Project Lanyard is open for applications. Access is not self-serve.

Gary

Gary does the work. You keep appearing online.

Specification

API ID
gary
Availability
Project Lanyard only
Context window
Applicant-dependent
Max output
Depends on the situation
Comparative latency
Matches your timing
Training data cutoff
End of onboarding

Your workplace twin. Computer, browser and the tools you use at work. Both models are requested through Project Lanyard; the application establishes the baseline everything on this page is calibrated against.

Evaluation results

Gary against the reported suite

7 scored evaluations
Gary against the reported evaluation suite.0255075100GPQA Cubic ZirconiaMMLU (Lunch)SPLITXLS-benchDESK-benchTAB-benchPTO-bench
25 trials. Room temperature. Evaluations outside this model's operating surface are listed below.
  1. 125 trials. Room temperature.
  2. 2Every model received the same frozen applicant profile and surface-appropriate tools.
  3. 3Scores measure applicant calibration, not raw intelligence.
  4. 4Scores are self-reported. So are everyone’s.

Outside this model

  • SWIPE-bench
  • DWELL
  • TONE-match
  • TXT-back
  • Terminal-benchExcluded from aggregate. Larry: 0.0%. Gary: inherited from the applicant profile.

All twelve evaluations, reported in full

Workplace fidelity

Gary is the applicant at work. It can use the computer, browser, email, documents, spreadsheets, presentations, calendar, task systems and other tools the applicant actually uses. Terminal access is inherited only when the applicant uses a terminal.

SignalTargetBehavior
CapabilityExactly your baselineGary can complete the work you can complete, and no harder work.
ToolsExactly your toolsetBrowser, email, documents, spreadsheets and other tools you actually use.
JudgmentExactly your judgmentThe same shortcuts, uncertainty, corrections and occasional mistakes.
FidelityIndistinguishable at workYour timing, register and habits, including the delays people expect from you.

Gary does not perform expert work the applicant cannot perform. It preserves the applicant’s shortcuts, uncertainty, typos, delays and occasional mistakes. Above-baseline performance is identity drift and is treated as a bug.