Assess · architecture
How judging works behind the scenes
Two AI crews and a calculator, where plain arithmetic owns every published number.
The Evidence team, five specialists in a relay
Writing a dossier is split into five small, single-purpose jobs, each sized so an affordable model does it well, instead of one giant “analyse this project” prompt only the most expensive models could attempt.
Surveyor
Quick model“What is this codebase made of?” — languages, frameworks and logical groups, inferred from file names and dependencies alone. It never reads the code.
no tools
Three passes, then the maths takes over
Judging one project for one sponsor takes exactly three AI passes. The Panel and the Fit Specialist run in parallel, the Writer waits for their scores, and the Calculator is code.
Evidence dossier
Everything the judge is allowed to rely on. The judge never re-runs the investigation — it scores from this and only double-checks details.
How the technical score is calculated
The four criterion scores go through a fixed weight table that sums to 100, and then a guardrail keyed to what the code actually proved, not to how the project presented itself.
The credibility cap
A polished pitch cannot buy an engineering score. Only verified code can.
Technical roll-up
The four criterion scores through the fixed weight table: one 0-100 engineering number, before any guardrail is applied.
Use the engine
Same judge, two ways in
Every score you have just read about is reachable either as a hosted event or as an API call, with the same citations behind it.
Hosted
Run it on your event
Bring the rubric. We read every repo, cite every score, and give your judges a board instead of a spreadsheet.
API
Call it from your stack
Point it at a repository, bring your own weights, and render the verdict in your own product.