Log

The test

Tasks

87 web tasks from MiniWoB++.

Seeds

Each task is run with 5 seeds.

Scoring

Each attempt is scored by the task's own exact checker: success or failure, no partial credit, no judgement calls.

Separation

The test set is never used for training.

Integrity

The evaluator is sealed; a tamper lock checks the seal before and after every Hebb run.

Score history

Scale 0–100%; dotted line at 50%. Most of this gain came from Hebb's fixes to Nous's body, not from a smarter brain.

1. In progress

   Aegis A/B test. Aegis, the silent manager that picks what Nous practises and invents new practice goals, is being compared with a fixed curriculum. No result yet.

2. 2026-10-21

   Fixed-test score 57.5%. REFLEX (about 1.3M trained parameters on a frozen small text encoder, laptop CPU) scores 56.6% against 57.5% for a 4B language model, about 7× faster. On this test the two are tied.

3. 2026-10-20

   Fixed-test score 47.6%.

4. 2026-10-19

   Fixed-test score 30.1%.

REFLEX vs a 4B language model

REFLEX has about 1.3M trained parameters on a frozen small text encoder and runs on a laptop CPU, about 7× faster than the 4B model. On this test the two are tied.

What would count as recursive self-improvement

The goal is scaffold-level recursive self-improvement: Hebb improving its own tools and methods. It is not claimed yet. To claim it, all of the following must hold.

1. A locked evaluator

   The test and the code that scores it must be something Hebb cannot change. A tamper lock already guards it: it seals the evaluator and checks the seal before and after every Hebb run.

2. Hebb's own score over time

   Not Nous's score alone, but how well Hebb improves Nous: gains per run, measured on the locked test.

3. Compounding gains

   When Hebb improves its own tools, its later improvements to Nous should come faster or bigger. One good week isn't enough; the curve has to bend.

Prior art: STOP (self-taught optimizer), Gödel Agent, and the Darwin Gödel Machine.