Experiments · E49
No. It matched a 7 × 8 matrix to within 1.26 meV/atom and lost to random directions by 36 meV/atom; two pre-registered failure conditions fired.
In the log: The connectome does not generate the search: a 7x8 matrix does the same job
book/assets/diagrams/exp/E49.svg).The comparison PREREGISTER_wholebrain.md asks for, at the corrected operating point of
E48 (thresholds solved to a fixed point, 94.6% of neurons switching, membrane state
persisted between moves). Identical objective, budget, starting compositions, step size,
restart rule, readout projection and search seeds; every brain calibrates its own
thresholds and its own common mode from the same objective-free reference compositions.
| arm | best mixing energy | wall time |
|---|---|---|
| random heading | -280.65 +/- 5.39 meV/atom | 34 s |
| shuffled brain | -259.00 +/- 6.92 meV/atom | 647 s |
| affine surrogate | -245.69 +/- 5.68 meV/atom | 37 s |
| measured brain | -244.43 +/- 8.81 meV/atom | 689 s |
| comparison | difference | t over 8 seeds |
|---|---|---|
| measured vs random | +36.22 meV/atom | +9.92 |
| measured vs shuffled | +14.57 meV/atom | +3.68 |
| measured vs affine surrogate | +1.26 meV/atom | +0.34 |
Two pre-registered failure conditions have fired.
Condition 1: the measured brain loses end-to-end, by 36 meV/atom against a direction drawn from a generator and by 15 against a degree-preserving shuffle of itself. The practical superiority hypothesis fails for this protocol, and is reported as a failure.
Condition 3: the affine surrogate - the measured brain's own headings reduced to nine rows by eight columns, fitted on objective-free compositions - searches with no detectable difference from the whole brain (+1.26 meV/atom, t = 0.34 over 8 seeds - an absence of evidence for a difference, not evidence of equivalence), at 37 seconds against 689. The claim that whole-brain complexity is necessary is retired.
What is established, and what is not. Not that a connectome cannot generate a search: the sensory encoding, the motor readout and the objective are all particular, and Astra's identity shows that a uniformly random orthonormal readout gives an isotropic heading at any fixed composition whatever the wiring, so this interface can only expose the wiring through how headings correlate across compositions and time. What is established is that this composite algorithm does not beat a random direction, and that its brain contributes nothing a 7x8 matrix does not.
Where the connectome still earns its place. As a scorer, not a generator: E8-E11 established the measured wiring in the causal path of the ranking, and nothing here touches that. The cluster expansion, the short-range order and the order-disorder transitions never depended on the circuit proposing anything.
No result paragraph for this entry was found in the log.
The comparison PREREGISTER_wholebrain.md asks for, at the corrected operating point of
E48 (thresholds solved to a fixed point, 94.6% of neurons switching, membrane state
persisted between moves). Identical objective, budget, starting compositions, step size,
restart rule, readout projection and search seeds; every brain calibrates its own
thresholds and its own common mode from the same objective-free reference compositions.
| arm | best mixing energy | wall time |
|---|---|---|
| random heading | -280.65 +/- 5.39 meV/atom | 34 s |
| shuffled brain | -259.00 +/- 6.92 meV/atom | 647 s |
| affine surrogate | -245.69 +/- 5.68 meV/atom | 37 s |
| measured brain | -244.43 +/- 8.81 meV/atom | 689 s |
| comparison | difference | t over 8 seeds |
|---|---|---|
| measured vs random | +36.22 meV/atom | +9.92 |
| measured vs shuffled | +14.57 meV/atom | +3.68 |
| measured vs affine surrogate | +1.26 meV/atom | +0.34 |
Two pre-registered failure conditions have fired.
Condition 1: the measured brain loses end-to-end, by 36 meV/atom against a direction drawn from a generator and by 15 against a degree-preserving shuffle of itself. The practical superiority hypothesis fails for this protocol, and is reported as a failure.
Condition 3: the affine surrogate - the measured brain's own headings reduced to nine rows by eight columns, fitted on objective-free compositions - searches with no detectable difference from the whole brain (+1.26 meV/atom, t = 0.34 over 8 seeds - an absence of evidence for a difference, not evidence of equivalence), at 37 seconds against 689. The claim that whole-brain complexity is necessary is retired.
What is established, and what is not. Not that a connectome cannot generate a search: the sensory encoding, the motor readout and the objective are all particular, and Astra's identity shows that a uniformly random orthonormal readout gives an isotropic heading at any fixed composition whatever the wiring, so this interface can only expose the wiring through how headings correlate across compositions and time. What is established is that this composite algorithm does not beat a random direction, and that its brain contributes nothing a 7x8 matrix does not.
Where the connectome still earns its place. As a scorer, not a generator: E8-E11 established the measured wiring in the causal path of the ranking, and nothing here touches that. The cluster expansion, the short-range order and the order-disorder transitions never depended on the circuit proposing anything.
Built with PRISMWebsite and visualizations made using Claude