Reasoning you can check
An applied AI research lab. Lemma, our reasoning engine, shows every step, so people can check AI instead of trusting it.
Sum of the first 100 odd numbers?
- Odd numbers 1, 3, …, 199
- Small cases are squares
- Proved by induction
- 100² = 10,000
Lemma engine
Conjecture
- Read contract 41 of 214
- Renewal: 1 March, 60 days' notice
- Calendar entry drafted
- Clause 14 may override 9.3
- Sent to Legal with reasoning
- Read contract 42 of 214
- No renewal clause found
- Read contract 41 of 214
- Renewal: 1 March, 60 days' notice
- Calendar entry drafted
- Clause 14 may override 9.3
- Sent to Legal with reasoning
- Read contract 42 of 214
- No renewal clause found
Agents
AI strategy
- Misuse uplift
- Reasoning faithfulness
- Prompt injection
Evaluated before release
Five steps to an answer you can check
- Is n² + n even for every integer n?n² + n = n(n + 1)n, n + 1 consecutiveOne of them is even
Decompose
Breaks a question into claims small enough to check one at a time.
- Factorisation holds
- Consecutive integers alternate parity
- Product with an even factor is even
Check
Verifies each claim with code, a proof assistant, a search or a second model.
0.97
Score
Gives every step a calibrated confidence, so people know where to look first.
- Small cases look like squaresReviewer
Review
Sends the steps it is least sure of to a person before anything counts.
Trace saved · trace_4f2a
Record
Keeps the whole trace: every claim, check and decision, ready to audit.
checked
Pick an example or ask your own question. Flagged steps open to show why.
What is the sum of the first 100 odd numbers?
One engine.
Every step on record.
Every Lemma answer comes back as a trace: each claim, the check behind it, its confidence, and what went to a person. There is no public API yet; these are sample traces.
{
"id": "odd",
"question": "What is the sum of the first 100 odd numbers?",
"steps": [
{
"kind": "claim",
"text": "The first 100 odd numbers are 1, 3, 5, …, 199.",
"confidence": 0.99,
"status": "verified"
},
{
"kind": "claim",
"text": "Small cases look like perfect squares: 1, 4, 9, 16.",
"confidence": 0.62,
"status": "flagged",
"flag": "Pattern, not proof",
"why": "A few matching cases suggest a pattern but do not prove it, so Lemma proves the general case by induction before this step counts."
},
{
"kind": "check",
"text": "Proving 1 + 3 + … + (2n − 1) = n² by induction",
"result": "Proof checked"
},
{
"kind": "claim",
"text": "Base case holds. Adding 2n + 1 to n² gives (n + 1)², so it holds for every n.",
"confidence": 0.99,
"status": "verified"
}
],
"answer": {
"text": "100² = 10,000",
"confidence": 0.99
},
"source": "recorded"
}start
Work with Conjecture
- Hypotheses with their evidence attached
- The cheapest test that could prove each one wrong
- Negative results kept on record
- Domain review before every project
Strategy and agents on Lemma
- Find where AI creates value, and what to do first
- Agents built and run inside your workflows
- Every action logged with its reasoning
- Your cloud or your premises