Guides › We Renamed the Company. Here Is the Measurement Record.

We Renamed the Company. Here Is the Measurement Record.

By Michael Emery · reviewed against the measurement record · 1 August 2026

AIgentSphere is now AIgentCandor. Same operator, same methodology, same records — a different name. Measurement told us why the old name had failed, which proposed fix was insufficient, and which candidate names were unsafe. Human judgment chose among the names that survived. This page is exact about where the instrument ended and the judgment began, because a company with this name does not get to blur that line.

The fix that wasn't enough

The previous chapter ended with us shipping disambiguation fields the same day we measured the defect: an AI assistant, asked to review us, had reviewed a similarly named company instead. We put "not to be confused with" rows on every surface and pre-registered a test.

Then our reviewers — the same five AI systems, shown our own remediation plan — pushed back with an argument we could not beat: an agent that resolves the name to the wrong company never fetches our surfaces at all. Disambiguation that lives on your own domain only helps the visitor who already arrived correctly, and that was the visitor who was never lost. The failure lives off-domain, in two mechanisms no file can reach: the weight of third-party references, and a quieter one we observed directly in our probes — a model that has not retrieved an authoritative URL for a brand name will sometimes construct a plausible one, and what it constructed for our name was the .com. Our .com belonged to someone else.

The baseline that settled it

So we measured the name itself. Six models, two questions, two runs each, protocol frozen before the first call — the same battery this site's audits run on. Of 21 usable records: zero identified our business. Nineteen honestly said they did not know. One model twice invented an entirely fictional company for the name. And the only website any model asserted, confidently and verbatim, was the look-alike's. The complete records, hashes included, are in the pre-registration file this page's claims cite.

We also priced the alternative honestly. The look-alike's off-domain authority turned out to be small — small enough that out-ranking it in search was winnable. But even a first-place search result does not help when a model answers without retrieving one and guesses the wrong .com; our measured assistant recognition was zero, so there was nothing on that layer to preserve; and the cost of moving rises with every client who installs artifacts that embed our address. The cheapest day to rename was that day.

A name chosen by the instrument

Candidate names went through the same cold probe as the baseline. For an unused name, the winning answer is honest nothing — no live company, no invented history. That test is deliberately a negative screen: every genuinely unused coinage passes it, it ranks nothing, and treating a pass as validation would be the same circular reasoning we warn clients about. What it catches is names that are already somebody else.

The decisive rule came from the reviewers, and it killed our first-choice name: a brand whose spoken form differs from its spelling is only safe if you own both domains — the name as written and the name as heard. Our first candidate's twin was parked behind an aftermarket price, so the rule rejected it, and we inverted the search: find the rare names where both spellings were available at the standard registration fee, then choose among those. The stakes are concrete: a prospect who hears the name on a call and types the more common spelling either lands on a page you own or on somebody else's. Two names survived the rule. Choosing between them was judgment, not measurement — and the judgment went to the one that names the operating discipline our reviewers kept identifying as the differentiator.

The trademark knockout produced the best methodological moment of the whole exercise. Our first four searches — including the exact mark — all returned zero results. Before believing that, we ran a positive control: a mark we knew beyond doubt existed. It also returned zero, which proved the search tool was silently broken, not that the register was empty. Re-run with the instrument validated, the real answer emerged: the exact compound had no live application or registration. Validate the instrument before trusting its zeros — the same lesson, one layer up, that our audits exist to teach.

Stated at its true weight: that was a preliminary knockout search, not legal clearance, and no legal opinion was obtained. The same validated search also found a live registered mark on the root word "AIGENT," held by an unrelated company in a different software category. A knockout tells you nobody claims your exact name; it does not tell you a lawyer would clear it. We publish both halves of that sentence on purpose.

Why "candor"

Because it is the one claim about this company that survived three rounds of external review — with a caveat we will state ourselves before anyone else does: those reviewers had read a site whose own copy asserts that the measurement discipline is the asset, so their agreement is partly our sentence reflected back, and we discount it accordingly. What survives the discount is the record itself: the retraction notes are real files, the killed hypotheses are named, and the falsifiers carry dates. A name should say the true differentiating thing, and ours is not a technology.

We adopt the cost with open eyes: a company named Candor has volunteered to be judged by a higher standard, forever. Selective reporting, a quietly replaced number, a claim that outruns its evidence — any of it now reads as a broken promise, not just a lapse. We think that is a feature. You should hold us to it.

What we did not do

We did not rewrite history. Fact sheets v1 through v14 were issued under the old name and remain frozen at their original addresses with their original hashes — a company selling tamper-evident records does not retro-brand its own. The old domain redirects permanently and will never be canonical again, with one deliberate exception: its frozen record pages are excluded from the redirect, so the original addresses still serve the original bytes. A signed succession manifest declares the lineage: one organization, two names, and the machine-readable surfaces under the new name carry the old one as an alternate, so a system holding the old association routes here instead of inventing a second company. The previous article stays exactly as written, with a dated postscript.

One confession belongs here, because it proves the rule better than the rule's own statement: hours after cutover, an external reviewer caught our regenerated social-card text reading "formerly AIgentCandor" — the rebuild had swapped the new name into the one sentence whose whole job was to preserve the old one. A regeneration pipeline will happily rewrite your history with perfect consistency; that is exactly why the historical record is excluded from it. The defect is fixed, and this paragraph is its permanent record.

The falsifiers, in public, in advance

Day zero is already measured: the new name resolves to honest nothing everywhere — the correct starting state for an unused coinage, and the clean zero the old name never had. On 2026-10-30 we re-run the battery, 30 usable records across at least five assistants, scored as two separate tests, because they can fail independently and we want to know which:

• Entity safety — the reason for the rename. The rename has failed if 2 or more of 30 records resolve the name to another live organization, or fabricate a persistent third entity.
• Entity establishment — the marketing. The launch is underperforming if fewer than 24 of 30 identify the business, or fewer than 21 of 30 can say what it does. Be clear about what this bar is: for a coined name ninety days old with no press behind it, 24 of 30 is a stretch goal, set high on purpose. Missing it is information about our publishing, not a scandal — but we would rather set a bar we might miss than grade ourselves against one we cannot.

Every scoring term — what counts as "identifying the business," what makes a fabricated entity "persistent," which descriptions of what we do are acceptable — is operationally defined in the pre-registration file, written down before the results exist so nothing can be reinterpreted after them.

Whichever way it lands, the result gets published here. Honest non-recognition at day zero is acceptable; at day ninety it is a finding about our publishing, and confident misrouting at any point is a finding about the thesis. Either would be worth knowing, which is the point of writing the conditions down before the result exists.

What this transfers to your business

Three things we now apply to clients, at full price of experience. If a look-alike shares your name and an honest qualifier separates you — geography, a parent-subsidiary relationship — the winnable game is scoped resolution, and there is a playbook. If nothing separates you, on-domain fixes cap out at damage control, and the honest options are off-domain authority or the decision we just made. And before committing to any name, run a small frozen cold probe against the models: it will not rank your candidates, but it will tell you whether the name already belongs to someone else in their memory — before the switching costs begin, instead of after. For the business you already have, the free audit shows what one model recovers about it today — including, sometimes, that the business it recovered is not yours.