TRACER RESEARCH LAB


Intelligence,
ensembled.

Tracer builds ensembles of specialist open models that work as one. Echo is our first foundational model built this way: one endpoint, the right amount of intelligence for every task, and it can be customized to your workload.

Live. OpenAI-compatible. YBacked by Y Combinator TRACER OSS 1K+

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.

01
DatasetA legally usable proprietary corpus
02
Capability AlphaCausal gain measured in a controlled evaluation
03
Economic translationExposure x monetization x useful life
04
Buyer 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.