The setup
Task sets, success criteria and baselines, written down before any run.
Research & developmentTestable questions
The research agenda
Intelligence should be
evaluated by outcomes.
Atlas investigates whether heterogeneous AI coordination can make work more reliable and efficient under explicit resource, quality, and authority constraints. These are questions we plan to test; no results are published yet.
One question. Multiple explanations.
Single modelFixed workflowAdaptive allocation
Matched constraintsSame tasks · resource budget · acceptance criteria
Independent evaluation
Proposed experimental structureResearch / WorkbenchControlled questions
Three questions.
Three test designs.
Choose a research direction to inspect the proposed experiment.
Proposed protocol
Bounded objective
Planning Contributor A
Execution Contributor B
Critique Contributor C
Independent acceptance
Example allocation
Does specialization beat simplicity?
Compare specialists assigned by task against a strong single-model baseline and a fixed multi-model workflow.
ControlMatched time, token/resource budget, task set and acceptance criteria
TestTask-dependent contributor allocation with separate evaluation
MeasureVerified completion, total cost, elapsed time and failures
What would challenge the hypothesisNo repeatable improvement over well-tuned simpler workflows under matched constraints.
How we judge resultsResearch method
Evidence is not
the model's opinion.
Strong comparative research needs baselines, honest failure reporting, and independent acceptance.
Protocol step / 01
Prespecify the work
Define the task set, success criteria and hold-out cases before running comparisons. Changing the goal after seeing results would undermine the evaluation.
01
Verified completion
Did the work satisfy acceptance criteria that were defined before execution?
02
Resource cost
What compute, tokens, time, and human intervention were actually required?
03
Independent judgment
Did checks or independent reviewers support the claimed outcome?
04
Reproducibility
Can the observation be repeated under a declared configuration and task scope?
Review / IndependenceAssurance before acceptance
Another model isn't
automatically a second opinion.
Atlas's planned reviewer-trial design would check who built the change, where the reviewer's model comes from, that the exact same version was reviewed, and whether the check actually ran. Explore the example cases.
Example protocol outcomeDisqualified
Self-review cannot establish independent judgment.
A reviewer who built the candidate is not independent. The reviewer-trial protocol rejects that pairing before scoring any findings.
Eligibility for a reviewer trial is not a merge permission or a finding of correctness. This is a schematic of protocol rules—not an actual review verdict.What we’ll publishOpen results
What we'll
publish.
The results
Completion, cost and time against a strong single model and fixed workflows, including where Atlas loses.
The failures
What broke, how it was caught and what changed.
The method
Enough detail for someone else to repeat it.