A State Named Curiosity Still Needs a Test
09 November 2026
Control-instrumentation research proposal

Abstract
The public Encyclopedia of Praxium describes an affect-like internal state derived from telemetry and used to adjust system behavior. Its scientific reader guide interprets this mechanism as gain scheduling with human-readable names, not evidence that a server experiences emotions. This proposal focuses on how such a controller could be tested: identify the measurement, the transformation and the intervention separately. No controller was implemented or evaluated for this article. [1, original pp. 6–7, §7]

The name is not the mechanism
The guide describes CPU load, memory pressure, latency, process fanout, fragmentation and errors being mapped into a high-dimensional internal state. Bounded readouts are assigned names such as curiosity, serenity or threat. The names can make an interface easier to discuss, but they do not specify the mapping or demonstrate a psychological state. [1, p. 6]

A technical record would need the input units, sample intervals, preprocessing, matrix or function parameters and output range. Without those details, two systems could display the same label while reacting to different evidence. Conversely, two differently named variables could implement the same numerical rule. Renaming a signal should not change its scientific status.

The intervention creates the claim
The reader guide gives examples: threat may reduce exploration, curiosity may broaden search and serenity may reduce reactive oscillation. It says the metaphor becomes testable when each variable has a telemetry mapping, bounded range, intervention logic and measurable effect. This is a requirement for an experiment, not a report that those effects have been established. [1, p. 7]

For a bounded proposal, choose one output: permitted search breadth. Define a baseline budget and a rule that adjusts it from an instrumented state while never exceeding a fixed resource ceiling. A change in search breadth is then observable. Whether that change improves task outcomes is a separate hypothesis. Calling the state curiosity does not settle either question.

Correlation can reverse the story
Suppose a named state rises whenever workload is high, and completion time also rises. The correlation does not establish that the state caused the delay. Both may be consequences of workload. If the controller then broadens search, its action can alter the telemetry that created the state, closing a feedback loop. Analysis must account for that loop rather than treating the state as an independent explanatory variable.

A proposed experiment would compare matched workload traces under a fixed-budget baseline and the adaptive rule. It would record task completion, search count, latency and resource use. Randomized ordering or replayed inputs could help separate controller effects from changing demand. No traces were collected here, and the proposal does not claim that an adaptive rule will outperform the baseline.

Bounded is not automatically useful
The guide emphasizes bounded manifold states because unbounded gains can destabilize a control loop. Boundedness is an important design constraint, but a bounded policy can still be wasteful, slow or badly calibrated. A sigmoid can limit a readout while saturating over most of the observed range, leaving the system insensitive to meaningful changes. [1, p. 6, §§6–7]

Calibration should therefore inspect both the input distribution and the action distribution. How often does the output sit near a limit? Does a small measurement error trigger a large decision change? Are there operating conditions absent from calibration? These questions concern the actual mapping. A visually smooth activation curve is not evidence of robust behavior in the deployed environment.

Missing telemetry is a condition
A missing sensor reading should not silently become a neutral or reassuring state. A declared fallback could retain a conservative fixed budget while marking the adaptive output unavailable. The exact fallback would depend on the application, but it should be specified before failures occur. Otherwise, an apparent emotional interpretation can hide an ordinary instrumentation fault.

The experiment should include delayed, missing and out-of-range measurements. Report how the policy responds and whether it stays within its fixed resource limits. A successful fallback is not evidence that the latent state correctly inferred the missing information; it shows that the system handled uncertainty without pretending to know more than it did.

Interpretation
A positive result would support a limited engineering claim: under the tested workloads, the declared mapping and intervention changed a measured outcome relative to the baseline. It would not establish subjective feeling, consciousness or a universal model of human affect. The source explicitly separates the gain-scheduling interpretation from phenomenal-experience claims. [1, p. 7]

Conclusion
Human-readable state names can organize a controller, but scientific accountability comes from the measurement and intervention chain. The productive question is not whether the server is curious. It is whether a specified bounded rule improves a specified task, under conditions that another investigator can examine.

Reference
[1] Encyclopedia of Praxium, Volume 0 Scientific Reader's Guide and Integrative Journal Review, Source-Bearing Research Edition, 6 September 2026, original PDF pp. 6–7, §§6–7. Only these scientific pages are included; private/source archives and embedded attachments are excluded. All experimental procedures above are proposals.
