Skill verification on Lemma means checking each skill against recorded evidence instead of a claim. This methodology sets out what counts as evidence, how it is weighed, how reliability and integrity checks work, how partners are matched and how mistakes are corrected. Each part describes what Lemma does today.
Peer confirmation from Sam T.
Confirmed after a 60 min session · 11 Oct
Correction: session voided after review
Logged first; the session stays in the log · 10 Oct
Learned: state management with Leo M.
60 min · 9 Oct · reported as ended early
Taught: hooks refactor with Priya K.
60 min · rated 5 stars · 7 Oct
Skill assessment passed · level 7.4
19 questions · integrity clear · 2 Oct
Recomputed
From all 21 evidence entries · ruleset v3
You can list any skill you want to learn or teach. Listing it adds nothing to its score. Only recorded evidence does.
Every piece of evidence is appended to a log. Scores are recomputed from that log by a versioned ruleset, so the same evidence always gives the same result.
Every score shows the drivers behind it. If a number cannot be traced back to the evidence that produced it, it does not appear.
Skill assessment passed · level 7.4
19 questions · integrity clear · 2 Oct
Taught: hooks refactor with Priya K.
60 min · rated 5.0 · 1.5x · 7 Oct
GitHub repo verified · checkout-form
42 commits by you · 9 Oct
Peer confirmation from Sam T.
After a 60 min session · just now
React SCI
80
Practice 23 · Proof 22 · Reliability 20 · Freshness 15
Five kinds of evidence count on Lemma: live sessions, adaptive assessments, peer confirmations, external proofs from GitHub, Figma or Vercel, and AI tutor practice. Each is recorded when it happens, by the system or by another person. None of it is typed in by you about yourself, and listing a skill adds nothing.
Priya K.
Taught · 60 min · 7 Oct
19 questions · pass 6.5
Sam T.
Confirmed React · 11 Oct
Walked me through a reducer refactor, clean and well tested.
alexchen/checkout-form
42 commits by the owner
GitHub · 1 of 2 connected proofs
useEffect cleanup
Learning and teaching sessions held in Lemma's video rooms. A session counts once both people have joined and it completes. Your partner's rating sets how much it adds, and teaching weighs more than learning.
A session rated one star adds nothing.
Timed assessments whose difficulty adapts to each answer, at two levels: a skill assessment and a longer certification assessment. Harder questions are worth more.
Only your best result counts. Failed attempts carry a small, capped penalty, and retries have a cooldown.
Another member confirms you have the skill, usually after a session together. You cannot confirm your own skills.
Only a few distinct people count toward a skill.
A GitHub repository or commit, a Figma file or a Vercel deployment, linked through your own account so ownership is checked against that platform.
Two per skill count. Remove one and the score is recomputed.
Drills passed in Lemma's AI tutor. They show steady practice between sessions.
A low cap. Practice supports other evidence; it cannot replace it.
Evidence feeds SCI, a score from 0 to 100 for each skill. Four components add points, each with its own cap: practice up to 25, proof up to 40, reliability up to 20 and freshness up to 15. A fifth, integrity, can only subtract, so no single kind of evidence can carry a skill alone.
Alex Chen
Full-stack developer · San Francisco · since Mar 2026
React
21 evidence · Peer confirmation · 11 Oct
TypeScript
15 evidence · AI tutor drill · 22 Sep
AI Application Development
9 evidence · Vercel proof · 19 Sep
Nobody edits a score by hand. When new evidence arrives, the skill is recomputed from its full evidence log. The ruleset that does it is versioned, so a past score can be replayed and checked.
The SCI explainer covers what each component counts and how the caps work.
Read the full SCI explainerA skill score says what you can do. Reliability and integrity checks say whether the record behind it holds up.
Penalties older than 30 days no longer count
Integrity checks
Completed sessions build reliability. Clean sessions add points up to a cap. No-shows and late cancellations in the last 30 days take points away; older ones no longer count. A long clean record earns a bonus that a single penalty removes.
Integrity checks look for score inflation: most sessions with the same partner, or confirmations traded back and forth inside a small circle. Friends who study together a few times do not trip them; systematic repetition does.
Questions have time limits and no going back, and recently seen questions are not served again. Each assessment carries an integrity score that drops with signals such as tab switching, pasting or leaving full screen.
Subscription plans change what you can access, never how you are scored. A free member and a paid member with the same evidence get the same score.
Matching suggests people to learn with or teach, based on the skills you have listed. Blocked people never appear, people you keep passing over appear less, and each suggestion comes with short reasons.
Can teach you React · Highly reliable · Both speak the same language & are in similar timezones
Can teach you React, Next.js · Reliable track record
Whether you can teach each other, or only one way, and whether they teach the skill you are focused on now. This carries the most weight.
How far apart your timezones are, so a session can happen at a reasonable hour for both of you.
Whether you share a spoken language. No overlap lowers the rank but does not hide the person.
Their session history. A record of turning up ranks higher; recent no-shows rank lower.
The ranking learns from sessions that went well for people with similar skills and goals, and gives newer partners a fair chance to be seen.
Automatic completion is right most of the time. When it is not, either person can report the session, a member of the Lemma team reviews its history, and the fix is added to the record instead of rewriting it. Voided sessions are marked, not deleted, and scores are recomputed for both people.
Report filed by Alex C.
“Leo left after 6 minutes; the session was marked completed.” · 9 Oct 19:12
Support case opened
Opened from the report · 9 Oct 19:12
Reviewed with session history
Room events and join records · 10 Oct
Decision: void
Logged before it is applied · 10 Oct
Credits refunded, both scores recomputed
Record kept, marked voided · 10 Oct
Original record, kept
React with Leo M. · 60 min
9 Oct 18:00 · policy v2
A session completes on its own once both people have joined. If one person never joins, it is recorded as a no-show for them.
A no-show that was not one, a session that ended early, inappropriate behaviour. A report opens a support case, and a person reviews it with the session's history.
A void refunds credits, reverses progress and stops the session counting. An upheld report leaves it standing. Either way, the correction is written to the log before it is applied.
Voided records are marked, not deleted, and scores are recomputed for both people. Each session keeps the policy version it was booked under, so later rule changes do not rewrite it.
Join the early-access list to build a skill record of your own, or read how each part of the product applies these rules.
Lemma Skills Passport
Full-stack developer
12 sessions · 6 proofs · updated 11 Oct
6 sessions · 5 proofs · updated 22 Sep
4 sessions · 5 proofs · updated 19 Sep
