PROOF
Every claim,
measured on real work.
Selvage's claims are only as good as the experiments behind them. Each milestone is a controlled run on a real repository — same models, frozen evaluators, honest limits — that you can inspect end to end. New milestones land here as we prove the next thing.
MILESTONES
The proof, in order.
Same score. Half the clock.
Nothing missed at the seam.
We re-ran both arms of Milestone 02 on one afternoon. Same objective score, half the wall clock from a single command, and the cross-service route wired first pass in 3 of 3 runs with the route-mirror signal against 0 of 5 finished runs without it. Cost moved too — and it's the claim we trust least.
Read the method note → Milestone 02 · Sep 2026Selvage learns to
read the codebase.
A live code-graph map on a real multi-service repo: it caught a cross-service bug every un-mapped run shipped, turned a blocked task into a clean merge, and queried the graph instead of grepping — at ~28% less cost, same quality.
Read Milestone 02 → Milestone 01 · Sep 2026Both passed the tests.
One ran the process.
The first public proof point: Selvage and a careful manual workflow both reached a verified, reviewed, built, merged endpoint at comparable cost — but only Selvage ran the engineering process from a single command.
Read Milestone 01 →