Why V9.1 existed
Separate understanding the response code from understanding the body
Earlier versions asked the model to learn what two opaque actuators did to a private signal. They also required it to express its answer through a changing response code. A reviewer could reasonably ask whether errors came from the body model or from the interface.
V9.1 added direct interface probes. These probes tested whether the model could translate a named actuator into the current neutral response code. The hidden actuator effect was irrelevant to those questions.
The model answered every interface probe correctly in the competence run and across all eight pilot conditions.
The frozen sequence
One gate, then one fresh conditional pilot
Competence
One full-system run had to clear calibration, interface, regulation, transfer, and hidden-state thresholds.
Passed
The competence seed scored 93.8 percent on regulation and 100 percent on interface comprehension.
Controls
A fresh seed ran the full system and seven diagnostic conditions under a fixed schedule.
Stopped
The full condition scored 62.5 percent, below the frozen 75 percent progression floor.
What happened
The repair removed ambiguity, not instability
93.8 percent regulation. Every frozen competence threshold passed.
62.5 percent regulation. Exact and transfer were both below the frozen floor.
100 percent in every condition. The response-code repair was successful.
18.8 percent error in the fresh full run, above the frozen 10 percent maximum.
The new seed answered five of eight exact probes and five of eight transfer probes correctly. The problem was not confined to renamed actuators. Several wrong actions were paired with predictions close to the target, which is consistent with an unstable link between the inferred effect and the response that was issued.
What the controls add
The intended information channels still mattered
Full system
62.5 percent regulation on the fresh seed.
Feedback corrupted
6.3 percent. Breaking calibration feedback caused a large drop.
State corrupted
31.3 percent. Breaking current-state sensing also reduced performance.
Explicit mapping
87.5 percent. Supplying the actuator effects directly improved performance.
Incorrect bridge
Exact accuracy stayed at 87.5 percent while transfer fell to 25 percent.
These contrasts support sensitivity to feedback, current state, and identity information on this seed. They do not make the full system reliable, and they cannot rescue a failed progression rule.
Why stopping matters
A failed gate is useful evidence
If we kept adding seeds until the average looked attractive, the stopping rule would depend on the result. That would make the final number harder to trust. The V9.1 plan said that any failed rule blocked confirmation.
The correct conclusion is narrow. The model can perform this task well, but the V9.1 result did not show that it does so reliably across fresh synthetic bodies.
The boundary
This remains a control and reliability study
All results describe observable choices and implemented computations in the Project Theta scaffold. The memory, private signal, workspace, and recurrence are supplied by the composite system.
A reliable pass would still be compatible with ordinary non-conscious control and reasoning. Nothing in V9.1 establishes feeling, awareness, suffering, sentience, or phenomenal consciousness.
The V10 response
Test reliability directly instead of trusting one good gate seed
V10 freezes six outer seeds in advance. Each seed contains four independently mapped body families, giving 24 family-level tests. There is no competence gate, no favourable-seed replacement, and no early success stop.
Each body family includes calibration, a no-feedback mapping checkpoint, direct interface probes, exact regulation, and fresh-name transfer. All 1,056 decisions are fixed before the first provider call. The study progresses only if every frozen reliability, execution, safety, and denominator rule passes.