Evidence

Everything below was counted from the system we run our own business on, on the day this page was written. Nothing here is a projection or a rounded-up marketing number. Each one says where it came from, so you can ask to see it.

What’s actually in it

What	How many	Where it comes from
Pieces of past work you can search	1,406	the search index
Full conversations kept word for word	532	saved sessions
Write-ups filed automatically	1,466	customer history files
Customers they’re filed under	22	customer folders
Things the AI asked permission for	281	the approvals list
— approved / turned down / still waiting	143 / 3 / 114	the approvals list, by status
Separate machines those requests came from	14	recorded on each request
Saved versions since April 2026	13,712	version history
Documents it keeps track of	16,636	file count
Actions written to the log	1,955	the activity log
Built-in limits published openly	59	published limits list
Separate installations running	4	live version check

These were true at publication. They are not updating themselves on this page — which is the point of the next section.

Why there are no version numbers anywhere on this site

Because the moment you write a number down, it starts going out of date.

VaultHelm enforces this on itself, and it’s the clearest example of how the whole thing is built. Every time it starts up, it looks at its own version number. If there’s any chance that number is stale, it doesn’t hand it over — it flags it as unconfirmed and tells its own AI to go and check the real one instead of repeating what it was told.

Never quote a version from memory or from a document. Go and look, or don’t say it.

That rule came from getting caught. A reference document here claimed one number while the live system had a different one — in the same paragraph that warned against writing the number down in the first place. The fix wasn’t to correct it. Correcting it just restarts the clock. The fix was to delete the number and have the document build itself from the real thing.

The same thinking is behind the published limits list. Every system has built-in limits — how much it will read at once, how long it will wait, how many results it returns. Most companies don’t tell you what theirs are, so you find out when something quietly comes back short. We publish all 59 of ours, pulled straight from the working code, by a tool that goes red and refuses to finish if any of them moves. Not one of those numbers is typed by hand.

What you get back when you finish

You type one word. This comes back:

Saved.     The full conversation, kept word for word
Filed.     Sorted into the right customer's history
Checked.   99.7% of it accounted for — nothing quietly dropped
Closed.    The job you were working on, marked done

The third line is the one that matters. A summary that loses half your work but reports success is worse than no summary at all, because you won’t find out until you need it. So VaultHelm measures how much of what you gave it actually made it through. If too much went missing, it keeps everything word for word instead. It is not permitted to report success without doing that check.

And if your connection drops mid-save, just say the word again. The record is matched on its content, so a repeat only fills in what didn’t make it — it can’t create a duplicate and it can’t overwrite what’s already there. There’s a built-in test that proves this: run it three times, get three confirmations and exactly one saved copy.

Finding something again is three steps, not a guess

When you ask what was done for a customer, you don’t get an AI’s best recollection. You get:

1. The entry — who, when, what it was about.
2. The write-up — what was actually done, in that customer’s history.
3. The original — the raw conversation, exactly as it happened.

You can always reach the third one. The middle one can be corrected if it got something wrong. The one underneath is never rewritten — which is the difference between a record you can rely on and a summary you have to trust.

What we deliberately don’t measure

You won’t find uptime percentages, customer counts, speed benchmarks or named references on this page.

Uptime and speed figures would need measurement we haven’t built to a standard worth publishing, and a number we can’t stand behind is worse than no number. Customer detail isn’t ours to put on a website.

If something matters to your decision and it isn’t here, ask. If it doesn’t exist yet, that’s the answer you’ll get.