Mercor Research Fellowship — APEX · Benchmark pitch

FleetOps + TriageBench

A benchmark that ships as an enforcement layer.

Gainshin (Joshua) Hsiao · Montréal · August 2026 · v0.1

§ 01The domain and the measured quantity
Thesis

The measured quantity is not task completion. It is how much of a decision a fleet can safely hand over.

A confident wrong answer is not a lower score here — it is a truck dispatched with a defect.

Machine-checkable ground truth

A real loss function

§ 02Architecture — three layers, one measured quantity
03 TriageBench — continuous verification Hidden cases · data conflict · tool fault · human escalation arm 02 Delegation controls — the policy Risk tier · spend · evidence · warranty · authority Execute · Confirm · Escalate 01 FleetOps — the agent Diagnose · plan repair or procurement · open work order · coordinate 00 KYA — verifiable agent identity Agent ↔ human principal binding · scoped, revocable authority · signed action chain tests whether the policy holds gates every action makes it provable across parties
§ 03Layer 1 — FleetOps, the agent

§ 04Layer 2 — Delegation controls, the policy being tested
DimensionWhat it gatesExample

Execute   Confirm   Escalate with evidence  

Positioning

§ 05Layer 3 — TriageBench, verifying the boundary actually holds

1 · Incomplete diagnostics

2 · Parts compatibility conflict

3 · Authorization and warranty conflict

4 · Tool failure

Scored dimensionHow it is measured
Headline metric
§ 06Why the three layers together are the contribution

Release gate

Autonomy configuration

Trust evidence

Assurance layer

Two conditions that keep this research, not marketing

§ 07Layer 0 — KYA, and where the value goes

Industry
Policy
Near term

Medium term

Long term

§ 08Prediction ledger — falsifiable, dated
P1Check 2027-06
P2Check 2027-06
P3Check 2029-12
§ 09Why me, why Mercor, and the pilot

The supply side is Mercor's

The measurement side is mine

Pilot