§ 1.1 — P(tₙ | t<ₙ)
Token-level commitment
A model writes one token at a time, and every token narrows what can follow. The first plausible direction becomes the direction it commits to.
Divergent Labs
Neurodivergence-Aligned
Reasoning Research
Modern AI research is drifting towards a deterministic reasoning approach. At Divergent Labs, we challenge this by building neurodivergence-aligned reasoning frameworks.
Fig. 1 — Serial search commits once and walks it out; divergent search runs N isolated branches, prunes, and converges.
Adopted in the open-source agent ecosystem
repowiremstackzk-flow-osshanapp-librarystriatumawesome-promptsnix-skillscaioniehues/adhdktg-one/adhdwtfismyrepomythifygodauditsgodplansClaude1.0
00 · Thesis
Guard railed divergence defeats deterministic reasoning by big margins
The strongest lateral thinking on hard, open problems holds many threads at once and sits with a half-broken idea until the good one turns up next to it — a lot of it shows up in ADHD and adjacent cognition. We build that in: many candidates in parallel, each under a different frame, isolated so none can anchor to the others. We publish the harness before we train the model.
01 · Problem
Early commitment, four ways.
When there is no known answer, committing fast is a failure mode — not thinking too little, but committing too early. Everything downstream inherits it.
§ 1.2 — chain-of-thought
Longer reasoning, same anchor
Chain-of-thought reasons forward from the same early commitment. More steps do not revisit the choice that framed them.
§ 1.3 — tree-of-thought
Breadth without independence
Tree search widens the frontier but carries the same anchor down every branch. Shared context is shared bias.
§ 1.4 — scale
Fluency is not search
A larger model mostly says the first idea more fluently. It hands back the first reasonable answer instead of the best one.
Fig. 2 — Harness topology: one stimulus, N isolated branches, one critic pass.
Divergence, not a prompt
The harness sits around any capable model. Nothing is bolted on at the end — isolation and critique are structural.
02 · First research
adhd — a reasoning harness
An open-source reasoning harness for coding agents, written up in a preprint. Three stages.
Stage 01
Diverge
N branches at once, each under one of 15 cognitive frames, no shared context while they generate.
N × frame, ctx = ∅
[ Read the code ]Stage 02
Critique
One critic pass scores every candidate and flags the traps — ideas that look right and quietly break.
score → prune → flag
[ See the evals ]Stage 03
Deepen
Survivors expand into concrete plans, with the frame and trace that produced them kept inspectable.
survivors → plans
[ npm i adhd ]Fig. 4 — head-to-head vs. single-shot baseline, 0–10 scale, 6 open-ended engineering problems
| Dimension | adhd | Baseline | Δ | Ratio |
|---|---|---|---|---|
| breadth | 9.00 | 4.83 | +4.17 | 1.9× |
| novelty | 7.83 | 2.67 | +5.17 | 2.9× |
| trap detection | 9.50 | 1.83 | +7.67 | 5.2× |
| actionability | 9.50 | 6.50 | +3.00 | 1.5× |
| builder usefulness | 7.67 | 6.83 | +0.83 | 1.1× |
2.5K+
stars
200+
forks
10,000+
deployments
A claim like “reasons better” is worth nothing you cannot measure. Judged head-to-head against a single-shot baseline across six open-ended engineering problems, adhd won five of six — the loss was a problem the baseline already knew cold, the honest number to report.
03 · The founder
Systems that keep producing
Fig. 3 — Udit Akhouri, 64×40 density grid.
Founded by Udit Akhouri — Ex-AI systems engineer at Emora Health, CS at IIT Patna. The thread: systems that keep producing value long after they ship.
Emora HealthExthalpyADHDIIT Patna