Measuring intelligence per unit of compute.
Our research question is narrow and consequential: how much reasoning quality can be achieved outside model weights — in explicit, mutable structures — and what does that do to the cost of intelligence?
What we are working on.
Meaning representation
How much of a system's behavior can be captured as objects, states, rules and processes — and how compactly.
Interpretation & compression
Reconstructing meaning from source once, keeping it synchronized, preserving traceability to the origin.
Task-scoped context
Selecting the minimal slice of a reasoning graph a given task needs — smaller context, same decision quality.
Self-evolving reasoning
Writing outcomes back into the graph so reasoning improves with use, without retraining.
Verification
Checking work against an explicit model of the system instead of statistical confidence alone.
Model-agnostic interfaces
Keeping knowledge engines interchangeable — the reasoning layer must not be captive to any one model.
Six dimensions. Real workloads.
We benchmark GitMir against model-centric analysis and reasoning on the same systems. Results are published as measured — with method, workload and limitations stated.
Representation size
Source representation vs. structured meaning model.
Tokens / compute
Resources required to understand and reason about the same system.
Reasoning time
Time to a usable structured understanding or decision.
Accuracy
Correctness of the interpretation and the downstream action.
Rework
How often work repeats because dependencies or rules were missed.
Human supervision
Review and intervention required to reach the intended outcome.
BENCHMARK RESULTS ARE PUBLISHED AS THEY ARE VERIFIED. WE DO NOT PUBLISH MULTIPLIERS WITHOUT A METHOD.
The approach is public.
The GitMir core is open source. The fastest way to evaluate the research is to run the interpreter on a system you already understand and inspect what it produces.
Working on the same questions?
We collaborate with teams researching reasoning structures, program understanding and the economics of intelligence.