I build public-interest data infrastructure and work on applied econometrics and
forecasting. Most of what I do starts from the same problem: a question people argue
about using numbers nobody can check.
So the data, the extraction code, and the verification steps are published alongside
every result — traced back to the source documents they came from.
Research
Competency Restoration Observatory
Live · Creator and maintainer
A free, public dashboard and dataset tracking how five U.S. states — Washington, Oregon,
Colorado, Texas and California — meet their legal deadlines for psychiatric competency
evaluation and restoration. Every figure is rebuilt from public court-monitor reports and
budget documents, with each number traceable to the source page it came from. The full
extraction pipeline and data are openly licensed and versioned.
The companion paper below analyses Washington's own trajectory over eight years, from
years of missed deadlines to sustained compliance.
Measurement validity of AI occupational-exposure scores
Under review · ACM FAccT 2027
Widely used indices claim to measure how exposed an occupation is to AI. I tested whether
they measure AI specifically, or general cognitive content. Two of the three most-cited
measures largely capture the latter; a third does not behave the same way. The distinction
matters, because these scores are increasingly used as inputs to labour-market research.
Model-selection instability in real-time forecasting
Under review (Statistics and Public Policy) · Companion note · Open-source tool
Leaderboards crown a single “best” model, but often several are statistically tied
— and which one you pick then propagates into decisions. This work introduces the
decision breakdown point: the fewest evaluation periods whose removal flips the
winner. It is applied to real-time U.S. official statistics, projected through a statutory
rule, and separately to a public machine-learning leaderboard. The diagnostic ships as an
open-source package.
Work in progress on predicting financial distress and closure risk among U.S. rural
hospitals from public CMS filings. Not yet published.
Publications
When the Best Forecasting Model Is a Tie: Pricing Model Choice Against Data Revision
in the Extended Benefits Trigger.
Introduces the decision breakdown point (k*) — the fewest evaluation periods whose
removal flips the pooled winner — with a block-bootstrap null. On real-time U.S. official
statistics, 11 of 12 series have multiple models statistically tied for best at
α = 0.10. Projected counterfactually through a statutory rule, the Extended
Benefits trigger, the call flips in 29 single months across 24 of 50 jurisdictions, 93 percent
of them knife-edge — though data revision moves a comparable call roughly 3 to 4 times more
often like-for-like. A control series shows selection can also be unambiguous.
SSRNUnder review · Statistics and Public Policy2026
One Task Flips the Top of a 4,497-Model Leaderboard: A Companion Note on the Decision
Breakdown Point.
The same diagnostic applied unchanged to the Open LLM Leaderboard v2 — 4,497 models across six
benchmark tasks. On the top-30 pool, k* = 1 of 6: deleting a single task flips the
top-ranked model to its binding rival. The board is nonetheless comparatively identified
against a 50-permutation null — a shakier crown than the paper's own series, on a board that
is not chaotic.
SSRNPreprint2026
Competency Restoration Timeliness in Washington Under Trueblood,
2018–2026.
A federal court found Washington's competency-restoration system unconstitutional in 2015,
for making defendants wait months in jail for a psychiatric bed. It did not turn around
until years of escalating enforcement had culminated in a contempt finding and a
$100 million fine. Using an open monthly panel of court-monitor data
(1,033 rows, January 2018 to May 2026), I analyse inpatient restoration timeliness as an
interrupted time series against the 7-day admission standard. The system was non-compliant
for six years — annual median waits of 28.0 to 63.5 days, compliance rarely above
30 percent, bottoming at 6 percent in March 2023 when the median wait peaked at 97 days
— then improved sharply, to a median of 5.0 days and 90.4 percent average compliance,
holding across 26 months. The gain concentrated in bed-dependent services while a
bed-independent comparator barely moved, and demand rose rather than fell. The improvement
coincided with sustained enforcement; a single-jurisdiction design supports association,
not causation.
SocArXivSSRNUnder review · Psychiatric Services
Do AI Occupational-Exposure Scores Measure AI?
AIOE and Eloundou (2024) largely capture cognitive content; Webb (2020) does not.
SocArXivUnder review · ACM FAccT 2027
Elsewhere
Happy to hear from anyone working on related questions — particularly people inside the
systems this data describes.