Why one physical state is not the same as many records kept in step
Like weights joined by springs. Each part's graded state pulls on its neighbours. A change that exceeds the part's own noise moves the others, and the coupling restores agreement. The parts are coordinates of one physical state.
Like separate ledgers kept consistent by a clerk. Each part keeps its own record, held in place by a margin, and a rule compares and corrects the records at discrete steps. Below the margin the records vary independently. They are many systems kept in step by a protocol.
A clocked digital chip keeps its bits as separate records, and its program can bring their contents into agreement. The paper's premise asks whether the carriers of overlapping contents also respond to one another above noise and restore agreement within the content's time window.
P is proposed as a necessary condition for unity of contents already taken to be conscious. It specifies the physical grain at which carriers and their readers must respond. A simulation can reproduce a coupled field's course while internal bits held within their margins fail this response test. A converter that samples a graded signal finer than its noise is judged by the same test at its actual downstream reader.
A bit is read through a noise margin: any voltage well above the threshold reads as 1, so small wobbles change nothing. That margin is what makes digital hardware reliable, and it is also what separates it from a coupled field. If content is read through a margin, small enough physical changes never reach it. The proof is short: continuity carries a small enough neighbourhood of each region's physical state into the neighbourhood on which the readout is constant. The result covers any content read from bits, whatever algorithm the bits run.
Real hardware is noisy, so the paper states the criterion on probabilities. The noisy margin theorem bounds the change in the next content's distribution by the margin exception probability: the chance that a noisy successor comes close to a threshold. The bound is attained. Near approaches count even when the bit remains correct, so this probability can exceed the logical error rate. A readout fails the response test when that probability is below the declared fluctuation yardstick and a change larger than the source's noise still fits inside its margin. A cortical spike time can pass the same test near threshold: one weak synaptic input shifts it by more than its jitter.
Each gate does pull a disturbed voltage back toward its rail, but only toward its own logical value. That restoring pull keeps each bit inside its margin and pulls no region toward agreement with another, so the chip's records stay separate even where its physics settles.
The line P draws is between symbols, contents read through a margin, and physical contents, graded quantities the system's own dynamics reads. It concerns how a substrate is organised, not what it is made of.
| Symbols (agreement computed) | Physical contents (agreement can be enforced) |
|---|---|
| Bits read well clear of their thresholds, including those in a transformer on a GPU | Spike times and membrane potentials read near threshold in cortex |
| Spike counts latched at a clock edge | Column currents of an analog crossbar settling through shared conductances |
| A neural circuit read only from all-or-none events latched far from threshold | Graded currents of an analog circuit or phases of coupled oscillators, when their readers and restoring coupling pass the test |
A switching node can be sensitive at its voltage threshold, but a clocked latch reads its restored bit outside the setup-and-hold window. The verdict depends on the actual reader, its noise and its integration time.
If each patch of cortex is tied only to its near neighbours, agreement passes from hand to hand and leaks away with distance. Picture the sheet as an electrical network in which each coupling is a wire: phase disagreement between two sites is controlled by the resistance between them. Under short-range coupling the bound grows with the logarithm of distance; sufficiently long-range coupling can sustain agreement across the sheet. Turning phase agreement into agreement in content requires an encoder connecting the two. Across centimetres that step rests on the companion paper's bridge assumption. The paper also predicts restoration within the content window and a cortical relaxation scale of 166 ms.
Assumption R links calibrated human reports of one simultaneous field to experienced unity under these interventions. Animal matching tests integration and the physical mechanism; it does not establish the experiential claim. Human reports are calibrated against simultaneous and alternating displays, and a switching model is tested on held-out reports; if the models cannot be distinguished at the measured time scale, the experiential comparison is inconclusive. Preserved unity reports when a bridge is certified to fail P challenge P and R together. An analog advantage favours P only if the rival model fails its predeclared prediction.
P is a proposed necessary condition for unity, not a derived result or a criterion for which contents are conscious. Its response bounds and bridge test do not settle why physical processes are experienced. The human bridge test depends on Assumption R and on controlling visibility, attention and report ability; the proposed animal preparation first tests the physical mechanism and behavioural integration. The paper's limitations section gives the assumptions and limits in detail.
Every formal result the paper states is machine-checked in Lean 4, in the same development as the companion paper, and an appendix of the paper maps each result to its declaration. The development declares no axioms of its own, and an audit target fails the build if one appears.
The Lean development, estimate and simulation scripts, tests and saved numerical summaries are in the project repository. Publication values are generated from those saved artifacts. The whole-cell recordings are available from the Kiritani Zenodo deposit; the repository fetches and verifies them but does not redistribute the recordings.