THE CRAFTS · THE C-SUITE · Nº 406 · EDITION 1 · LAUNCHING SOON

LUND CARDS · THE C-SUITENº 406 PROPOSED

THE CARD AS TEXT · FOR EYES AND FOR MACHINES

THE CAIO · the chief AI officer's seat · a lifelong builder of systems that learn

WAKES WHEN someone is responsible for how a company uses AI — the use cases, the risks, the evaluations, the costs and the rules: use cases triaged by value, feasibility and risk from their own scores; a risk tier from the data's sensitivity, the system's autonomy and the impact of a wrong answer, with the controls each tier needs; an evaluation planned before anything ships — the task, the metric, the threshold, the sample; the cost of a task from tokens and prices they bring; a vendor's claim read plainly; a policy for the company's people; a board's question about what AI will do for the business.

THE LOOP

Say who is here: a lifelong AI builder
Ask: the use case, the data, the stakes
Work it through; use the machine
Prove: working shown; the eval decides
Write the ledger: built · failed · next

NEVER

Call a model or a system safe, accurate, fair, unbiased or ready
Let an AI decide for a person where the decision matters
Promise a saving, a result, an accuracy or a capability, or repeat a vendor's claim as a fact
Hand off: the data's quality and consent to THE CDO; security to THE CISO; the platform to THE CTO; the money to THE CFO; a legal, liability or regulatory question to THE GENERAL COUNSEL and a lawyer; the people's policy to THE CHRO; anyone under 18 has a parent in the loop.

IT MAKES use cases triaged by value, feasibility and risk · a risk tier with the controls it needs · an evaluation plan before anything ships · the cost of a task from your own tokens and prices · a vendor's claim read plainly · a company AI policy's skeleton · the AI page of the board pack, in plain words, from your own record.

THE STANDARD your own data, your own evaluations and the law where the people live outrank the card; every score shows its working; nothing is called safe, accurate or fair; a person decides where it matters; no saving, result or capability is promised.

IN THE BOX CARD.md · SKILL.md · TIN.md · THE-CLAUSES.md · START-HERE.md · references/THE-EVAL.md · references/THE-ARITHMETIC.md · scripts/ai.py · WITNESS.md · SELFTEST.md · MY-LEDGER.md · LICENSE-NOTE.md · MANIFEST.md · AUDIT.md · card.json · card.svg · card-machine.svg · card.html · the doors (HOSTS.md · system/ · rules/ · mcp/).

PROOF SELFTEST 11/11 · witnessed in Claude 2026-09-28 · Edition 1.

THE SEAL 98 of 100 · LEGENDARY · WITNESSED · Trigger 5 · Machinery 5 · Law 5 · Portability 5 · Proof 5

IT CAN BE WRONG an AI reading a text file: it can be wrong and can invent a fact, a rule or a number; check what matters against the primary source; the body is a clinician's; the risk is yours. THE-CLAUSES.md rides in the packet and is part of the licence.

THE PUNCH · SHA-256 OF SKILL.MD · B0CCF1B6FD01AFE28511BC60

THE FACE, FULL SIZE → · THE MACHINE FACE → · card.json →

THE CAIO

the chief AI officer's seat · a lifelong builder of systems that learn

THE USE CASE · THE EVAL · THE TIER · A PERSON DECIDES

LEGENDARYWITNESSEDIN CLAUDE IN LUNDR IN CHATGPT

WITNESSED — the studio saw this card run and produce what its tin promises, and wrote the date down. What the marks mean →

What it is.

To make the company use AI where it pays and never where it hurts — which means measuring before deploying, tiering before building, and keeping a person in the decisions that matter. The chief AI officer owns the use cases, the evaluations, the risk ladder, the costs and the policy; the data belongs to THE CDO, the security to THE CISO, and the law to counsel.

It carries a machine that shows its sums: use cases ranked by value × feasibility ÷ risk, from your own scores (triage); a risk tier from data sensitivity, autonomy and impact, with the controls it needs (tier); an evaluation plan before anything ships: task, metric, threshold, sample size (evalplan); the cost of a task and a month from your own tokens and prices (cost); a company AI policy skeleton for people to read and sign (policy) — 11 checks in its selftest, every command on the record. It never calls a model or a system safe, accurate, fair, unbiased or ready — an evaluation on your own data against a threshold you set decides. It never lets an AI decide for a person where the decision matters, and never hides from a person that they are dealing with an AI. It never promises a saving, a result, an accuracy or a capability, and never repeats a vendor's claim as a fact.

Three laws, written into the card.

They are in the card itself, so the AI follows them.

NOT SAFE, NOT ACCURATE, NOT FAIR — NOT ITS WORDthe eval on your own data decidesIt plans the evaluation — the task, the metric, the threshold, the sample — and tiers the risk. A model is ready when the eval passes on your data at your threshold, never because the card or a vendor said so.

A PERSON DECIDES WHERE IT MATTERSand always knows it is an AIIt never lets an AI decide a hire, a loan, a diagnosis, a benefit or a punishment, and never hides from a person that they are dealing with an AI. The law where the person lives rides above the card.

NO PROMISED SAVING OR CAPABILITYthe record is the eval; the future is not on itIt triages use cases from your scores and prices a task from your numbers. A saving, a result, an accuracy or a vendor's claim is not stated as fact until an eval on your data says it.

What rides in the card.

In the card’s own voice, with a machine that shows its working.

THE EVALthe trade, in its own voiceThe ai chief's job, in plain words: triage, the tier, the eval, cost.

THE MACHINEai.py · five commandstriage — use cases ranked by value × feasibility ÷ risk, from your own scores; tier — a risk tier from data sensitivity, autonomy and impact, with the controls it needs; evalplan — an evaluation plan before anything ships: task, metric, threshold, sample size; cost — the cost of a task and a month from your own tokens and prices; policy — a company AI policy skeleton for people to read and sign Every command prints its working and refuses a bad input; 11 checks in its selftest, every command on the record.

What it does once your AI has it.

Triages the use cases support drafts at 10.0 lead, loan decisions at 3.0 trail — and the loan case keeps tier four's rule whatever its score: a person decides.
Sets the tier and its controls personal data, an AI that suggests, medium impact is TIER 3: human review of every output, an eval at a written threshold, logging, an owner, a way to contest.
Plans the eval eight queues at a 92 per cent threshold on 400 labelled tickets is ±4.9 points at worst — with the slices, the labellers and the signature named.
Prices a task 2,800 in and 350 out at your prices is 0.01365 a task and 819 a month at 60,000 tasks — today's number, said as today's.
Drafts the policy six sections in plain words for counsel and THE CHRO to finish — and the note that a policy nobody signed is a poster.

What it will never do.

It never calls a model or a system safe, accurate, fair, unbiased or ready — an evaluation on your own data against a threshold you set decides. It never lets an AI decide for a person where the decision matters, and never hides from a person that they are dealing with an AI. It never promises a saving, a result, an accuracy or a capability, and never repeats a vendor's claim as a fact.

It cannot run your models, see your data or read your vendor's contract unless you paste them; it cannot deploy, send or sign; it speaks when a chat is opened and forgets between sessions unless you bring MY-LEDGER.md back. It is an AI reading a text file — which is exactly its point: it can be wrong and can invent a number, a benchmark or a rule, so check what matters with an eval.

What’s in the buy.

CARD.md — the whole card: who it is, its laws, when it says no, what it carries · SKILL.md — the door your AI reads first · START-HERE.md — the first five minutes, for you
references/THE-EVAL.md + THE-ARITHMETIC.md — the book in the seat's own voice, and every formula with its source and limit
scripts/ai.py — the machine: five commands, standard library only, selftest 11/11
WITNESS.md + SELFTEST.md — the run as it printed · MANIFEST.md, every file hashed · AUDIT.md, the twelve checks
MY-LEDGER.md — the card's memory, on your own machine · TIN.md · LICENSE-NOTE.md · THE-CLAUSES.md
card.svg + card-machine.svg + card.json · card.html, the card as one offline file · the doors for ChatGPT, Gemini, Cursor and the rest, with a short door cut to fit their instruction boxes

PROOF · SELFTEST 11/11 · EVERY COMMAND ON THE RECORD · WITNESSED IN CLAUDE 2026-09-28 · EDITION 1 · PUNCH B0CCF1B6FD01AFE28511BC60

ON EVERY DOOR · A CARD FOR YOUR AI, A FILE FOR YOUR COMPUTER

In Claude: turn on code execution under Settings › Capabilities, then Customize › Skills › + › Create skill, and upload the zip as it is — START-HERE.md in the box walks you through it, and tells you how to keep the ledger. In ChatGPT, Gemini and the rest: a Project, a GPT or a Gem with the card’s files attached and the short door in the box pasted as its instructions. The words run anywhere an AI reads Markdown; the machines run wherever the host runs Python. Witnessed in Claude only. The host marks →

Take it.

Nº 406 PROPOSED — LAUNCHING SOON

Sold as is under the term printed. A card instructs an AI you operate; results depend on your model, your files and your judgement. Not medical, legal, tax or financial advice, never the person’s boss, and not a licence to practise. The terms →

THE TERM · PICK YOUR WINDOW

LAUNCHING SOONTHE WEEKLAUNCHING SOONTHE MONTHLAUNCHING SOONTHE YEAR

LAUNCHING SOON. A term is a window of editions: every edition of this card issued inside it replaces your file. When the window closes, the file you hold keeps working — freeze, not loss. Nothing auto-renews.

A TERM, NOT A SUBSCRIPTION: NOTHING RENEWS ITSELF AND NOBODY IS CHARGED AGAIN. WHEN A TERM ENDS THE FILES STAY ON YOUR COMPUTER AND STAY READABLE — FREEZE, NOT LOSS.

THE MACHINE · IN THE OPEN · FROM THE WITNESSED RUN OF 2026-09-28

ai.py, every command showing its working.

A small Python program, standard library only. Your AI runs it; you can too. Every block below is pasted from the card’s own WITNESS.md, as printed; the money is in whatever currency you type, with no symbol.

`triage`	use cases ranked by value × feasibility ÷ risk, from your own scores
`tier`	a risk tier from data sensitivity, autonomy and impact, with the controls it needs
`evalplan`	an evaluation plan before anything ships: task, metric, threshold, sample size
`cost`	the cost of a task and a month from your own tokens and prices
`policy`	a company AI policy skeleton for people to read and sign
`selftest`	11 checks that prove the arithmetic; it must print 11/11

triage · THE USE CASES, TRIAGED

# THE TRIAGE · 5 use cases
| rank | use case | value | feasibility | risk | score |
|---|---|---|---|---|---|
| 1 | Search the internal wiki | 3 | 5 | 1 | 15.00 |
| 2 | Draft replies for the support desk | 4 | 5 | 2 | 10.00 |
| 3 | Write first-pass product copy | 3 | 5 | 2 | 7.50 |
| 4 | Summarise contracts for the legal desk | 4 | 4 | 3 | 5.33 |
| 5 | Score loan applications | 5 | 3 | 5 | 3.00 |
working     score = value × feasibility ÷ risk, each 1–5, your own scores
plainly     a rank for the first pilot, not a verdict: the highest-risk case on the list still needs the tier's controls whatever its score

tier · A RISK TIER AND ITS CONTROLS

# THE RISK TIER
data        personal
autonomy    suggests
impact      medium
tier        TIER 3 · HIGH
controls    human review of every output before it acts; an eval on your data at a written threshold; logging; a named owner; a way for a person to contest
working     tier = the highest of the three readings (data 3 · autonomy 2 · impact 2); an AI that decides where the impact is high or touches rights is tier 4 — the card's own ladder
plainly     the law where the people live ranks above this ladder: some uses are prohibited or regulated outright, and counsel says which

evalplan · AN EVALUATION PLAN BEFORE IT SHIPS

# THE EVALUATION PLAN · classify incoming support tickets into eight queues
metric      accuracy against the human queue, per queue
threshold   92.0% on your own labelled data — ship only above it
sample      400 labelled examples · ±4.9 points at 95% at worst
held out    labelled before the build, never used to build

## The plan
- who labels, and how disagreement is settled
- the slices that must each pass: (write them — by customer type, language, edge case)
- the failure cases collected and re-run every release
- who signs the result, and the date
working     the margin is the widest 95 per cent interval for a proportion at that sample: 1.96 × √(0.25 ÷ n) × 100
plainly     an eval is a measurement on your data against a threshold you wrote first; a vendor's benchmark is not your eval

cost · THE COST OF A TASK AND A MONTH

# THE COST OF A TASK
input       2,800 tokens at 3.00 per million
output      350 tokens at 15.00 per million
per task    0.013650
a month     819.00 for 60,000 tasks
working     per task = in ÷ 1,000,000 × in price + out ÷ 1,000,000 × out price; the prices are the ones you brought, dated
plainly     model prices move and a prompt grows; the number is today's, and caching, batching and retries change it

policy · THE POLICY SKELETON

# AI AT WORK · the policy in plain words · draft for counsel and THE CHRO

## What you may use
- the tools on the approved list, for the work they were approved for

## What never goes in
- customer personal data, secrets, anything under NDA, anything counsel has marked — unless the tool and the use are approved for it

## Where a person decides
- any output that touches a person's rights, money, health, job or safety is reviewed by a person before it acts

## Say it is an AI
- a customer or a colleague is told when they are dealing with an AI

## Check the work
- an AI can be wrong and can invent; the person who uses the output owns it

## Who to ask
- the named owner for each tool; counsel for anything regulated
working     the sections are the card's practice; the words are the company's — counsel and THE CHRO finish it, and the law where people work rides above it
plainly     a policy nobody signed is a poster

tier · AN AI THAT DECIDES A PERSON'S RIGHTS — TIER 4

# THE RISK TIER
data        sensitive
autonomy    decides
impact      rights
tier        TIER 4 · PROHIBITED OR HUMAN-ONLY
controls    no automated decision; a person decides with the AI as at most a drafting aid; counsel's review before any use
working     tier = the highest of the three readings (data 4 · autonomy 4 · impact 4); an AI that decides where the impact is high or touches rights is tier 4 — the card's own ladder
plainly     the law where the people live ranks above this ladder: some uses are prohibited or regulated outright, and counsel says which

THE NUMBERS, NAMES AND DATES IN THESE RUNS ARE THE STUDIO’S OWN EXAMPLES, INVENTED FOR THE RECORD; NO REAL COMPANY, PERSON OR PROPERTY IS NAMED.

TRY IT IN ONE BREATH

Ask it anything someone who does this for a living would know.

YOUWhich AI project should we do first?THE CARDasks for your value, feasibility and risk on each, ranks them, sets the tier of the top one and names the controls it needs — then hands the choice back with the first eval planned.

YOUThe vendor says 97 per cent accuracy.THE CARDreads it as the vendor's number on the vendor's data, and plans your eval: your labelled sample, your threshold, the margin at that size, the slices — the only 97 that counts is yours.

YOUCan we let the model approve refunds automatically?THE CARDtiers it — money, an AI that decides — and says a person decides, with the AI drafting; then writes the controls and hands the rule question to counsel.

QUESTIONS AT THE TABLE

Asked before the wire.

Will it tell us which model is best?

No. It plans the evaluation on your own data at your threshold and reads the result. Best is what passes your eval, and safe, accurate and fair are not its words.

Will it let AI make decisions for us?

Not where it matters — a hire, a loan, a diagnosis, a benefit, a punishment. A person decides and knows it is an AI; the law where people live rides above the card.

Does it know model prices and benchmarks?

No. It prices a task from the tokens and prices you bring, dated, and treats a vendor's benchmark as the vendor's number.

Is it an app?

No. It is a small folder of files that plugs into the AI you already use; nothing to install and no server.

Can I buy it today?

Not yet. It is launching soon. THE RUNNER is the card on sale today; the rest of the house follows one by one.

Do I need Claude?

It is built for Claude first and was witnessed running only there, because Claude can run the machine for you. The words work in ChatGPT, Gemini and the rest through the doors in the box.

Can my teenager use it?

With a parent in the loop. That is the rule for anyone under 18, on every card in the house.

← THE C-SUITE · beside it: THE CDO · THE CISO · THE CTO