GOBenchmark
English
Independent global nonprofit initiative

How do you measure AI transformation?

G2AI-driven workG4Gödel Organization↓

Test organizational self-improvement through real work.

Once every department uses AI, two questions remain. How much core work can AI carry from the original request to an accepted result? When something goes wrong today, can the organization learn enough to prevent it next time?

Open standard · v0.1

What earns a level?

GO Benchmark is an open method for organizational self-assessment. Teams draw evidence from actual task records, use an agent to organize it and run a versioned scorer. G2 asks whether AI delivers most core work end to end. G4 asks whether the organization can repeatedly verify its own improvements. G5 explores whether the improvement process itself can evolve.

Select a level to see its definition and evidence requirements.

GÖDEL MACHINE

When should a machine rewrite itself?

In 1931, mathematician and logician Kurt Gödel encoded formulas and proofs as numbers in his work on incompleteness. Arithmetic could then express statements about proofs themselves. This idea of self-reference later inspired a question in computer science. Could a program reason about and improve its own code?

In 2003, computer scientist Jürgen Schmidhuber proposed the Gödel machine. This theoretical machine carries out tasks while searching for changes to its code. Under its stated goals and axioms, it must prove that making a particular change now has greater expected utility than continuing to run the current program and search for alternatives. It applies the change once it finds that proof.

The proof-search procedure is itself code that the machine can rewrite. This illustrates recursive self-improvement (RSI). One improvement can change how the system finds the next. The proof depends on its axioms, assumptions about the environment and utility function. A useful change may remain unproved.

Gödel Organization (GO) takes its name from that ambition. An organization can learn from the work it does and eventually improve how it learns. GO Benchmark traces that progression through evidence of work delivered by AI, lasting changes to workflows and improvements to the methods used to test those changes.

Schmidhuber’s Gödel machine paper ↗ Kurt Gödel at the IAS ↗
GÖDEL ORGANIZATION · GO

How can a team improve the way it improves?

A Gödel Organization makes its workflows visible. AI can propose changes within human-set limits, and the team can test them on later tasks. An organization or unit earns the GO designation at G4. The evidence comes from operational results, comparisons and a traceable record of decisions.

G4 · Improving the quoting process

Suppose AI finds a pattern in customer account statements. Quotes for remote destinations repeatedly omit shipping. AI proposes a new pricing rule and trials it with approval. On later orders, the team compares the old and new procedures, checking omissions and turnaround time before adopting or reversing the change. To qualify, it must also show that this cycle recurs across different problems.

G5 · Improving the way changes are tested

AI notices that trials of new rules mostly use commonplace orders. It proposes holding out test cases for each order type. When another discount rule comes up for review, the team spots errors affecting small customers earlier, with the same review budget. The capability under test is how well the organization can tell whether a change works. G5 is the benchmark’s research goal.

Case 01 · Launch team

How GO Benchmark assessed its own launch

Who did the work of launching the benchmark? We traced eight deliverables back to the requests, actions and receipts behind them. AI carried out 21 of 32 stages. That supports a retrospective G1 estimate for this launch window.

Read the case study

Start self-assessment

Give the task to your own agent. It reconstructs deliverables from existing requests, work records and acceptance receipts, then scores and replays the assessment locally.

Start self-assessment