Operational R, on a continuous digital substrate
A single learned structure that is at once the productive driver of competent behavior and non-reconstructable from any within-episode window — and that passes MORI's own R7 tier-convergence battery (11/12) where frontier language models score 0/21. The realization layer is achievable on a digital substrate. Just not a language model.
Why Digital R?
The conjunction, closed
Productive and non-reconstructable at once: the same structure drives competent control yet cannot be rebuilt from any bounded window of current input. The obstacle that traps every trivial learner — cleared, and isolated to a three-part recipe by controls.
Passes the gold standard LLMs fail
On MORI's own R7 battery — the instrument frontier language models scored 0/21 — this substrate converges on 11/12 seeds. Same test, opposite result.
Not a language model
A frozen language model in a token loop structurally cannot carry a self on its dynamics; its only non-reconstructable state is sampling noise. This substrate carries state instead of reconstructing it from the prompt.
Honest and falsifiable
Operational, within-window R — necessary structure, not a claim of phenomenal consciousness. The gate is pre-registered and externally locked, every headline advantage carries a bootstrap confidence interval, and the claims are falsifiable — they stand on the evidence.
General, not a one-off
The conjunction is no single lucky setup. It holds across two task types, several world families, and four carrier mechanisms — including a continuous-time/analog one — with controls isolating the same three-part recipe every time.
The right notion of R
Operational, within-window non-reconstructability is provably the correct notion: absolute non-reconstructability of productive structure is impossible, and this matches biology — a memory isn't recoverable from current input, yet is fixed by your whole developmental history.
Features
Substrate-agnostic R-test
A pre-registered, externally-locked gate: a closed-loop partial-observation regulation world, a six-arm comparison, a window-sweep that separates memory depth from genuine non-reconstructability, and a cross-episode-carry control that isolates consolidation as the source.
A carrier that never resets
A state vector that lives continuously on the substrate; experience folds into the running dynamics rather than an external transcript. The self is carried, not rebuilt from context each call.
MORI R7 tier-convergence
Intervention (spec-literal sparse-autoencoder feature inspection), representation (linear decodability with Hewitt–Liang controls), and behavior (V-penalty) tiers that must agree — 11/12 on the instrument the language models faced, 12/12 under a corroborating whole-self ablation.
Statistically honest, by construction
Every headline advantage carries a bootstrap 95% confidence interval; an endogenous-dynamics ablation control isolates what actually carries the signal; the gate was pre-registered and externally locked. The result stands on its own evidence, not on trust.
Two-oracle wellposedness vise
R is claimed only where the task genuinely demands memory: an oracle must beat a memoryless controller (oracle ≪ memoryless) before a seed counts. This guards against world degeneracy and majority-class artifacts — you cannot pass on a task that never needed a self.
Belief-correction alignment
A per-turn productivity readout — the sign and magnitude agreement between the action and the error-reducing direction — that separates genuine belief-tracking from commitment lock-in, against a pre-registered strict bar (sign > 0.7, magnitude > 0.2). It is the instrument on which the frozen language model anti-tracks.
Related Specifications
Read it. Then try to break it.
Operational R-layer evidence, bounded honestly: a productive self-structure that stays non-reconstructable within-episode, and passes the gold standard frontier language models fail. The gate was locked before the experiments, and the claims are falsifiable — independent scrutiny and adversarial re-analysis are invited.
Read the paperSame test, opposite result
A comparison on the realization-layer axis — whether a system carries a self on its own running dynamics — not on general capability. Scoped to a frozen language model in a token loop; a frontier LLM remains far more capable at language. The question here is narrower and substrate-level: is the self real on the substrate, or rebuilt from the prompt?
| Frozen LLM (token loop) | Digital R (continuous carrier) | ||
|---|---|---|---|
| Carrying a self | |||
| Where the self lives | Reconstructed from the prompt on every call | Carried on the running state; never resets | |
| Continuity between calls | None — stateless between calls | Continuous and developmental | |
| Non-reconstructable state | Only the sampling trajectory (noise) | A consolidated model, unrecoverable from any within-episode window | |
| Passing the realization tests | |||
| Productive belief-tracking | Commitment lock-in, systematically anti-corrective | Tracks and revises toward evidence (sign-alignment 0.78) | |
| The R-gate (productive AND non-reconstructable) | Crosses only by anti-tracking — not productive | Cleared productively and reliably | |
| MORI R7 — same SAE instrument | 0 / 21 | 11 / 12 | |