Research / Publications
Research on how AI systems allocate work and create value.
We publish practical work on model routing, data economics, and the systems that connect technical performance to commercial outcomes.
Research 01 / Research paper / August 2026
Pricing Capability
Could Spirit Aviation's deidentified data-and-software package have been underpriced at $10 million? Under the paper's stated assumptions, the selected price implies a 0.236% break-even hurdle. Separate conditional buyer-specific scenarios produce values of approximately $50 million to $200 million. Spirit's Capability Alpha remains unmeasured.
01DatasetA legally usable proprietary corpus
02Capability AlphaCausal gain measured in a controlled evaluation
03Economic translationExposure x monetization x useful life
04Buyer valueUse value, PIU, and strategic extensions
$10MSelected bid
0.236%PIU at 5% scope
$50M-$200MConditional use value
The use-value range assumes a 1% improvement in airline economics and 0.5%-2.0% Google capture. It is a conditional scenario, carries no fair-value claim, and leaves Spirit's Capability Alpha unmeasured.
Research 02 / Open-source paper / April 2026
TRACER: Learning When to Defer
Can an AI system learn which classification calls still need an LLM? TRACER trains a lightweight surrogate from the model's own production traces. An acceptor serves predictable inputs locally and defers uncertain cases to the original teacher.
01Production traceInput plus teacher label
02Local surrogateFast fixed-label prediction
03Acceptor gateEstimate teacher agreement
ConfidentAnswer locally
UncertainAsk the LLM
83.2%Banking77 local
.959Teacher agreement
0%MNLI refused
Paper results at target alpha .95. CLINC150 reached full local coverage but missed the .95 target on the unseen test set. MNLI was refused in every tested configuration.