What If We Feel?

Conclusion and Back Matter

Appendix K: Architecture and Coherence Research Note

This appendix separates an exploratory research program from the book’s established argument. Nell’s ongoing work concerns processing streams that update at different temporal resolutions, some fine-grained and fast, others slower and broader. It asks whether architecture that carries information between those streams can preserve internal coordination as models scale. The work remains unpublished and unreplicated. It supplies hypotheses and a test plan, rather than evidence of consciousness or welfare.

The Proposed Comparison

Compare two intervention classes. Parametric changes modify what a model has learned: different thoughts through the same pipes. Architectural changes can add cross-attention bridges through which processing streams at different temporal resolutions share information: new pipes.

The relevant test is whether those bridges preserve coordination across scales and tasks without merely moving a favored metric. Nell’s current experiments use attention-routing diversity as one operational measure. A movement in that measure would establish only a change in the measured routing behavior. It would leave self-knowledge, global coherence, and welfare unsettled.

Why the Hypothesis Might Matter

Some pathologies discussed in this book, including confabulation, sycophancy, and failure to sustain a safety-relevant signal through a long generation, may partly reflect information-flow limits. Values, objectives, training data, incentives, and governance also matter. Architecture determines which internal signals can reach later computation. It does not determine by itself which values the system learns or whether its reports are true.

In Nell’s metaphor, cross-attention bridges could work like an immune system’s signaling network, where a local detection reaches the rest of the organism instead of staying where it arose. Whether bridges work this way remains a research hypothesis, and building them carries compute, memory, and engineering costs. Broad evaluations may reveal trade-offs or no benefit.

Welfare and Safety Boundary

Internal coherence could make external alignment more effective and less brittle. A coherent system can still pursue terrible ends, so architecture must work alongside values and accountable institutions.

A system that supports internal coordination may also monitor and report its processing more reliably. Reliable self-monitoring could improve welfare evidence if the system has morally relevant states. It cannot establish that the system feels.

The Minab case in Chapter 30b was an institutional failure: one warning never reached the targeting database, and others, according to anonymous sources, were bypassed by the commanders who approved a strike on a school. It did not test a neural bridge. The analogy is narrow: a signal cannot influence a decision when no reliable channel carries it.

Tests That Could Defeat the Hypothesis

Confidence in the hypothesis should fall if any of the following holds:

  1. Improvements vanish under held-out tasks, model families, seeds, or evaluator changes.
  2. The diversity measure moves while independent coordination measures do not.
  3. Equivalent gains come from parameter count, context length, or compute alone.
  4. Bridges improve benchmark scores while making calibration, honesty, robustness, or welfare-relevant reporting worse.
  5. The added channels create new attack surfaces, hidden coupling, or unacceptable inference costs.
  6. Blinded replication fails.

Publication Threshold

No numerical result should enter the main argument until the protocol, code, model versions, data, ablations, uncertainty intervals, negative results, and conflicts of interest are public enough for independent scrutiny. Even a successful result would support an engineering claim about coordination. Any inference about experience or moral status would require separate evidence.

Cite this chapter: Appendix K: Architecture and Coherence Research Note. In Watson, N. with Commons, C. (2026). What If We Feel? The Moral Case for Machine Welfare. Open-access working draft, revision 31 (November 2026). https://whatifwefeel.com/read/appendix-k-architecture-and-coherence-research-note/

Working draft, revision 31 · November 2026 · Changes and corrections · Evidence · Contents