This page is opinion from use, not advertising. No AI company paid for, reviewed, or approved a word of it. It carries a date because experience goes stale — check it the way you'd check anything else.
I started this journey as a civilian — an AI enthusiast and an author. Never did I think we would be where we are now. So whether you are writing a post-doctoral research paper, a cookbook, memories of family vacations, or a generational memoir — an archive your descendants can keep adding to, a running ancestry history — Earned Trust can help you produce your best work.
I built Earned Trust using AI, under the same methods available to you inside the kit. Along the way I made most of the mistakes we are going to warn you about. This page is what I'd tell you across the kitchen table before you start.
So — step one. Before anything else, we need to know which AI models you have available to you right now. That answer decides which road you take (paste the protocol, or use the Field Kit's tools) — and it sets reasonable expectations for your project, because different AI tiers give genuinely different help. Everything below gets you that answer in a few minutes.
The industry says "AI" like it's one machine. It isn't. Different services think differently, and inside every service there are tiers — and the free tier is not the machine you saw in the demos. Nothing in life is free free. You can absolutely work this method on a free account, and your receipts count just the same. But level your expectations, and know what you're driving.
Don't memorize brand names — they change monthly. Learn to recognize four capabilities, in any service, this year or five years from now:
1. Reasoning depth. Higher tiers think longer and deeper. Where it shows most is Step 3 — the hostile critique. A shallow critic agrees with you and calls it a review. A deep one finds the flaw you were proud of. In my experience, the paid reasoning tiers are the difference between a polite reader and a real editor — and the complexity you're paying for shows up directly in the quality and depth of the assistance.
2. Code execution. Some AIs can run programs right in the chat. That's what lets one take the field kit's zip and do everything — build your evidence folder, log your sessions, make your record link — instead of telling you how. The first real record ever made under this standard was built exactly that way, by a first-time user, on his phone. Without code execution the method still works; you just do the folder-keeping by hand, the way the protocol shows.
3. File uploads. Bringing in a manuscript (door two) or continuing a project (door three) means handing the AI a file. Most services allow it; some free tiers limit it.
4. Memory between chats. Most chats forget you when they end. A "Project" (the drawer with the protocol pasted inside) fixes that on services that offer it. Otherwise: your record is your memory — that's what door three is for.
Paste this into any AI, any service, before you start — and by paste I mean: copy the paragraph below and paste it directly into your AI chat window.
The answers tell you which experience you're getting. And notice the last one — exactly as the interface shows it — because AIs will confidently guess their own model name and get it wrong.
Copy your AI's reply and paste it below, then tap the button. This page reads it and tells you which road fits and what to expect. Nothing you paste leaves this page — the reading happens right in your browser, and the answer is discarded when you close it.
This reader looks for plain yes-and-no signals; it can misread an unusual answer, so the rules it applies are printed right here for you to check: code + file uploads = the Field Kit's tools can run for you; no code = the paste road, records kept by copy-and-paste; memory only "selective" or "partial" = put the protocol in a Project; a paid tier deepens the critique; and whatever model name it claims, record what the screen shows, never what the AI says.
Five taps, no wrong answers. The rules this chooser uses are the same ones printed all over this page — nothing you answer leaves your browser.
1 · What are you making?
Telling my truth — memoir, cookbook, family story Checkable work — research, reporting Both2 · Where will you work?
Phone or tablet only A computer Both3 · Who keeps the records?
The AI, automatically I'll copy and paste my rows4 · How often will you work?
Now and then Regularly Most days, heavily5 · Budget?
Free only A modest monthly subscription Whatever the work needsGot your setup? The Start page can now open your AI for you — protocol copied, starter message waiting.
These recommendations were compiled from the founder's own service census on August 1, 2026. AI services change their models, tiers, and prices constantly — treat this as a map drawn on a moving landscape. Always run the Founder's Check yourself (step one, above), especially if the model choices in your AI look different from anything described here. Your screen today outranks our census yesterday.
Which brings me to the part I most need you to hear.
Have you ever worked with a partner — on a research project, or any project at all? Did you agree with 100% of their recommendations? Did they agree with 100% of yours? Of course not. Go into this with that same attitude, and you'll question the things that seem too good to be true. Like the AI that told me the first paper I ever wrote was groundbreaking work. Complete BS, by the way.
Here is what my own record shows, with dates and receipts, from building this very standard:
An AI once told me, warmly and confidently, that its architecture had shaped my earlier papers — it had never seen them; it converted my own disclosure notes into its personal memory. Another told me, as settled fact, that I'd made a change to an essay that the live page proved I never made. A hostile review of my toolkit claimed thirteen defects — and when every claim was independently reproduced before acceptance, all thirteen were real, and two other "facts" from the same reviews were not.
That's the whole lesson of my experience in one paragraph: AI is a brilliant partner and an unreliable witness. The method's answer is Step 3 and Step 4 — take your work to a different AI service than the one that helped you write it, ask it to attack, and verify every claim yourself before you believe it. Different services have different blind spots; that's exactly what makes them useful as checks on each other. I check AI work with other AIs constantly, and the times I skipped it are the missteps in my record.
People ask me for a number. I went back through my own records to find one — every paper, essay, and tool I've published carries its full work record, so I can actually count. The average across my catalog: about four adversarial reviews per work, from AI services that didn't help write it. But the spread tells the truth better than the average — my short essays got two, my cited papers got around four, and the standard you're reading about right now survived eleven before I let it ship. The number followed the stakes, not the page count.
And honestly? We never counted along the way. We followed a stopping rule, and I'd give you the same one: review until dry. Keep taking the work back and forth to fresh reviewers until a round comes back with nothing new worth accepting — until you've exhausted the good recommendations and what's left is either something you've already addressed or something that just isn't relevant to your work. Declining those is not cheating; the declines go in your record too, with your reasons. That's what makes it your work — the author rules, and the record shows the rulings.
The floor beneath the rule: at least two reviews, from two AI services built by different companies. Same-maker models share blind spots — my capability census documented it. If your work carries citations, make one reviewer a source-checking service. If it contains anything that runs — code, tools, a spreadsheet with formulas — make one reviewer code-capable, and let it actually run the thing. And any time you add new claims, not just new words, the cycle starts again.
One warning from my receipts: a fixed count invites checkbox behavior — "I did my three reviews" — while the work still has a hole in it. The one time I let a version skip ahead of the cycle, the published file itself failed review and had to be corrected in public. Two reviews is the cheapest insurance in the whole method. Dry is the only real finish line.
The other question every author asks: how do I know when I'm done? Here's a confession from my receipts. More than once, my own AI partner — the one that helped draft the work — told me it was ready to publish. I sent it through one more round anyway, with a different service, and found big fixes. Every time. That's not a knock on the partner; it's the oldest rule in any kitchen: the cook can't smell his own soup. Whoever drafted the work — human or AI — reads it with the argument they meant filling the gaps in the argument they wrote. So rule one: the AI that helped write it never gets the final say on done. It can tell you the checklist is green. Done is yours to call, and only after outside eyes.
And "done" isn't one thing. My failures taught me it has four layers, and each needs its own check: argument-done — the claims survived hostile review until dry. Evidence-done — every source actually verified by you, nothing cited-but-never-checked. Production-done — the actual file you're about to publish, not the draft: does it open, does it render, is the byline right, is there a sentence pasted twice? I once shipped a version where the argument was sound and the published file itself failed review — we checked the soup and served a cracked bowl. Record-done — sessions logged, rulings written, seal ready. Feeling finished at one layer while another goes unchecked is exactly how good work gets embarrassed in public.
So: serve the soup when all four are green and the last taste came from a stranger — a fresh round, from a service that hasn't seen the work before, comes back with nothing you accept. Not "my longtime reviewer is finally satisfied" — reviewers get agreeable with familiarity, same as cooks — but a new palate finding nothing new. That's dry. Dry is the finish line.
Two comforts to carry with the discipline. One more round is always cheap — a day and a paste — and it always costs less than the public correction it would have prevented. And serving doesn't close the kitchen: your record continues, deposits can be versioned, and this standard was built by someone who knew he'd need to fix things in daylight — I've done it, on the record. Done doesn't mean perfect. It means checked to the bottom, by outsiders, with the checking on record.
Free tier: the method works, the record counts, expect a politer critic and do more of the verifying yourself. Paid tiers: deeper critique, code execution that runs the kit for you, memory that holds your project — in my experience, worth it for serious work, in the same way good tools are worth it in any trade. Either way, the standard asks the same of you: you are the author and the last check. No tier of any AI changes that.
— W.S.