benchmarks evals
The Same Behavior Can Hide a Different Machine
Finite interventions can separate identical outputs from different causal mechanisms—but missing internal access creates uncertainty no sample size removes.
Summary
Finite interventions can separate identical outputs from different causal mechanisms—but missing internal access creates uncertainty no sample size removes.
Mechanism-level audits ask how a system computes, not only what it emits. In four- and five-unit stochastic networks with known ground truth, the authors study when integrated information can be identified and certified from finite, noisy interventions. A recurrent and a feed-forward network produced identical delayed input-output behavior while carrying 0.28 and 0 bits respectively; internal recordings separated them in 4 to 53 trials. When units were omitted from one-step recordings, however, the resulting 0.30-to-1.02-bit precision floors persisted regardless of sample size. The measure also changed sharply near partition boundaries, although one-sided lower-bound certificates remained possible. These small synthetic systems do not establish machine consciousness; they show that measurability depends on experimental access and design, not behavioral fit alone.
Why it matters
Finite interventions can separate identical outputs from different causal mechanisms—but missing internal access creates uncertainty no sample size removes.
Limits and context
- Mechanism-level audits ask how a system computes, not only what it emits.
- These small synthetic systems do not establish machine consciousness; they show that measurability depends on experimental access and design, not behavioral fit alone.
Key claims
Finite interventions can separate identical outputs from different causal mechanisms—but missing internal access creates uncertainty no sample size removes.
Qualification: Mechanism-level audits ask how a system computes, not only what it emits.
Evidence: source-2026-11-08-002
Sources
- arXiv preprint 2611.08469, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.