An appendix to It Starts With Safety
Appendix. Revision 1. ~2,600 words. Written 3 August 2026, revised 5 August 2026. A cluster spoke of There is no such thing as AI safety.
The field guide walks through five layers and tells you what a complete version of each contains. It does not explain why there are five, what the writing actually accomplishes, or what the guide is still missing.
This appendix does those three things. The last section changes the guide, and I would rather amend it in public than quietly.
Why five
There is a principled answer and a contingent one, and I want to give both.
The principled part
The five layers answer five different questions that cannot be folded into each other, and they have to be built in order.
What do we believe? That is the Statement. Without it, everything underneath reads as arbitrary paperwork, and the first hard case turns into an argument with nothing to settle it. Most policy failures I have watched were failures of grounding. A rule existed, someone found it inconvenient, and there was nothing to appeal to.
What will we do, and what will we refuse? That is the Policy. A belief you have not turned into a commitment is a feeling. The refusals carry most of the weight here, because a policy that only permits is not a policy.
What is actually true here? That is the Context, and it is different in kind from the others. The other four are decisions. This one is a finding. It is the list of what tools are really in use, by whom, under what vendor terms, and what your data actually is.
It is also the layer people most want to skip, because doing it honestly turns up things nobody wanted turned up. Every engagement I have run has produced at least one uncomfortable discovery here. That is not a failure of the layer. It is the layer working.
What governs the specific situations? That is the Rules. Belief and policy do not tell a staff member whether they can paste a particular thing into a particular tool on a Tuesday afternoon. Rules do. This is the layer people actually use.
What happens when it breaks? That is the Response Plans. This is separate rather than a clause in the Rules because failure is a different mode. Rules describe normal operation. Response describes the hour after normal operation stops.
The order is not tidiness. You cannot write Policy without a Statement to ground it. You cannot write Rules without knowing the Context. You cannot write a Response Plan without Rules naming what should not happen.
The contingent part
These five are a synthesis, not a discovery. They compress what the governance literature says an organization needs into the smallest number of documents a small leadership team can actually finish in two weeks.
Another practitioner might draw the lines at four or six. What matters is that the five questions get answered, and that the answers are written, dated, and ratified. The architecture is defensible. It is not sacred.
What writing does that thinking does not
People tell me they already know all this and their staff are careful. Writing it down feels like paperwork.
Here is what the writing actually does, and these are six different things.
It turns an opinion into a commitment. An unwritten position can change silently. Anybody can revise it, under pressure, with no record that it moved. A written one has to be argued with. This is most of the value, and it is why boards ratify things.
It makes the disagreements surface now. Somewhere in the drafting, two senior people find out they do not agree about something that matters. That disagreement is there either way. Writing it down means the argument happens on a Tuesday in a room, instead of at nine at night while something is going wrong.
It gives everyone the same thing to point at. Twelve people carrying twelve private versions of the policy cannot coordinate. The document is not mainly there to be read. It is there so a conversation has a referent.
It makes delegation possible. A part-time summer staffer cannot read your mind. They can read half a page. Everything you need a nineteen-year-old volunteer to get right has to exist outside your head.
It is the evidence. The legal question about board oversight is about process and documentation, not technical understanding. Good faith that was never written down is close to worthless in that inquiry, and in an underwriting conversation.
And it gives you a refusal you can cite. This is the one people underestimate. A vendor will pitch the thing you refused, persuasively, often to a board member rather than to you. We decided this and the board voted on it ends a conversation that I have concerns does not.
What it costs
The free path costs no money. It costs other things and I would rather name them.
Leadership time. Four to seven people plus one decider, over two weeks, doing unfamiliar work. Twenty to forty person-hours if it is facilitated, and more if it is not, because the drafting is the slow part.
The cost of finding out. The Context layer will surface something. A tool nobody approved. Data somewhere it should not be. A staff member who has been doing something for a year. That discovery has a cost, sometimes a personnel cost, and it is the real reason organizations avoid this layer.
Political cost. Named refusals mean telling somebody no. Deciding where efficiency will not be the criterion means giving up an efficiency somebody wanted. These are real losses and they will be argued about.
Maintenance. Annual review of the Statement, quarterly refresh of the inventory, updates after any incident. Small, ongoing, and the first thing to lapse.
And the cost of doing it badly. A guidebook that is written and then ignored is worse than none, because it sets a standard you are visibly failing to meet. In an oversight inquiry, a policy you did not follow is worse evidence than a policy you never wrote.
That is the strongest reason I know not to start unless you mean to finish, and I would rather say it than sell a document that becomes a problem.
Why it is still hard, even free
The guide is free and complete. Everything in the paid version is in it. So why does almost nobody finish?
Not because of money, and not because of missing information. I spent three and a half years showing this technology to organizations, mostly for free, and watched astonishment produce almost no change in what anyone actually did. Knowing what to do was never the obstacle.
Five obstacles, and money removes none of them.
The blank page. Writing a two-page organizational statement about AI is a kind of writing nobody in a nonprofit has done before. There is no example in the building. People who write grant proposals fluently freeze in front of this.
Nobody owns it. It is not development, not operations, not the executive director alone. Work that belongs to everyone belongs to nobody, and it stays on the list.
It surfaces conflict. When the disagreement shows up, the easiest thing is to defer that section. The deferred section is always the one that mattered.
There is no deadline. Nothing outside the organization forces it to be finished. It is always next month, and next month is genuinely busier.
And it needs a decision, not a document. The writing is the easy part. The hard part is deciding what you refuse. Deciding is unpleasant and it can be put off forever without anything visibly going wrong. Until it does.
Free removes a barrier that was not the binding one. That is not an argument against free. The guide should exist and should be free, because some organizations will use it and they should be able to. It is just an observation about which constraint actually holds.
What the guide is still missing
I researched this properly for the first time in July 2026 and found four gaps in my own work. I am naming them here, adding them to the guide, and adding them to engagements already delivered.
1 · There is no crosswalk to the emerging standard
This is the most important one and the cheapest to fix.
The NIST AI Risk Management Framework appears to be on the same path the NIST Cybersecurity Framework took. Over about a decade, that one became the working definition of reasonable practice, cited by regulators, used by courts to define the standard of care, and used by insurers in underwriting.
Right now a completed guidebook is a well-made document that an outsider has to judge on its own terms. If it carried a page showing which layer answers which part of the NIST framework, it would be legible to an underwriter, a regulator, or a lawyer without translation.
That costs one page. It turns we wrote a careful document into we meet the framework everyone is converging on. To a board, an insurer, or a foundation officer, those are not the same claim.
Adding it, and going back to do it for guidebooks already delivered.
2 · There is no measurement layer
The five layers cover belief, commitment, inventory, rules, and failure. None of them covers how you would know whether any of it is working.
In practice this means a running record. Decisions made and what came of them. Near misses that did not become incidents. A periodic check of AI-assisted work for accuracy. A note of what the Rules got wrong.
Without it, the guidebook is a snapshot of what a leadership team believed during one fortnight. The renewal schedule says to revisit it, but revisiting with no evidence produces the same document with a new date on it.
Adding a sixth layer, or a required appendix. Working name: the Record.
3 · There is no record of who has been trained
Who was trained, on what, and when. That is a question an insurer will ask, and a board member after an incident, and a lawyer.
It is also the honest link between this stage and the ones after it. A guidebook governing people who have never used these tools is a set of rules about a country nobody has visited. That is not a reason to write it later. It is a reason for the record to show when the formation happened.
Adding a competence register. It is a table and it costs nothing.
4 · Nothing talks to your insurance or your contracts
Given what changed in the insurance market this year, this is now a live gap.
The Context layer lists tools and vendor terms. It does not ask whether your current policy excludes this, what your broker said, or what your vendor contracts say about indemnity, rights to train on your inputs, and where your data lives.
Adding an appendix that sets the tool inventory against the insurance position and the contract position. It is one conversation with a broker and one pass through vendor terms. Given that exclusions started landing in January, it may be the most valuable hour in the whole process.
And one I am less sure about
The Rules cover disclosure inside the organization. There is nothing public that tells the people you serve how you use AI.
The case for adding it: transparency rules are exactly what took effect in Europe on 2 August 2026, and American state legislatures tend to start in the same place. Also, an organization whose whole asset is trust should probably say out loud what it does, before somebody asks.
The case against: a public statement is a commitment that can be held against you, and a small organization may reasonably decide it is not ready.
Offering it as optional. The organization decides.
What this does not do
The list I would want a board to have before hiring anyone, including me.
It does not make your use of AI safe. It makes your practice governable.
It does not create insurance coverage. If your policy excludes this, the document does not override it. It may help at renewal. It may not.
It does not prevent an incident. It lowers the odds on the foreseeable ones and shortens the response to the rest.
It does not stay true. It describes one moment, and the moment moves faster than the review schedule. That is the weakness I am least able to solve.
It is not a substitute for forming people. That is what comes next, and the guidebook alone should not be oversold.
Everything described here is free. If you would rather have it facilitated, that is a thousand dollars and two weeks, and the reason it works is not the writing. It is that a date exists, someone is expecting the material, the blank page is gone, and there is an outside person in the room when the disagreement shows up. Both paths produce the same ratified document, and I would rather you finish it yourself than pay me and not finish.
The four additions above are being made to the free guide as well.
Before this publishes
Nothing in this file needs editing. One thing needs deciding.
This appendix commits me to four pieces of work: the NIST crosswalk, the Record layer, the competence register, and the insurance appendix, all added to the free guide and retrofitted to guidebooks already delivered. Publishing corrections to your own product is the strongest thing you can do. Publishing corrections you then do not make is the worst, and it is worse than never having written the appendix.
So each of the four is either confirmed or cut before this goes live. Softening one is not an option, because a softened commitment is the version that quietly does not happen.