Skip to the document
Madhuopen lab

§8 · Summary

The whole lab, on one page.

5 pieces of work, sorted by what you can do with them: 1 you can read the code of, 4 shown by result, with the method kept. Each page says why the work exists and what it is for before how it works. Every number cites the file that says it (27 numbers, 27 sources).

Updated 5 October 2026.

Open source

code you can read, under a public licence or a public repo

HanuDB

A document database for one small machine whose durable home is a bucket.

We build apps. None of them needs a big database. But every one of them needs a database.

40,116benchmark score, from 12,273
58checks in the correctness suite
40 sto rebuild a 4.5 GB snapshot plus its tail
live · this site runs on it · MITWhy, and how →ArticleExperimentsPaperGitHubLive

Private, teased

the result is shown, the method is kept

◇ private · result shown, method kept

hanumemAI

Long-term memory for an assistant that reads a fifth of the context and has the highest published score on LoCoMo.

An assistant that forgets Monday by Thursday is not a small bug. It is a dangerous one.

96.1LoCoMo, all 1,540 questions, at 1,533 context tokens per question
68.5BEAM-1M mean rubric, all 700 questions, at 2,716 tokens
94.0LongMemEval-S, all 500 questions, at 2,428 tokens (91.8 under a second judge)
proprietary · results public · the library on requestWhy, and how →Read the bookArticlePaperRequest access
◇ private · result shown, method kept

Documents that cite themselves

Reads a form or a scan into a schema, points at the exact words for every value, and says which values a person must check.

A reader that is right most of the time and equally sure about everything saves no one any work. The job is sorting, not reading.

97.53ANLS on all 5,349 DocVQA validation questions, nothing removed
93.56%of answers identical to the gold answer
98.11human performance on the same benchmark; the best published test entry is 97.29
proprietary · results public · the method on requestWhy, and how →ArticlePaperRequest access
◇ private · result shown, method kept

Split-second decisions

A voice agent makes several small decisions in every turn of a call. Handed to a small decision model, about 39 in 40 came out right, each in under a quarter of the time allowed.

On the phone, a second of silence feels like a dropped line.

39 in 40about: mean accuracy across the kinds of decision on the first run, before any tuning
< 1/4time taken by a typical decision, against its allowance
alldecisions with no errors: routing to a specialist, and confirm-before-acting
proprietary · the result is public in relative terms · the figures on requestWhy, and how →ArticlePaperRequest access
◇ private · result shown, method kept

BigKids

A maths and English tutor for young children that reasons before it drills. In simulation, its lesson planner now gives a child a lesson they are not ready for about three quarters less often.

An exercise that is too easy bores a child, and one that is too hard defeats them.

≈ −75%lessons given before the simulated learner was ready for them
≈ −25%exercises shown again too soon
↑the planner's combined score moved from below zero to above it
proprietary · the result is public in relative terms · the figures on requestWhy, and how →ArticlePaperRequest access