the decision just created. A first-order win with an unnamed later bill is not a decision. It is a delay. CASCADE, on this site, is the personal habit that fires that pause in your own head. DOWNSTREAM is the organisational protocol for delayed, displaced, previously unowned costs. They are companions, not synonyms. What are downstream effects owns the definition. The operator article owns the longer map. This page is the card. Both books are in production and not for sale.

Articles · 8 min read

utm_source=chatgpt.com: how to see ChatGPT and other AI traffic in GA4

utm_source=chatgpt.com is a tag ChatGPT adds to the links in its answers, so Google Analytics 4 (GA4) can record chatgpt.com as the source of a visit. To see these visits, open Reports, then Acquisition, then Traffic acquisition, and look for the AI Assistant channel Google announced in May 2026, or switch the dimension to Session source and type chatgpt into the table's search box. GA4 only records clicks: an answer that cites your page without a click leaves no trace there.

Articles · 6 min read

SENSE: a 90-day strategic intelligence layer

A strategic intelligence layer is a standing system that turns market, competitor, and internal signals into sourced claims a founder can act on. SENSE is one implementation in five moves: Scan, Evaluate, Navigate, Simulate, Execute. It has no Decide phase: the system prepares the choice, a person owns it. In ninety days you Diagnose, Deploy, then Evaluate-and-decide. Watchlist, assumption register, cadence, and kill rule are artifacts inside those gates, not a second calendar. The unit of work is a claim with a confidence grade, not a longer PDF.

Articles · 4 min read

STACK: version the agent OS like you version the app

STACK is a five-layer architecture for repositories where humans and agents share work: Structure, Toolchain, Agent configuration, Connection, Knowledge and quality. The agent operating system should be versioned like the application. If prompts, MCP servers, and skills live outside the repo, or live in five places inside it, you have agentic repo debt. The fix is ownership and a single source of rules, not a smarter model.

Articles · 7 min read

STACK vs AGENTS.md: where instructions end and controls begin

AGENTS.md is a plain Markdown file in a code repository that tells AI coding agents how the project works: setup, tests, conventions. It is an open format stewarded by the Agentic AI Foundation. Agents read it as context, so it guides but cannot enforce. STACK, from The Agentic Codebase, is a five-layer map of what a repository needs once agents work in it; AGENTS.md lives in its Agent configuration layer. Rules that must hold every time belong in controls: permission rules, hooks, sandboxes, branch protection and automated checks.

Articles · 2 min read

Stoic systems: judgment as infrastructure, not a mood

Stoic systems, on this site, are decision infrastructure, not a temperament. The named system is STOIC: Selective Attention, Temporal Distance, Outcome Detachment, Intentional Architecture, Continuous Review. These are five trainable skills for governing judgment under pressure. This page is an operating definition, not ancient-philosophy scholarship or clinical guidance. Stoic Systems for Decision-Making is in production and not for sale.

Articles · 7 min read

Being stoic is not Stoicism: the philosophy was never about hiding your feelings

Being stoic, in everyday English, means not showing feeling, especially under pain or distress. Stoicism is an ancient philosophy that taught something different: much of what disturbs us comes from the judgments we make about events, and those judgments can be examined and corrected. The Stoics did not teach people to hide feelings, because hiding a reaction leaves the judgment untouched. Modern psychology calls the closer skill cognitive reappraisal: changing how you think about an event rather than masking your response to it.

Articles · 9 min read

Technical GEO: the plumbing that decides whether AI can use your page

Technical GEO is the part of generative engine optimization (GEO: making pages usable as sources in AI-written answers) that a developer fixes rather than a writer. It covers whether AI search crawlers may fetch the page, whether the text is in the HTML they receive, whether the page can be indexed, whether snippet settings allow quoting, and whether the page has a stable address, a visible date and a named author. Google says its AI features need no special markup or AI files.

Articles · 5 min read

The Agentic Author: Stop Writing Books. Build a Production System.

The Agentic Author is the operator manual for BooksOS, a seven-stage nonfiction production pipeline where AI agents run stages and the author certifies quality gates. It covers evidence tracking, red-team review, bilingual native editions, and tool-agnostic design so authors ship serious work without surrendering the byline.

Articles · 3 min read

The Agentic Codebase: Version the Agent OS Like You Version the App

The Agentic Codebase is a production architecture book for technical founders and senior builders shipping agentic software. It introduces the STACK five-layer model so AGENTS.md, rules, skills, MCP, hooks, evals, and shell automation become versioned repository infrastructure that survives tool churn and team growth.

Articles · 8 min read

The Multi-Agent Organization: Close the Coordination Gap With CHORUS

The Multi-Agent Organization is an operator manual for 2-to-15-person teams that already mix human operators and specialized AI agent workers. It installs CHORUS, a six-skill coordination protocol: Charter, Handoff, Orchestrate, Review, Update, Sync. The book treats agents as workers with contracts, not as a chat demo or a vendor runtime, and it runs in the tools you already have.

Articles · 7 min read

ChatGPT, Perplexity, Google AI Overviews: three engines, three doors to your page

ChatGPT search, Perplexity and Google's AI Overviews each reach your page through a different door. ChatGPT search needs OpenAI's OAI-SearchBot crawler to be allowed. Perplexity needs PerplexityBot. Google's AI Overviews and AI Mode draw on the ordinary Google Search index, so they need Googlebot, indexing and snippet eligibility, with Search Console's generative AI setting left on include. Opening one door does nothing for the others, and none of the three publishes how it picks the sources it cites.

Articles · 6 min read

TRACE vs RAG: one fetches the text, the other checks what it proves

RAG (retrieval-augmented generation) is a software design: before an AI model answers, it fetches passages from a document collection and writes from them. TRACE is a research method from The Verified Research Loop: Target, Retrieve, Assess, Cross-check, Export. RAG answers the question 'which text did the model use?' TRACE answers 'what does the evidence prove, and may we act on it?' RAG can feed TRACE's Retrieve step. Neither one proves a conclusion true on its own.

Articles · 8 min read

Upstream vs downstream: one river picture that explains oil, supply chains, software, and decisions

Upstream means earlier in a flow, closer to the source; downstream means later, closer to where things end up. The picture comes from rivers and works across fields: in oil, exploration and production versus refining and sales; in supply chains, suppliers versus customers; in software, the original project versus the products built on it; in data pipelines, the task before versus the tasks after; in decisions, the choice versus its later effects. The labels are relative: the same team can be downstream of one group and upstream of another.

Articles · 7 min read

Using AI agents effectively: one job, a ceiling, a log

An AI agent is already defined on this site: a system that holds a goal, uses tools, and continues across steps under a charter. Using one effectively means giving it one job, the smallest tool set that job needs, a written refuse-list, a hard ceiling on steps and spend, and a human on send, spend, delete, and speak-as-the-company. Writes should be draft-first or reversible, and an unfamiliar tool error should stop the run rather than start a retry loop. Optimisation is not a bigger model. It is a tighter charter, fewer write tools, a log you can open on Tuesday, full review of early writes, and a permanent human gate on irreversible calls. If it cannot refuse and cannot stop, it is not an agent you can run.

Articles · 4 min read

What are downstream effects?

A downstream effect is a consequence of an operational decision that is delayed, lands on a different surface than the metric you optimized, and had no owner in the room when the decision was made. Downstream consequences is the same river with slightly wider language. Downstream affect is the verb; the noun you want is effect. You do not need to predict every consequence. You need a way to name the next bend before commitment. The first-order metric can be genuinely green. Partner margin, a regulator queue, an agent network, or a customer renewal are typical landing surfaces.

Articles · 7 min read

What does 'likely' mean? Ask five people and you get five numbers

'Likely' has no fixed number. In a survey of about 1,700 people, answers for common probability words spread widely, and in a 2020 study of 881 Dutch speakers, nine in ten put the Dutch word for likely anywhere from 57% to 91%. So two people can agree that something is 'likely' while holding very different odds. When the number matters, say it, together with the event, the date and how it will be judged: 'likely (about 70%) that we ship by 30 June'.

Articles · 7 min read

What is a confidence grade? Why "high confidence" is not 90%

A confidence grade is a short label, such as high, medium or low, that records how well the evidence you have checked supports one specific claim, with the reasons written beside it. It is not a probability: the IPCC tells its authors that confidence should not be read as a probability, and US intelligence standards forbid mixing the two in one sentence. It is not permission to act either. That also depends on what a mistake would cost.

Articles · 6 min read

What is a deal architect?

A deal architect is an operator who designs a high-stakes deal as a structure, not as a persuasion contest. The work is power, alignment, constraints, terms, and future options. Optimizing one visible number (price, close date, headline valuation) without those five is not architecture. It is a local maximum. The named system on this site is PACT. The book The Deal Architect is in production and not for sale. The definition is already usable.

Articles · 8 min read

What is a dominated AI model? Better and cheaper is only half the test

A dominated AI model is one where another model scores at least as high on the same quality test, costs no more for the same kind of work, and is better on at least one of the two. Undominated.ai, an independent AI price index, adds a third check: the alternative must also cover what the first model can do, such as context length, output length, and input types, or the index names what you would give up. The verdict tells you what to test, not what to buy.

Articles · 6 min read

What is a knowledge graph? Named nodes, typed edges, a walk you can audit

A knowledge graph is a store of named nodes and typed edges: this account depends on that service, this claim cites that source, this person owns that system. You query it by walking relations, not by hoping a similar paragraph appears. It is not a mind map, dashboard, or vector index. RAG fetches a passage. A graph stores how things connect. Microsoft's GraphRAG is a graph-index method for global questions across a corpus; its GitHub project is largely in maintenance mode as checked 2026-08-27. That method is not the definition of a knowledge graph. In a company the useful graphs are boring: ownership, dependencies, evidence. Keep the edge vocabulary closed and give every edge an owner and a date, or it will go stale while sounding certain. If documentation cannot answer 'what breaks if this changes,' the missing object is often an edge.

Articles · 3 min read

What is a model cascade? Escalate with a stop rule

A model cascade is a planned chain of models for one task: start at the cheapest tier that could do the job, escalate to a more capable model only when a written trigger says the output is insufficient, and stop when a written rule says stop. A fallback is different. A fallback is recovery after the primary model has already missed a latency, quality, or availability target. Retrying the same flagship model is neither. On this site a cascade is a routing pattern inside The Model Portfolio, which is live. It is not the CASCADE framework from The Second-Order Thinker, a different book that is still a manuscript and not for sale.

Articles · 2 min read

What is a personal health agent? Data architecture, not a clinician

A personal health agent, on this site, is a non-medical data architecture: it organises signals you already have, detects change against a baseline you named, and surfaces questions worth taking to a professional. It is not a clinician, not a diagnosis engine, and not treatment. The named system in AI Agents for Personal Health is AGENT: Assemble, Generate, Evaluate, Notify, Tune. That book is in design and not for sale. This page owns the search 'the anatomy of a personal health agent' as an operating object, not as medical advice. If you need care or have health concerns, consult a qualified healthcare professional; this page is not medical advice.

Articles · 5 min read

The first screen an answer engine can quote

This page is the craft. The customs collision lives on AEO meaning. AEO, answer engine optimization, is the craft of writing a page so an answer engine can quote a clean capsule: a direct answer near the top, named entities, visible sources, and no invented FAQ schema. GEO is the system around the crawl. AEO is the page. They fail together when the first screen is a brand manifesto. Authorized Economic Operator is a different AEO, in customs. Google ended FAQ rich results on 2026-05-07; a fake FAQPage is a liability, not a tactic. If the first screen cannot be quoted, the rest of the page will not save it.

Articles · 2 min read

What is agent orchestration? Coordination, not a graph demo

Agent orchestration is the assignment of work across specialised human and AI workers so ownership, handoffs, review, and state stay visible. Four artefacts make it inspectable: a named owner, a handoff packet, a review gate, and a shared state surface. A runtime graph can implement routing, but it does not establish those controls by itself. CHORUS is one named protocol for this operating layer. Why multi-agent systems fail owns the coordination gap. The Multi-Agent Organization is live.

Articles · 2 min read

What is agent strategy? Selection, sequence, exposure, exit

Agent strategy, on this site, is the portfolio choice above any single worker: which jobs merit agents, which dependency must be proven first, how much aggregate send, spend, delete, or speak-as-the-company authority the portfolio may hold, and what evidence pauses or retires a worker. What is an AI agent owns the worker. Using AI agents effectively owns the single-worker charter, ceiling, and run log. Why multi-agent systems fail owns coordination. This page owns selection, sequence, aggregate exposure, and exit.

Articles · 2 min read

What is AI team dynamics? Mixed teams, human accountability

AI team dynamics, on this site, is the social operating system of a mixed human and agent team. Productivity is allowed to compound only if accountability stays human. The named system is SYNERGY: Shared intent before tools, Yield to strengths, Negotiate interfaces, Evaluate loops, Review as operating rhythm, Guardrails and kill-switches, Yield ownership, not control. It is not a vibe about collaboration, not a seat-count of bots, and not a replacement plan. Why multi-agent systems fail owns the coordination gap. This page owns the query AI team dynamics. The AI Team Dynamics is in production and not for sale.

Articles · 5 min read

What is an AI agent? A definition founders can use

An AI agent is a system that takes a goal, uses tools, and acts across more than one step without a human prompting each one. A chatbot answers a turn. An automation runs a fixed script. An agent is only accountable when it has a charter: inputs, outputs, an authority ceiling, and a review gate. Without those, you have a fluent wrapper, not an operator you can put on a roster.

Articles · 3 min read

What is an AI agent charter? Seven fields or it is a demo

An AI agent charter is the written contract that makes an agent accountable. Seven fields, the same ones used for human operators in CHORUS: purpose, inputs, outputs, authority ceiling, success criteria, review requirement, and escalation. An AI agent takes a goal, uses tools, and continues across steps. Without the charter you cannot say what it may not do, which means you cannot hand it a credential. The refuse-list lives next door in override doctrine: money, identity, irreversibility. This page owns the seven fields. What is an AI agent owns the worker. Using AI agents effectively owns the operating loop.

Articles · 7 min read

What is an evidence ledger?

An evidence ledger is a claim-level register: what was asserted, which source supports it, how strong that source is, and what would retire the claim. It is not a bibliography dumped at the end, not a reading list, and not a folder of PDFs. The unit is one claim, not one source. The grade is a ceiling on the verb: an internal export can carry "in our data, this quarter" and cannot carry "in the market." The retire condition names something you would trip over in normal work, not "if it turns out to be wrong." Operators keep one so a number cannot wander between a chapter, a pitch deck, a back-cover line, and an answer-engine capsule. If the claim cannot point at a row, it is not ready to print, pitch, or paste.

Articles · 6 min read

An MCP connector is the plug. Permission is a separate ceiling.

An MCP connector is the attachment between a host (ChatGPT, Claude, Gemini, Grok, a desktop app, a CLI) and an MCP server. The server exposes tools, resources, and prompts. The host opens one client per server. Local connectors usually start a process over stdio. Remote connectors speak HTTPS (Streamable HTTP or SSE) so a cloud host can reach the server. ChatGPT custom MCP is remote-only. Claude Desktop can run local stdio servers from a config file. Grok custom connectors need a public HTTPS URL; localhost must be tunnelled. Gemini managed agents and Gemini CLI can attach remote MCP servers. Start first-connect diagnosis with transport, restart, and auth, not the model. Menus move. Recheck first-party docs. A connector without visibility, mutation, and a log is still a hole.

Articles · 4 min read

What calibration means: matching confidence to evidence

Calibration, on this site, is the practice of matching the strength of a judgment to the strength of the evidence, then updating when the facts move. Confidence is a report, not a personality trait. A loud claim with a thin source is a calibration failure, not conviction. The work is to score the claim, write what would change your mind, and revise the score when the evidence moves. It is not camera calibration, not model-weight calibration as a product claim, and not a psychometric test you can buy. The named system is CALIBRATE, in The Calibrated Mind, live in English and Dutch paperback and hardcover.

Articles · 6 min read

What is CHORUS? Coordination, not a choir

CHORUS is a six-skill coordination protocol for teams of 2 to 15 people who share work with specialised AI workers: Charter, Handoff, Orchestrate, Review, Update, Sync. A charter is seven fields. A handoff is a six-field packet. They are not the same artifact. It is not a choir, not an alphabet song, not CrewAI, and not a graph of model calls. If the receiver reconstructs the job from a chat log, you have a paste, not coordination. The book that owns the name is The Multi-Agent Organization, which is live.

Articles · 9 min read

What is citeturn0search0? The ChatGPT leftover that shows an answer was copied without its sources

citeturn0search0 is a ChatGPT citation marker that survived a copy and paste. Inside ChatGPT, that spot is a clickable link to a web page the answer used. Pasted elsewhere, the link is lost and the code stays. Read against OpenAI's developer guide to the same format, turn0 is the first tool call (such as a web search) and search0 is the first source it returned. The string shows a passage came from a search-backed ChatGPT answer. It does not show which page, or whether the claim is true.

Articles · 2 min read

What is founder resilience? Capacity, not a mood

Founder resilience is trained capacity under load, not a mood that returns after a holiday. On this site the named model is RESILIENT, from The Psychology of Founder Resilience: nine capacities you train and budget, which are reappraisal, emotional granularity, somatic regulation, identity flexibility, learning orientation, interpersonal support, executive function, narrative integration, and titrated exposure. The four the book puts in public first are sleep, recovery, decision quality, and the social cover that lets a founder tell the truth. RESILIENT is an authorial dashboard, proposed and applied, not a validated clinical scale and not medical advice. The book is live. This page is the name. The capacity article is the spec.

Articles · 2 min read

What is GRAIN? The inspectable retrieval loop

GRAIN is a retrieval loop with five named moves from The RAG Engineer: Gather corpora, Rank and rerank, Assemble context, Inspect failures, and Navigate freshness. If you cannot inspect the retrieved and supplied passages, you do not have inspectable retrieval. You have a fluent guess. It is not a metaphor for grain size, not a cereal brand, and not the sentence 'we added a vector store.' RAG names the architecture. GRAIN names the production boundaries. The RAG Engineer owns the chapter-length build and is live in English and Dutch paperback, hardcover, and Kindle. This page is the letters.

Articles · 3 min read

What is hybrid search? Two indexes, one inspectable hit

Hybrid search is retrieval that runs two signals over the same corpus, a sparse lexical index such as BM25 and a dense vector index, computes each independently, and fuses the two rankings before the model writes. Lexical search catches exact identifiers, error codes, and rare names. Vector search catches paraphrase. Either alone drops a class of honest questions, and 'we added embeddings' is not hybrid. In GRAIN, the loop from The RAG Engineer, hybrid retrieval lives in Rank, and it does not excuse Inspect: if you cannot open the passage that was supplied, the fused score is decoration. The RAG Engineer is live. This page is the split, not the full loop.

Articles · 8 min read

What is LEVEL? A five-step check before you act on a claim

LEVEL is a five-step check for deciding how far to rely on one claim before you act on it. Locate the claim in one checkable sentence. Estimate what being wrong, or waiting, would cost. Value the evidence: what kind it is and how good. Expose the gap between that evidence and what the decision needs. Lean accordingly: act only as far as evidence and stakes allow, with an owner and a review date. It comes from the book The Evidence-Aware Life.

Articles · 5 min read

What is LLM grounding? The sentence has to open a folder

LLM grounding means that every factual claim in a generated answer is supported by specific evidence the model was given for that answer. The support can come from one passage or from several passages read together, and the evidence must have been present in the generation-time context. A valid source link is not enough on its own: the cited passage has to entail the claim. Terminology varies by vendor. Google's Search guide calls retrieval-augmented generation, or RAG, a technique also known as grounding, while Google Cloud describes RAG as a commonly used grounding technique. This site keeps an operational distinction: RAG retrieves and assembles evidence, and grounding tests whether the answer follows from it.

Articles · 4 min read

What is llms.txt? A manifesto, not a ranking switch

llms.txt is a proposed Markdown file, named llms.txt, published at a site's root or at any path within it. It gives agents and retrieval clients a short description of the site or section plus curated links to more detailed, preferably machine-readable pages. Reading it is voluntary. The only required element is an H1. It is not robots.txt, not an access control, and not a replacement for a sitemap. Google's guide to generative AI features in Search, last updated 2026-07-10 and re-read 2026-09-04, says Google Search does not use such files and that creating one neither helps nor harms visibility or rankings, including in generative features such as AI Overviews and AI Mode.

Articles · 3 min read

What is model routing? A policy, not a default

Model routing is a written policy that sends each task to a named model for a named reason. The objective is typed first and the model is chosen second. A route you can defend carries a mandate, a cost class, a privacy class, a latency class, and a fallback. Sending every call to the flagship model because last quarter's demo did is a default, and a default is a habit with a bill attached. On this site the discipline is ROUTE, from The Model Portfolio: Register, Objective typing, Utilize policy, Track, Evolve. The book is live in English and Dutch paperback, hardcover, and Kindle. This page is the definition. The book owns the operating loop.

Articles · 4 min read

What is RAG? When AI looks up a source before it answers

RAG, retrieval-augmented generation, is an AI method. A language model does not magically know your contracts, handbook, or last Tuesday's report. RAG makes it look up a specific passage in a collection you gave it, then write the answer from that passage. Retrieval is the product. The sentence is the last mile. If you cannot open the source behind the sentence, you do not have RAG. You have a chat window with extra text pasted in.

Articles · 8 min read

What is ReasonKit? Structured decisions through MCP

ReasonKit Think is a free, open-source add-on for AI coding assistants. It helps the assistant turn a decision into a map of options, test the important claims against evidence you or the assistant supply, stop when critical evidence is missing, and export a record of how the decision was reached. It connects through MCP (Model Context Protocol), the open standard for linking AI apps to outside tools, and runs on your own machine. Len P. van der Hof, who writes this site, built it.

Articles · 8 min read

What is ROSTER? How one person manages a team of AI agents

ROSTER is a framework from The Agent Dream Team for a founder who works alone but runs several AI agents. It names six decisions to make for every agent: Roles (one job with written limits), Objectives (a result you can grade), Skills and tools (only the access that job needs), Triggers (the event that starts it), Evaluation (a written scorecard) and Rotation (when to upgrade, merge or retire it). It is a management method, not software.

Articles · 5 min read

What is SENSE? Five moves, not another unread summary

SENSE is an authorial five-phase design grammar for strategic agent systems, introduced in AI Agents for Startup Strategy: Scan, Evaluate, Navigate, Simulate, Execute. A strategic intelligence layer is the standing system that turns signals into sourced claims and brings them to a decision cadence; SENSE is one way to design and diagnose that system. Scan collects against a strategic thesis. Evaluate turns a signal into a sourced claim with stated confidence. Navigate changes what happens next. Simulate rehearses plausible futures and their signposts. Execute delivers the work to an owner, a deadline, and a decision moment, usually as prepare and present rather than autonomous action. SENSE deliberately has no Decide phase: a person keeps judgment. Deliberately partial systems are valid when the missing phase and the human handoff are explicit.

Articles · 5 min read

What is STACK? Five layers, not a vendor diagram

STACK is a five-layer architecture for repositories where humans and agents share work: Structure, Toolchain, Agent configuration, Connection, Knowledge and quality. Structure is the map an agent can navigate on minute one. Toolchain is the commands it may run. Agent configuration is versioned standing context. Connection is MCP with contracts and least privilege. Knowledge and quality is memory, evals, and CI, so a model upgrade does not silently lower the bar. It is not a vendor 'AI stack' diagram. The book that owns the name is The Agentic Codebase, which is live.

Articles · 8 min read

What is TRACE? Five steps to check research before you act on it

TRACE is a five-step method for checking research before someone acts on it: Target (tie the question to a decision), Retrieve (get the actual source documents), Assess (judge what each source can prove for this exact claim), Cross-check (compare sources claim by claim and trace them to their origin), and Export (hand over a memo that separates what is supported from what is still open, with an owner). It comes from The Verified Research Loop. It shows what the evidence supports; it does not prove a conclusion true.

Articles · 7 min read

When not to use an AI agent: five signs, and what to use instead

Do not use an AI agent when you can write down every step in advance, when a single prompt does the job, when nobody can say what a correct result looks like, when you cannot limit what it can touch, or when a mistake cannot be undone and no person approves first. Use a script, one model call, a draft-only assistant, or a person instead. An agent, a system that chooses its own next step and tools, earns its cost only when the next step depends on what it finds.

Articles · 6 min read

When to use RAG: named passages, not a vector store

Use retrieval-augmented generation when a language model must answer from a private or changing corpus and you need to open the passage behind the sentence. Skip building a RAG system when an authoritative lookup or one-off document read will do, or when the durable question is how entities relate. Before choosing an architecture, set two preconditions: how fast a changed document becomes answerable, and whether retrieval enforces who may see what. Naive retrieve-then-generate, hybrid lexical plus dense search, agentic retrieval as a tool, and graph-index methods such as Microsoft's GraphRAG are different jobs, not a ladder of sophistication. Effective means the answer uses the current authorised version, opens the supporting passage, and refuses when support is absent. The RAG Engineer is available now.

Articles · 7 min read

Why am I not mentioned in ChatGPT?

ChatGPT does not name you for operating reasons, not because the model likes someone else, and there is no published citation ranking to buy. ChatGPT Search can only quote a public URL that OAI-SearchBot may fetch; GPTBot is training, not Search. Test in a fresh Search chat, more than once. Log three outcomes: absent points to access, linked but unnamed points to weak entity cues near the claim, and apparent reuse without credit is a lead to check, not proof.

Articles · 6 min read

Why multi-agent systems fail: the coordination gap

Most multi-agent systems fail as coordination, not as models. A 2-to-15-person team adds specialised agents faster than it adds charters, handoff packets, review gates, and a shared view of state. Org charts name reporting lines. They do not name who holds context at the boundary. CHORUS is one protocol for that gap: Charter, Handoff, Orchestrate, Review, Update, Sync. A graph of model calls is not orchestration.

Projects

MindSesh

Browser-based, neurobiologically grounded sessions for focus, NSDR, breathing and stress recovery.

Projects

ReasonKit

A Rust-based reasoning engine for structured, auditable and verifiable AI workflows.

Projects

Undominated.ai

Independent AI inference price index. Capability first, then price. Names the models that are both worse and dearer. No cut of inference.

Concepts

AEO

AEO, answer engine optimization, is the craft of writing a page so an answer engine can quote a clean capsule: a direct answer near the top, named entities, visible sources, and no invented FAQ schema. GEO is the system. AEO is the page. They fail together when the first screen is a brand manifesto. Google ended FAQ rich results on 2026-05-07; a fake FAQPage is a liability, not a tactic. If the first screen cannot be quoted, the rest of the page will not save it.

Concepts

Agent orchestration

Agent orchestration is the assignment of work across humans and specialised agents so that handoffs, reviews, and state stay visible. A graph of model calls is not orchestration. Orchestration is who is chartered, what packet travels, what must be reviewed, and who sees the board. Without those four, adding agents adds waiting. CHORUS is one implementation of that operating layer. If you cannot point at the owner, the packet, the gate, and the board, you have a demo.

Concepts

AI agent

An AI agent is a system that takes a goal, uses tools, and acts across steps without a human prompting each one. A chatbot answers a turn. An automation runs a fixed script. An agent is accountable only when it has a charter: purpose, inputs, outputs, authority ceiling, success criteria, review, and escalation. Without those seven fields, it is a demo that can spend money. A wrapper around a chat model is not this. Neither is a scheduled script.

Concepts

BooksOS

BooksOS is a gated production system for serious nonfiction: Brief, Bible, Map, Draft, Edit, Package, Launch. Agents do specialist work inside the gates. Judgment, voice, and accountability stay human.

Concepts

CALIBRATE

CALIBRATE is the practice of matching the strength of a judgment to the strength of the evidence, then updating when the facts move. Confidence is a report, not a personality trait. A loud claim with a thin source is a calibration failure, not conviction.

Concepts

CASCADE

CASCADE is the personal cognitive companion to DOWNSTREAM: trace the next bend before you commit, in the meeting, in your own head. DOWNSTREAM is organizational. CASCADE is the habit that makes the protocol fire when nobody is watching. It is the pause that names who pays later, on a different surface, before you say yes. The book that owns this name is still in production; the habit is already usable. A first-order win with an unnamed later bill is not a decision. It is a delay.

Concepts

CHORUS

CHORUS is a six-skill coordination protocol for teams of 2 to 15 people who share work with specialised AI workers: Charter, Handoff, Orchestrate, Review, Update, Sync. Org charts name reporting lines. CHORUS names coordination lines.

Concepts

CITATION

CITATION is an eight-part framework for becoming a source answer engines name: Cite, Identify, Triangulate, Anchor, Track, Iterate, Optimize, Normalize. Search eligibility is not a citation. Ranking in ten blue links is a different job. Answer engines quote the first clean capsule they can verify. Invented FAQ schema is not a shortcut; Google ended FAQ rich results on 2026-05-07. The book that owns this name is still being drafted; the framework is already how this site is built. A mention without a checkable source does not count.

Concepts

DOWNSTREAM

DOWNSTREAM is an operational protocol for consequences that arrive later, on a different surface, with no owner in the room: Detect, Observe, Weigh, Navigate, Stream effects, Trace, Review, Effects on stakeholders, Assess systemic risk, Mitigate. It is an operating heuristic, not a validated risk instrument.

Concepts

Evidence ledger

An evidence ledger is a claim-level register: what was asserted, which source supports it, how strong that source is, and what would retire the claim. It is not a bibliography dumped at the end. Operators use it so a number cannot wander between chapters, books, or a back-cover line. If the claim cannot point at a row, it is not ready to print, pitch, or paste into an answer. The ledger is the product. The reference list is a rendering.

Concepts

GEO

GEO, generative engine optimization, is the work of becoming a source that answer engines can retrieve, interpret, extract, and verify. It is not a bag of hidden prompts. Eligibility, extractable answers, original value, and evidence that survives scrutiny are the levers. Ranking in ten blue links is a different job. Google does not require schema, chunking, or an llms.txt file for AI Overviews. A page that cannot be quoted cleanly will not be cited, even if it ranks.

Concepts

GRAIN

GRAIN is a retrieval loop: Gather, Rank, Assemble, Inspect, Navigate. Gather governs the corpus: sources, permissions, and versioning. Rank retrieves candidate passages and reranks them for the question actually asked. Assemble builds the context the model may see. Inspect traces a wrong answer back to the chunk and retrieval decision behind it. Navigate keeps the index fresh: TTL policy, re-ingest, and change detection. If you cannot point at the passage, you do not have retrieval. You have a fluent guess.

Concepts

Human in the loop

Human in the loop means a named person holds an authority ceiling and a review gate. It is not a human glancing at a dashboard. The human charters the work, can stop a tool call, and owns the outcome. If no one can fail the output against written criteria, the loop is decorative. A vendor slide that says HITL without a named reviewer is a slogan. The name, the ceiling, and the fail criteria have to live in the runbook, not in the deck.

Concepts

LEVEL

LEVEL is a five-move check for grading a claim before you change a belief, a habit, a roadmap, or a bet: Locate, Estimate, Value, Expose, Lean. Locate names the claim in one sentence. Estimate the stakes of being wrong. Value the evidence: name the rung it actually sits on, not the rung the packaging implied, then check its quality within that rung. Expose the gap between that support and what the decision requires. Lean with a next step proportionate to that gap and those stakes. The Evidence Ladder's seven rungs are kinds of support, not a quality scale. Confidence is not a grade.

Concepts

MCP

MCP, the Model Context Protocol, is a contract for how a model calls tools: what it may see, what it may change, and how the call is logged. The official split is host, client, and server, with tools, resources, and prompts as the primitives. Auth is optional; least privilege is not. A server without a contract is a hole in the repo. Founders should care about the interface and the permission, not the logo on the protocol. A clean server with no charter is still a hole with nicer plumbing.

Concepts

Override doctrine

Override doctrine is the written list of decisions an agent may never take: spending above a ceiling, speaking as the company, changing production data, or closing a judgment the founder still owns. It is the complement of a charter. The charter says what the agent does. The doctrine says what it must refuse. A refuse-list that lives in a Slack thread is not doctrine. It is a hope. Write it before the first tool that can send, spend, or delete.

Concepts

RAG

RAG, retrieval-augmented generation, means the model must fetch named passages before it answers. The retrieval is the product. The generation is the last mile. If you cannot inspect which chunk was used, you do not have RAG. You have a language model with extra context stuffed in the prompt. A vector store is a component. GRAIN is the loop that keeps the answer honest. Adding embeddings is not a finished system. Naming the passage is.

Concepts

RESILIENT

RESILIENT is nine trainable capacities for founder pressure: reappraisal, emotional granularity, somatic regulation, identity flexibility, learning orientation, interpersonal support, executive function, narrative integration, and titrated exposure. Resilience is infrastructure, not a personality trait.

Concepts

ROSTER

ROSTER is a six-skill protocol for a one-person company that runs named specialist agents: Roles, Objectives, Skills and tools, Triggers, Evaluation, Rotation. Roles give each agent a charter. Objectives make the charter measurable. Skills and tools bound what the agent may use. Triggers decide when it runs. Evaluation scores output instead of vibe. Rotation retires, merges, upgrades, or migrates a role on scorecard evidence. One chatbot that never hands off is not a team.

Concepts

ROUTE

ROUTE is a five-skill discipline for governing a mix of language models: Register, Objective typing, Utilize policy, Track, Evolve. Register catalogs every model with mandate, cost class, privacy class, and retirement status. Objective typing names what a task actually requires before a model is chosen. Utilize policy writes the routing rules, cascades, and fallbacks. Track measures cost, latency, and quality per route. Evolve promotes, demotes, and retires on a cadence. A leaderboard pick is a collection. ROUTE is a portfolio.

Concepts

SENSE

SENSE is a five-move strategic intelligence layer: Scan, Evaluate, Navigate, Simulate, Execute. It turns agent work from a productivity shortcut into a sourced claim that changes a watchlist, an assumption register, or a Monday decision.

Concepts

STACK

Concepts

Strategic intelligence layer

A strategic intelligence layer is the standing system that turns market, competitor, and internal signals into sourced claims a founder can act on. It is not a dashboard wall and not a weekly PDF. The test is whether a claim changes a watchlist, an assumption, or a Monday decision with a confidence grade attached. SENSE is one implementation: scan, evaluate, navigate, simulate, execute. If competitive work is still a founder reading Slack at midnight, you do not have a layer. You have a habit that will miss a week.

Concepts

Agent operations

How to run and retire production agents so credentials, triggers, and spend stop when the agent should stop.

Concepts

Agentic Codebase

Repository architecture that keeps AI-agent work production-safe: layers, review gates, and durable interfaces that survive real deployments.

Concepts

AI Agents

Architectures and operating patterns for agentic AI systems that founders and engineers can deploy without cargo-cult complexity.

Concepts

Downstream Effects

Delayed, displaced consequences of operational decisions: second-order effects, partner and regulatory landing surfaces, and the audit that makes the next bend visible.

Concepts

Founder Resilience

Operator-grade resilience systems for high-stakes founders who need clear thinking under pressure, not motivational slogans.

Concepts

Marketing

Choice architecture, persuasion literacy, and honest commercial practice.

Concepts

Multi-agent organization

Small teams where human operators and specialized AI workers share goals, handoffs, and review gates.

Concepts

Positioning

Positioning as a decision about who you refuse and what you will not claim.

Concepts

Undominated.ai

The independent AI inference price index.

Concepts

TRACE

TRACE is a five-step research loop: Target, Retrieve, Assess, Cross-check, Export. Target writes the research brief before any search runs. Retrieve obtains the source bodies, not one confident pass. Assess appraises what each body can bear per claim; tiers are source roles, not credibility weights. Cross-check records what each source supports or contradicts, and maps shared origins. Export is a versioned decision memo a co-founder can act on. Fluent collection is not research. TRACE is the verification structure.

Books

AI Agents for Personal Health

Deploy Intelligent Systems to Monitor, Analyze, and Optimize Your Biology

Books

AI Agents for Startup Strategy

How Founders and Operators Deploy Agentic AI to Move Faster and Think Smarter

Books

Decision-Making Under Chaos

A Practical Toolkit for High-Stakes Decisions Under Uncertainty

Books

Leadership Under Load

Awareness-Based Practice for Founders and Operators Under Sustained Pressure

Books

Marketing Without Manipulation

How to Sell with Integrity in a World of Persuasion Tactics

Books

ML for Agent Builders

Auto-Training, Fine-Tuning, and Eval Loops for LLM-Powered Systems

Books

ML for Strategic Founders

Read Models, Not Math - Decision-Grade Machine Learning for Operators

Books

The Multi-Agent Organization

How Humans and AI Agents Share Work Without Losing the Plot

Books

The Psychology of Founder Resilience

How High-Stakes Leaders Stay Grounded, Clear, and Effective Under Pressure

Books

Stoic Systems for Decision-Making

A Practical Framework for Clear Judgment Under Pressure

Books

Strategic Solitude

How Deliberate Alone Time Sharpens Decision-Making

Books

The Agent Dream Team

Build Your Autonomous Specialist Roster for a One-Person Company

Books

The Agentic Author

The BooksOS Method for Writing Serious Nonfiction with AI Agents

Books

The Agentic Codebase

The Five-Layer Architecture That Makes Agent Work Survive Production

Books

The AI Risk Register

Build a Living Register That Names, Owns, and Tests Every AI Risk Before It Names You

Books

The AI Team Dynamics

Designing Mixed Human and Agent Teams So Productivity Compounds and Accountability Stays Human

Books

The Attention Architect

Reclaiming Cognitive Sovereignty by Delegating Distraction to AI

Books

The Authentic Operator

Voice, Identity, and Trust When AI Writes With You

Books

The Calibrated Mind

How to Hold Judgments at Their Right Strength Under Uncertainty

Books

The Cognitive Budget

Spend Focus, Judgment, and Emotion Like Capital

Books

The Constraint Founder

Stoic Limits for Bootstrapped Ambition

Books

The Crisis Command

COMMAND: Calm, Assess, Map, Mandate, Unify, Notify, Decide. The Trained Response

Books

The Deal Architect

Design High-Stakes Deals That Create Value and Protect Optionality

Books

The Delegated Mind

Stoic Judgment in the Age of AI Delegation

Books

The Downstream Effects

Trace the Delayed Consequences of Your Decisions Before They Cost You

Books

The Evidence-Aware Life

How to Think in Levels Instead of Opinions

Books

The Findability Engine

How to Be Cited, Trusted, and Discovered in the Age of AI Search

Books

The Founder CFO

Cash, Metrics, and Calm Between Rounds

Books

The Founder Sales System

Pipeline, Discovery, and Deals Before You Have a Sales Team

Books

The Fundraising Operator

Term Sheets, Power, and Calm Under Dilution

Books

The Human-in-the-Loop Life

How to Stay Essential in an Age of AI Automation

Books

The Identity Forge

The Operating System for Identity That Survives Growth

Books

The Judgment Gym

Maintain the Thinking That Makes You Worth Hiring When AI Thinks Faster

Books

The Knowledge Scale

Engineer Organizational Memory So Your Company Stops Forgetting What It Already Learned

Books

The Meaning Operator

Applied Existentialism for Founders: Why You Build, What It Costs, What It Is Worth

Books

The Model Portfolio

How to Route, Mix, and Govern LLMs Like a Strategic Asset Allocator

Books

The Negotiation Operator

Evidence-Based Deal-Making for Founders

Books

The Org Memory

Build a Company That Remembers What It Learns

Books

The Performance Integrator

One Integrated Protocol for the Body, Mind, Energy, and Team That Runs Your Company

Books

The Predictive Founder

When to Trust Your Gut, When to Override It, and How to Keep Score

Books

The Pricing Mind

Behavioral Economics of Price: How Buyers Experience Value and How to Design That Experience

Books

Proof in Public

Trust Velocity for Evidence-First Founders

Books

The RAG Engineer

Retrieval Systems That Agents Can Actually Trust in Production

Books

The Red Team Founder

Adversarial Thinking for Operators Who Cannot Afford Surprise

Books

The Scale-Up Operator

Build the Operating System That Runs the Company Without You in Every Room

Books

The Second-Order Thinker

Trace the Full Cascade of a Decision Before You Commit to It

Books

The Experimental Founder

N=1 Protocols for Cognitive and Physical Performance

Books

The Simulation Table

War Games for Teams Too Small for a Strategy Department

Books

The Talent System

Build a Talent Operating System That Compounds in the Agent Era

Books

The Turnaround Protocol

RESCUE: Diagnose, Stabilize, Cut, Unify, Execute or Exit with Dignity

Books

The Verified Research Loop

Deep Online Research for Entrepreneurs Using AI Without Trusting It Blindly

The knowledge atlasLEN P. VAN DER HOF

Featured research

Start with the strongest signal.

AI Systems

Introducing Undominated.ai: the independent AI inference price index

Best first. Then price. The board that names every model something else beats without a trade-off.

Len P. van der Hof 25 August 2026

Latest

Recent articles.

SEO, GEO and AEO

Is there an ideal word count for AI citations?

Your page has a length. The passage that answers the question has to make sense without the rest of it.

Len P. van der Hof 16 October 2026

Founder Performance

How to run a pre-mortem in 20 minutes

Asking "what could go wrong?" turns every doubt into a challenge to the plan. Declaring "it has failed" makes explaining the failure the task. The result is a list of risks, not proof of a better decision.

Len P. van der Hof 16 October 2026