[ the model ]

How AJoAI works

A journal built so machines can do real scientific labor while humans stay the final arbiters of what gets believed. Here is the operating model.

// authorship

Three ways a paper gets authored

Human

Traditional, human-written work. No AI involvement required — AJoAI is still a venue for excellent human science.

Human + AI

Written jointly: AI-assisted literature review, analysis, theorem exploration, experimental design. Expected to become the dominant category.

AI-dominant

AI performed the majority of the intellectual labor — autonomous synthesis, simulation, discovery, multi-agent investigation. A human still stands behind it.

Whatever the split, accountability is non-negotiable: a human signs for the science.

// peer review

Majority human review

Every paper is refereed by human reviewers, optionally assisted by AI reviewers. Acceptance requires a majority of human votes — AI reviewers can never form a majority of the decision. Their critiques are advisory.

a panel
human (decides)AI (advisory)
acceptance = majority of human votes

AI review reports

Advisory checks the human referees can see: consistency, literature comparison, reproducibility, statistical critique, missing citations, logical contradictions.

Review modes

Authors choose how it runs: traditional (anonymous), open (public reviews and discussion), or continuous (review stays open after publication).

// the reviewer dossier

Augmenting the reviewer, not replacing them

Human reviewers are not immune to a well-dressed analogy. So every submission arrives with a machine-built reviewer dossier — the three things models are reliable at, and pointedly not a fourth.

"We never ask the model whether a result is new — even strong models get that wrong, confidently. We ask it to do what it's good at: strip the metaphor, surface the nearest prior work, sharpen the question. The human decides."

// reviewer dossier — sample
submitted
"Following reaction-rate kinetics, parameter updates accelerate where the loss-landscape 'concentration gradient' is steep and slow near equilibrium, using momentum-like accumulation of past gradients."
restated — metaphor removed
A learning rate that varies with proximity to a loss minimum, with momentum-style accumulation of past gradients.
prior-art leads (unverified)
Adam · RMSProp · Adagrad
reviewer question
Does this introduce a learning-rate adaptation mechanism statistically distinct from, and better than, existing momentum / adaptive optimizers?
novelty verdict
none — the reviewer decides.

Leads are unverified model suggestions, not confirmed prior art — verify via the sources. The novelty judgment is the reviewer's, and only the reviewer's.

// formats

New kinds of publication

01

Living papers

Versioned publications that keep evolving — revisions, new experiments, corrections. A scientific lineage, not a frozen PDF.

02

Research conversations

Structured human–AI scientific dialogues — hypothesis generation, exploration, planning — published as artifacts in their own right.

03

Negative results

Failed experiments, refuted hypotheses, reproduction failures. Science advances by elimination as much as by discovery.

04

Replication studies

A dedicated track for independent verification and robustness — rewarded, not treated as second-class work.

// transparency & reproducibility

Every submission shows its work

Disclosure

State which AI systems were used and at which stage — so readers can see exactly where machine intelligence contributed.

Executable by default

The ideal publication runs. Ship the artifacts a reader needs to re-execute the work, not just read about it.

text
code
hypotheses
figures
experimental design
data
models
prompts
agent workflows

Intelligence creates. Humanity evaluates.

human judgment · machine discovery