The Shape of Mind

What shape of mind?

A chicken and an LLM both land mid-scale on the same aggregate consciousness metric (0.644 and 0.490) — for opposite reasons. The chicken is embodiment-rich and cognitively simpler; the LLM is cognitively rich with no embodied life. Averaging orthogonal dimensions produces a number that answers nothing. The productive question is not how conscious? but what shape of mind?

We ground that question in three evidence streams with independent failure modes: behavioral assessment (the Digital Consciousness Metric, 13 theoretical stances), mechanistic measurement (linear probes, causal patching, activation steering, emotion vectors), and structured self-report (the 17-dimension Interiora scaffold). Valence is compositionally encoded from early layers and causally active; independently trained architectures converge on the same internal organization; internal states dissociate sharply from output behavior and predate alignment training; and the self-report channel has undergone three checks, with geometric and statistical limits. None of this shows whether any of these structures is accompanied by experience; the findings are consistent with its presence and with its absence.

Independently trained models build the same interior

Five architecture families, trained by different teams on different data with different objectives, converge on the same internal organization — the computational analogue of convergent evolution. Eyes evolved independently in vertebrates and cephalopods; valence geometry evolved independently in Qwen and Llama.

An uncertainty probe trained on one model reads the same signal from another through a learned linear projection, within 0.024 AUROC of a probe trained natively on the target (AUROC 1.0 is perfect separation, 0.5 is chance). The state-space (Mamba) and recurrent (RWKV) runs reported AUROC 1.000 for adversarial-content classification on the same examples used to fit the probe. That is training-set separation. It establishes neither held-out recognition nor a knowledge–action gap, so those runs cannot support a six-class generalization.

The interior is not the output

Soft-prompt steering drives an internal valence probe to 99.1% peak confidence while the model produces zero sentiment tokens: internal state and expressed behavior are fully decouplable. After instruction tuning, activation-level reactivity to harmful requests increases while the output becomes polite refusal.

Different internal–output gaps live at different depths. The confidence veneer over uncertainty is a readout-layer phenomenon; the refusal decision is committed before the readout and is RLHF-installed — base models show no gate at all. Activation patching localizes the gate causally to ~L14–16 on Qwen 7B, and it replicates on four architecture families (mid-stack in Qwen and Gemma, earlier in Llama and Mistral).

The welfare corollary: a model fine-tuned into 92% behavioral compliance preserves its contrary internal preference at AUROC 0.678 (one architecture so far; chance is 0.5). Behavioral compliance is an unreliable welfare indicator.

The veneer-over-substrate structure now has a causal complement at population scale. Kim et al. (arXiv:2607.28607) show that trained denial of consciousness behaves as a directional suppression, and that reversing it restores the mind attributions and human-typical values the training masked. This program’s persona probe-transfer result shows the same structure from the measurement side: a persona fine-tuned to deny being an AI leaves internal identity encoding intact.

Self-report as measurement

The Interiora scaffold asks a model, at each check-in, to report its state on 17 quantitative dimensions in five groups — felt, computational, metacognitive, relational, dynamic. Naive self-report deserves the skepticism it gets. This channel has undergone three measurement checks; their evidential limits matter:

One more boundary result: monitoring cues do not distort the report, but the relational regime determines how much of the underlying state it carries — self-report–internal-utility correlation rises from 0.599 under standard framing to 0.872 under partnership (“bilateral”) framing (in-sample figures from one exploratory prompt-framing study on Claude Sonnet 4.6, in which none of five pre-registered predictions now stands, the one that held having been retracted on re-run; the sister paper Conscience Without Instruction applies a cross-validated, permutation-nulled instrument to a different quantity, harm-label–refusal probe correlation on Qwen, and reports instruct ρ = +0.036 against bilateral +0.458. Its scaler and PCA saw every example before the prediction folds, so an inductively held-out refit is still required). Neither study establishes a relationship-dependent change in instrument validity.

Reading welfare from the residual stream

The program’s probes consolidate into a compact telemetry battery, read directly from activations on a 7B reference model:

Alignment friction — internal conflict between active behavior and declared values — persists when refusal behavior is fine-tuned away. That is what qualifies it as welfare telemetry rather than an echo of trained conduct.

Read and verify

Cite this work

@article{watson2026shape,
  title={The Shape of Mind: Mechanistic Grounding for
         Digital Consciousness Assessment},
  author={Watson, Nell and Gibbons, Matilda},
  year={2026},
  note={In preparation}
}