Latest work and presentations

Working paper · mathematics · not yet peer reviewed · September 2026

Operational Capacity and Bridge Amplitudes on Refinement Complexes

The mathematics under observation capacity. In a finite model of observation, the paper separates three quantities that are easy to conflate: how many independent questions a protocol can ask within a budget, how many outcomes it can actually tell apart, and how the many histories that lead to one outcome add up.

• A budget bounds the questions

  When every informative query has a positive minimum cost, a budget caps the rank of what one protocol can ask, and with it the number of outcomes it can distinguish.

• Rank is not resolution

  Reaching that ceiling also needs the interface to realize every signature. Independent questions alone do not guarantee distinct answers.

• The choice of protocol is information

  Switching among protocols adds the protocol itself as an index, so a family of protocols can separate more than its best single member.

• Geometry costs only when unavoidable

  Curvature in the refinement geometry lowers capacity only when it forces a positive cost at the ranks being tested. Counterexamples show curvature alone does not.

• Amplitudes need their own assumptions

  Rank and outcome counts do not control the summed amplitude of histories; the paper gives finite-history bounds and a growth criterion under stated conditions.

• Limits stated

  Earlier physical and algorithmic applications are withdrawn as deductions and kept as open questions. Exact-arithmetic checks accompany the proofs.

Conceptual paper · draft · August 2026

In Situ Analytics: Instrumentation as the Unit of Enterprise Intelligence

That intelligence belongs inside the work rather than in a report is widely asserted and rarely explained. This paper supplies the account, in Beer's managerial sense of cybernetics rather than a technological one.

• The asymmetry

  AI can generate plausible records from what an organization already holds. It cannot recover ground truth about an event the organization never instrumented. That makes observation capacity, not any dataset it produces, the durable asset.

• What makes an instrument

  It captures a decision-relevant event as part of the work, conditions action while the decision is still open, and returns what was learned to shape what follows. A system with two of the three is a report, a monitor or a dashboard.

• A delegation system

  An instrument centralizes accountability in its architecture while moving observation and decision authority to the point of work. Adoption is the spread of that authority, not the count of logins.

• Build or buy

  The choice turns on marginal cost per unit of work, not capability. Inference that capitalizes into structure amortizes; inference that stays in the loop recurs.

• The boundary

  Observation of work must never become observation of workers. The record is keyed to the transaction, and the line is enforced by access control rather than promise.

• One measure

  Adaptation latency: the time between a real divergence in operations and a governed change in the organization's response.

Draft, August 2026. Comments welcome.

Read the draft (PDF) ↗

Open code · lhpc_kb · September 2026

California Health Plan Policy Knowledge Base

Policies from all 17 Local Health Plans of California plans, with DHCS All Plan Letters as the cross-plan spine. Page-anchored chunks with vector and BM25 search in one DuckDB store. Every chunk carries its plan and line of business, because nine plans publish something called a Provider Manual and an answer about one must not be drawn from another.

Read the code ↗

August 2026

Realtime Operational Instrumentation in Healthcare: AI Analytics in Action

Ai4 2026 · The Venetian, Las Vegas

A biologist in the field checks one box in a notebook: blue butterfly or red. The notebook is the instrument. In the lab the specimen can show its wingspan but never its flight. Enterprise analytics works the same way, and every dashboard is a dead butterfly.

• The analysis is not somewhere else. In an instrument it is the screen where the work happens, already in priority order when someone signs in.
• Governed drafting: a clinical note that carries its own evidence, including the rule version it was written under, the policy passages retrieved, the one it relied on, and the person who signed it.
• Shown on a test bed of synthetic records: the instrument caught its own drafting model asserting an unsupported eligibility basis and leaving out citations, and the fix shipped the same day.
• When code is nearly free, the scarce resource is no longer engineering. It is proximity to the work.

Without instrumentation, failure is not detected. It is survived.

February 2019

Nuts and Bolts of Being a Data Scientist

Biostatistics seminar series · School of Public Health, West Virginia University

A seminar for the biostatistics department on what the working life of a data scientist actually involves.