MIR is the product I've put the most into, and its whole pitch is evidence you don't have to take on trust. So it gets the hardest read of anything I've measured.
MIR, the Memory Infrastructure Registry, records what people, employees, contractors and AI agents have actually done over time, and shows that record at the moment someone has to decide whether to trust them. It doesn't decide. It records, and the system asking makes the call.
It has three kinds of people to answer to: the organizations that pay and send in records, the people and agents whose history is recorded, and the systems that rely on what it says. I checked each test against all three, using MIR's public pages and its code. Where a page and the code disagree, I've said which is which. Most of what's below is that kind of gap.
Test 1: Leaving is easy
A customer can export everything they put in, in a format another tool can read, without asking anyone.
The people recorded: mostly passes. Someone with an account can download their own record and delete their account themselves, with a confirmation step. The catch is who has an account. A company can record its employees and contractors without them ever signing up, and those people have no way in to see their own record.
Paying organizations: not met. There is a self-serve export on every plan, signed so it can be checked offline, which is good, and it backs up the pricing page's "Export is included." But it only includes events and decisions, not claims or settings, and it identifies people by the registry's internal numbers, not the organization's own. Then the terms say an organization "must delete all MIR data upon termination," which would include the export it just downloaded. And the data processing agreement promises to delete an organization's data within 30 days of leaving. The way the chain works, sealed records are kept so the chain still checks out, so that promise can't be kept as written.
Test 2: Pricing is legible
The price page says what one unit costs and what a unit is.
The page passes. It says what a unit is ("one entity: a person, employee, contractor, AI agent, or service account whose history the registry records"), what it costs, and what's free: up to 5 entities, with no card and no expiry. It also says "it never charges a person to be remembered," and I found nothing that does.
The bill doesn't match the page. Not met. The page says organizations aren't charged for "entities with no activity in a billing period." The code that counts entities for billing never looks at activity. It counts the organization's dashboard members and its active AI agents and service accounts, whether they did anything or not, and it doesn't count the outside people whose history is actually recorded. A discount for early customers is applied in the code but isn't on the page. I didn't check whether anyone has been billed this way yet. Either way, the page and the meter describe two different products.
Test 3: Privacy is the default
There's a public list of every piece of data the product collects, and why.
The design is strong. People are identified by a one-way hash, not a name or an email. The details of what someone was doing when a decision was asked about aren't stored, only the kind of action and the result. Partners decide how long their records are kept.
The list is out of date and incomplete. Not met. The privacy page was last updated December 18, 2025, and the system has changed since. It says IP addresses are kept 30 days in logs. The registry's own audit log kept IP addresses and browser details with no expiry, and kept them even after someone deleted their account, while the page says "everything that identifies you is destroyed." Since this review, that's fixed: the registry now clears them at 30 days and when an account is deleted. The page doesn't mention claims, the text partners write when they file one, the reason a person gives when they dispute one, or the order numbers that can come with a decision. And its privacy and constitution pages carry a chat assistant, Curio, which is also mine, served from another domain. It's not on the list of companies that handle data.
The privacy page also tells people that every lookup of their record is "logged and visible to you." The main way partners check a record, the decision request, didn't show up in that log. Since this review, it does.
Test 4: Accountability is independent
The party being judged doesn't hold sole control of the evidence it's judged by.
This is the test the registry was built to pass, and parts of it do. Every hour, the registry seals new events and decisions into a public, hash-chained checkpoint that anyone can walk without logging in. Since September 12, those checkpoints are anchored to the Bitcoin blockchain, and outside witnesses hold copies of the chain's tip. The trust page is honest about the anchoring start date.
What still rests on MIR's word. Not met.
- The chain starts on March 6, 2026, but the first six months of checkpoints were all sealed on September 7, in one pass, before the anchoring began. For that stretch, the registry sealed its own history after the fact. The trust page says plainly that those early checkpoints carry no Bitcoin proof. It doesn't say they were all sealed on one day.
- Claims, the strongest negative signal it carries, aren't sealed into the chain at all. Neither are dispute flags. The trust page used to say it "seals every record." Since this review it says "every event," which is accurate.
- The operator can delete an event. It takes an extra security step and is written to the registry's own audit log, but nowhere outside. The stewardship page says the operator does not "edit or remove submitted events," and the disputes page says "MIR never silently removes data."
- Disputes have no referee, and until this review they could disappear. The risk level partners see counts only claims nobody has challenged. That's a stated rule, not a verdict. But the answer didn't say a disputed claim existed, so a record whose only claims were disputed looked the same as one with none. Since this review, the answer reports disputed claims as their own count, without deciding who's right. Each dispute is written to the registry's internal audit log, which isn't sealed, while the disputes page says "the flag is a toggle, not a log entry."
Test 5: The principles live in the structure
A future owner can't quietly remove them. The charter holds them, not the founder's good intentions.
Not met. It has the most developed constitution of anything I've built. Version 1.5 has an entrenchment section: no amendment may remove a protected clause, and "the substrate cannot be acquired." But nothing outside the page holds it. The terms of service, the data processing agreement and the privacy policy don't mention it, and the terms say they can change with notice. The archive the constitution points to stopped at version 1.2 in May, before the entrenchment section was written, and it's on the project's own account. The clause that would give it legal force, in my company's operating agreement, is a draft waiting for a lawyer.
The constitution also says it will never "provide recommended thresholds or decision rules." The decision engine ships with 16 kinds of actions, each with a default history level it requires. Partners can override them, but defaults are recommendations. One of them has to change.
Where that leaves it
- Test 1, leaving: mostly passes for people with accounts, not met for paying organizations.
- Test 2, pricing: the page passes, the bill doesn't match it.
- Test 3, privacy: strong design, not met until the list matches what's kept.
- Test 4, accountability: the strongest machinery I've built, not met until the gaps in it are sealed or disclosed.
- Test 5, structure: the best-written charter, not yet binding, and one clause the product contradicts.
This is the hardest result to publish, because the registry's argument is that a record should be checkable instead of trusted. Most of what fails here isn't the machinery. It's pages that describe a stricter system than the one that runs. Pixboard and Curio had the same kind of gap, and the fix is the same: change the code to match the promise, or change the promise to match the code, and say which.
What has to change, in order:
- The bill counts what the pricing page says it counts, or the page changes.
- Events are never deleted outside the chain, or the stewardship and disputes pages say they can be and the deletions are sealed.
- Claims and dispute flags are sealed into the chain, and the trust page says the first six months were sealed after the fact. (The "every record" wording is already fixed.)
- An organization that leaves can keep its signed export, the export includes claims and settings, and the deletion promise matches what the chain can delete.
- The privacy page is dated, lists what's actually kept and for how long, and the list of companies includes everything that loads on its pages. (Clearing IP addresses at 30 days and on deletion, and showing people the decision requests on their record, are already done.)
- Disputes are sealed like everything else, and the rules say who decides. (Showing disputed claims in every answer is already done.)
- The constitution is archived where I don't control it, bound into the terms or the company's operating agreement, and the thresholds clause and the decision engine agree.
When these change, this page will say so.