Critical Density

The problem you only see after it is too late

Your AI estate grew the way every estate grows: one sensible deployment at a time. An assistant here, a fine-tuned model there, a vendor feature switched on by default. Every individual decision was reasonable. Nobody chose to cross a line, because there is no line you can see. The dependency builds quietly, the curve stays smooth, and the day it tips, no single event triggers an alarm.

Then something goes wrong, and you discover the part that actually hurts: you cannot say which layer did it. The base model, the fine-tune, the corpus, the orchestration, the prompt, each owned by a different party, each able to change behaviour without telling the others. When you go to reconstruct what happened, the record is scattered and will not join.

# incident review: AI system produced a wrong answer that was acted on
$ reconstruct --request req-8841 --window 90d
  base-model log .... vendor cloud, UTC, key: trace_id
  orchestration log . our stack, UTC+2, key: session
  retrieval log ..... data team, no clock, key: none
RESULT: cannot reconstruct — no joinable timeline
# evidence everywhere. an account of what happened: nowhere.

That capacity, the ability to say who did it, is not recovered after the incident. It is a property you build in before, or it is absent forever. This book is how you build it in.

An instrument, used and then set down

The book reasons about your AI estate through the Kessler Syndrome, the orbital-debris cascade where each collision makes the next more likely until a band of orbit becomes unusable. It is a clean, emotionally neutral model of how dependency accumulates, how thresholds get crossed without an alarm, and how accountability dissolves once the harm is done. It lets you think clearly about your own systems without first picking a side in the tired argument over whether AI is salvation or catastrophe.

It is an instrument, not a thesis. The orbit story leads the first third and then recedes on purpose, until the governance substance carries the page on its own. By Part II you are reading audit practice, not space. No predictions about how AI ends. No hype, no doom. An operating manual.

What is inside

The book. Ten chapters, three beats each. First the mechanism, stated as a general law. Then the translation, what that behaviour looks like in your AI portfolio, named as an observable you can go and check. Then the move, one concrete artifact with an owner, a meeting, and a first action. Every chapter ends with something to do on Monday. Ships as a print-ready PDF (A4), a reflowable EPUB, and a browser HTML copy.

The toolkit. Five board-ready artifacts that assemble into one system, each as a ready-to-print PDF, a browser HTML, and editable Markdown:

• The Density Register The join key. One scored row per AI touchpoint. Everything else attaches to its rows, and building it surfaces the touchpoints nobody knew were live.
• The Attribution-Readiness Checklist + Traceability Architecture Spec 25 scored items opening with the Licence-Plate Test, plus the seven requirements that make the hard items cheap. It tells you, before an incident, whether a system can be attributed at all.
• The RACI-for-AI One named person in the Accountable cell per system, with lifecycle columns that close the six accountability seams, an Article 25 role-shift check, and an orphan check.
• The Audit Question Bank Around sixty evidence-scored questions across eight sections. The examiner's lens, run formally by internal audit or as a self-assessment by the risk officer.
• The Governance-Cadence Control Spec The meta-control: the review calendar, the assumptions register, and the refresh matrix that schedules the other four, so the toolkit stays a living system instead of a folder of decaying snapshots.

The system at a glance

The artifacts arrive one chapter at a time, but they are one system. The Density Register is the spine; the others attach to it; the cadence keeps them all current.

        THE DENSITY REGISTER  one row per AI touchpoint (the spine)
                 |             |             |
            scores each    names each   examines each
                 v             v             v
   ATTRIBUTION      RACI-for-AI     AUDIT QUESTION BANK
   can we trace?    who owns it?    can we evidence it?
                 \             |             /
                  THE GOVERNANCE-CADENCE CONTROL
                  schedules the refresh of all the above

The Licence-Plate Test, in one screen

The free front door is the sharpest instrument from the book: score one system, read the band, find the gap before an incident does.

$ licence-plate-test --system "bid-evaluation-assistant" --criticality 4
  item 1  base model + version pinned ...... 0  vendor serves "latest"
  item 4  vendor change vs our change ....... 0  not joinable
  item 9  clocks + ids join across parties .. 0  three clocks disagree
  ... 12 items scored, 0-2 each ............. 9 / 24
BAND: RED  (red <12 / amber 12-18 / green 19+)
GATE: criticality-4 on a RED → does not deploy

Get the full 25-item checklist and the other four artifacts — $59 →

Who this is for

• CIOs, CTOs, and platform leads who own the AI estate and answer for it
• Heads of internal audit, risk officers, and compliance leads who have to attest
• AI program owners standing up governance that survives a real portfolio
• Anyone preparing for the EU AI Act who needs the controls, not another explainer

The examples lean toward EPC and energy, the seat the author operates from, but nothing is sector-locked. A bank examiner, a hospital's risk officer, and an insurer's model-risk lead all find their own register rows in these pages.

What makes this different

Pricing and updates

$59 for the book in three formats and the full five-artifact toolkit. Free updates: when a load-bearing fact shifts (the EU AI Act dates are still moving), the updated files re-deliver to every buyer automatically. Team use included: share with your platform, audit, and risk teams, no per-seat fees.

Get Critical Density — $59

Frequently asked questions