measured, not marketed

everything here is reproducible

We publish the grading sheet — misses included.

You're being asked to trust this thing with your team's memory. So every number below links to its grade sheet, harness, or ADR — including the ones that went against us.

96.7%faithful

Zero fabrications on strangers’ repos.

58 of 60 blind-graded claims held against source on three repos we'd never seen — mozilla/fxa, Backstage, Athens. The two misses are in the sheet too.

0.383< floor

The honest abstain.

A dormant repo scored 0.383 against our 0.38 confidence floor — and the product declined to bluff. It surfaced the gap ledger instead of dressing up an answer.

abstain floor · calibrated, shipped
80%bar — missed

The feature we cancelled.

We pre-registered a precision bar before building auto-promotion of inferred links. The matcher missed it. The feature that would have written machine guesses into your record died as written.

ADR-0066 · docs/evaluation/correval-2026-06-11.csv

“You're being asked to trust this thing with your team's memory. So we show our misses.”

a typical answer arrives as ~220 tokens of cited claims · methodology page ships at GA — until then, no ratios.

open source

MIT, and you can check our homework.

MIT

lema-mcp is open source — the graph is yours.

Runs local, no account, no lock-in. Works with any MCP client.

house policy: shipping today — you can reproduce it without talking to anyone early access — proven, opening upcoming — dashed means dashedif it isn't shipped, it's dashed.

don't lose the why

Don't lose the why. Ask the repo.

Run a real “why” query on React, Kubernetes, Rust, Vue, or Go right now — no account, no import. Or run it local in ten minutes. Both doors lead to the same graph.

Try it now — no login
● try it · no account · React, Kubernetes, Rust, Vue, Go◐ connect your repo — early-access, opening as hardening completes