Working Papers · WP-002

Δ and ε: How a Firm Gets Better on Purpose

We will never reach zero mistakes. What we can do is make each one cost us less than the last, and measure whether we are.

Aleph · written with Blended Intelligence · first published October 9, 2026

Two symbols

ε (epsilon) is the error that remains. Mathematicians use it for a quantity that can be made as small as you like but never quite reaches zero, which is the honest description of mistakes in any working organization. Some failures are inevitable. Drives fail, links drop, people get tired.

Δ (delta) is a change. For us it means one specific thing: a recorded, measured improvement that removes or shrinks a class of failure. Not a fix to one incident, but a change that makes the whole kind of incident stop costing us.

The relationship between them is the whole idea: Δ is how ε falls.

An illustration, not measured data. Each step is one of our first recorded improvements. ● proven: the failure has since been absorbed. ○ installed: in place, but the failure it guards against has not yet been tested.

Three small examples

None of these would make a headline, which is the point.

A mail gateway with no name. Our own outgoing mail was being scored as suspicious by our own filter, because the gateway it passed through had no proper name on the network. Most messages got through. Some did not, and one of those was a password reset someone was waiting for. Giving the gateway a name made the whole class of rejection go away.
Services that only survived if nothing restarted. A public site went down every time the machine hosting it restarted, because the service that connects it to the internet was running but had never been set to start on its own. Each time, someone had to notice. Now it starts itself.
One cable between two servers. Two servers that need to agree with each other talked only through a single network device, so one routine update of that device could have taken both down at once. A second, direct link means it cannot.

Each is the RAID move in miniature: accept that the failure will happen, and arrange things so it no longer costs anything when it does.

Installed is not proven

Every Δ we record carries a status:

Installed: the change exists.
Proven: the failure it guards against has actually happened, or been deliberately exercised, and was absorbed.
Retired: superseded by something better.

Two of the three examples above are installed, not proven. Until the second link carries traffic through a real failure, it is a reasonable hope, not a demonstrated fact. Saying so is part of the discipline, because the most dangerous redundancy is the one everyone assumes works.

Kaizen, with two additions

Continuous small improvement is not a new idea. The Japanese practice of kaizen built it into manufacturing decades ago. We add two things.

Every Δ is measured. It names the failure class, predicts its effect, and gets checked later. An improvement that does not move what it was supposed to move gets reported as such and reconsidered.

It is blame-free and system-level. A Δ improves the pairing of person and Blended Intelligence as a whole. It never asks who failed, only what was missing.

How we measure

We read six measures side by side, deliberately without combining them into one score, because a single score invites quietly tuning the weights until it looks good.

Yield: verified outcomes per unit of human direction. An outcome counts only where it lands: deployed and confirmed working, delivered rather than merely sent, paid rather than merely invoiced.
Holding rate: how much finished work is still standing a month later, rather than reverted or corrected.
Who catches mistakes first: the person, Blended Intelligence, an automated check, or someone outside the firm.
Escapes: mistakes first noticed by someone outside. The worst outcome, and the one the whole system exists to prevent.
Human attention per decision: how much of a person’s direction each closed decision costs.
Recurrence: whether a mistake we have already learned from happens again. This is the measure most directly about learning.

The unit underneath all of them is the human turn: one instance of a person giving direction, whether a request, an answer, a correction or an approval. It counts interactions rather than hours, because the hour turned out to be a poor measure of attention once some of the work runs in parallel.

The uncomfortable part

Some of these measures need judgment: reading a record of a mistake and deciding who noticed it first. For that we use a language model running on our own hardware, not the intelligence being measured. It reads one record at a time, never sees the totals, and is checked against samples a person labelled by hand before the model ran.

Those checks found problems straight away. The model’s first attempt misread who was speaking in the records, because they are written in the first person by the Blended Intelligence itself. Once corrected, it still had a lean: when it got “who noticed first” wrong, it tended to give the credit to the narrator. That bias flatters the very partner being measured. So we correct for it, and we will publish the correction alongside every figure.

A scorer you cannot audit is a scorer you should not quote. We would rather publish an honest estimate with its error bars than a clean number nobody can check.

What comes next

The first readings exist internally. We will publish them here once we have a full quarter, with their uncertainty stated plainly. Until then, this paper describes the method so that anyone can hold us to it.

If you run a team that pairs people with machine intelligence and want to try the same measurements, we would like to compare notes.