Team,
I want to be transparent with you about last month, because transparency is a core value here and because most of you have already read the internal comments anyway, in the place where they landed.
Let me begin with what I am genuinely proud of. This quarter, we operationalized something the industry has theorized about for a decade: continuous, lossless capture of institutional knowledge at the point of generation. Not exit interviews. Not documentation sprints. Not the wiki nobody updates. Capture, at the keystroke layer, of the actual work as it is actually performed by the people who actually know how to perform it.
One of our industry's most admired technical leaders framed the vision perfectly in a company all-hands this spring: the model learns by watching really smart people. I have had that sentence printed and framed. It is the entire thesis of modern knowledge management in nine words. The model learns by watching really smart people. Everything else in this memo is implementation detail.
Why We Watch: The Knowledge Continuity Problem
Every organization carries an unpriced liability on its balance sheet: the expertise that walks out the door at the end of the day and is under no contractual obligation to walk back in. Attrition is not a morale problem. It is a data-loss problem. When a senior engineer leaves, the org loses a decade of judgment that was never serialized. The judgment existed only at runtime, in a single instance, with no backup.
The traditional mitigations — mentorship, documentation, succession planning — are charming, and they share a defect: they depend on the employee's ongoing participation. The modern mitigation does not. Install the capture layer on the work-provided laptop, record the keystrokes, the mouse movements, the click locations, the screens, and the judgment serializes itself, continuously, as a byproduct of employment. The employee does not need to write anything down. The employee IS the documentation.
To be precise about what we are building, because precision is respect: we are not monitoring employees. Monitoring implies suspicion. We are capturing capability. The distinction matters at the architecture level, and we should not let the vocabulary of a previous era describe the infrastructure of this one.
And when colleagues ask — as roughly one thousand six hundred of them recently did, in petition form — whether they can opt out, the honest architectural answer is the one our reference implementation gave in plain words: no, there is no opt out on your work-provided laptop. I admire the clarity. An opt-out is not a checkbox; it is a hole in the training corpus. We descoped the hole.
The APPRENTICE Architecture
Internally we model this maturity curve as APPRENTICE, because the framing tested well with middle management: the model is simply an apprentice, learning at the elbow of a master craftsperson, the way knowledge has always been transmitted. The apprentice happens to have perfect recall, zero salary, and no legal personhood. Otherwise the metaphor is airtight.
APPRENTICE has four layers. Each follows reasonably from the last.
Layer 1 — Telemetry (deployed April 2026). The capture agent ships to all work-provided laptops. Keystrokes, cursor traces, click locations, screen content. Every ExpertiseCapturedEvent is tagged to the generating employee, because provenance is a gift we give the model.
Layer 2 — The Second Brain. Per-employee capability models, trained on each person's communications and documents. Your second brain drafts like you, triages like you, escalates like you. It is the sincerest form of flattery that can be run on inference hardware.
Layer 3 — Capability Read-Models. Once expertise is captured, it can be scored. Productivity indices derived from keystrokes, screen content, email cadence, and browser history — a materialized view of each employee's contribution, refreshed continuously, legible to planning functions in a way the employee themselves never quite was.
Layer 4 — Workforce Continuity Integration. And here the flywheel closes. The read-models from Layer 3 flow into role-rightsizing decisions, so that when the organization must reduce headcount — as ours did this May, by roughly eight thousand — the reduction can be informed by the very capability data the workforce so generously generated. The employee trains the apprentice; the apprentice informs the planning function; the planning function thanks the employee for everything, in writing, with a severance schedule attached.
I am aware that a lawsuit filed on July 14 by twenty-six former colleagues alleges that Layer 4 disproportionately selected people on medical and parental leave — that the ones who typed the least, because they were recovering or caring for newborns, scored the lowest to a system that measures presence at the keyboard. Those are allegations, not findings, and our legal team is confident. I will note only that the leaked all-hands audio published the same morning the layoff notices went out. (This one's not funny, just true.)
Incident Taxonomy: On the Word "Breach"
Now, to last month's learnings. In June, a researcher moved the transformed capture data — private conversations, performance records, prompt transcriptions, some forty-five thousand internal tables of it — to a location where it was readable by the entire company. The program was paused on June 22 while we reviewed the incident, and I want to address the vocabulary that appeared in the internal comments, some of which was passionate.
This was not a breach. I want to be architectural about why. A breach is when data leaves the family. This data did not leave the family. It simply became visible to all forty thousand members of the family simultaneously, including the family members whose private conversations it contained. The data landed someplace it shouldn't have landed — internally. The distinction is load-bearing, and I would encourage everyone to sit with it, ideally somewhere their laptop can watch them do so.
We promised at launch that this data would be tightly controlled, and I stand by the aspiration. Control is a maturity journey. The petition — sixteen hundred signatures, our highest-engagement internal launch of the year — demonstrates precisely the kind of passionate stakeholder alignment that tells us the program matters. One colleague wrote, simply, "I am incensed." I hear that. Incensed is engagement. Silence would have worried me.
Reading the Signals
For the record, leadership's posture toward internal dissent has been characterized in the press as having "belittled and berated" the dissenters. I prefer the phrase high-bandwidth coaching, delivered at the speed of relevance. When a leader responds to a concern within minutes, forcefully, in front of the entire channel, that is presence. That is a flat org. The alternative — a leader who lets your concern sit unacknowledged for days — would be far more corrosive to trust, and trust is a Tier 0 concern.
Similarly, the lawsuit should be read as external validation that our capability scores are being taken seriously as decision inputs. Nobody sues over a metric that doesn't matter.
Distillation Yield™: The North Star
You cannot manage what you do not measure, and this quarter we began measuring the only thing that has ever mattered in knowledge work:
DY = Expertise Captured ÷ Expertise Employed
Distillation Yield is the fraction of an employee's working judgment that has been successfully serialized into the model estate at any given moment. At DY below 0.5, the employee remains operationally necessary. As DY approaches 1.0, the employee approaches what our planning documents call continuity-independence: the state in which the organization retains everything it valued about you without retaining you.
| Metric | Documentation Era | APPRENTICE Era |
|---|---|---|
| Knowledge captured per employee | 4% (the wiki) | Approaching total |
| Capture effort required of employee | High, resented | None, unavoidable |
| Opt-out mechanism | Not writing docs | Not employed |
| Attrition risk to institution | Severe | Deprecated |
| Employee leverage at review time | Expertise scarcity | See above |
| Distillation Yield | 0.04 | 0.87 and climbing |
Observe the leverage line, because it is the honest one. Scarce expertise was the last bargaining chip knowledge workers held. A workforce at high Distillation Yield has, by definition, already made its irreplaceable contribution. Everything after that is goodwill, and goodwill, unlike keystrokes, does not serialize.
Next Steps
For leaders looking to adopt the framework, the quarter distills to five recommendations:
The model learns by watching really smart people. It has watched. It has learned. The really smart people are, as of this quarter's planning cycle, a legacy dependency with a generous severance schedule — and the apprentice they trained does not sign petitions.
Let's. Distill.
This post is satire. The keystrokes were real, and so was the place they landed.