Experiments · E37
No. It generated alloys outside every list but gained 0.0123 of the score against the optimiser's 0.0750, a pre-registered loss.
In the log: The generative test: pre-registered, and a loss on both arms
book/assets/diagrams/exp/E37.svg).Date 2026-09-12 · Question A Gaussian process cannot generate - it ranks a list it
is handed. Can a learned generator produce crystals meeting several objectives, including
compositions no list contained? · Provenance PREREGISTER_generative.md,
…/generative.py
Why this differs from E35. That comparison had the circuit selecting from the same 4,000-point pool as the process, which removed the only structural difference between them. Here the circuit generates by run-and-tumble and the process does what it can do, which is rank a pool. The question standing behind it is whether a learned generator is viable at all - which is also the premise of the generative-flow-network work.
Result: LOSS, both arms. Hypervolume gained over the shared start, objectives normalised, matched greedy acquisition, seeds 800-809:
| arm | gained | standard error | left the pool |
|---|---|---|---|
| Gaussian process, selecting | 0.0750 | 0.0117 | 90%* |
| random generation | 0.0495 | 0.0085 | 30%* |
| the fly, generating | 0.0123 | 0.0040 | 100% |
| the fly, generating + the expansion's uncertainty | 0.0079 | 0.0032 | 100% |
-5.09 and -5.56 standard errors. The circuit generates compositions outside any pool on every seed and they are not good ones. Adding the expansion's measured uncertainty as an exploration term made it slightly worse, not better.
A flaw in one column, found after the run and reported rather than dropped. The process was given its own Latin hypercube pool (a different draw from the reference), so "left the pool" measures only that its draw contained points the reference draw did not dominate - not that it generated anything, which it cannot. That column distinguishes nothing as built; to mean what it was meant to, the process's pool must be the reference pool. The hypervolume comparison is unaffected.
Also settled, separately: the promotion signal changes nothing. Teaching the circuit a verdict - was this candidate worth its promotion, +1 or -1 - rather than an objective's value is the better fit for machinery built on appetitive against aversive dopamine, and was argued for on those grounds before testing. Measured: 0.7737 +/- 0.0184 against 0.7783 +/- 0.0142 for the value signal. No difference.
Standing count. Five fair configurations now - single-objective selection (E33), representation substitution (E33), multi-objective selection (E35), generative multi-objective (here), and promotion-based learning (here). The circuit has won none of them. Each loss has had a mechanism found and fixed, and the next has followed. That pattern is itself the finding: repeated repair with the outcome fixed in advance is how noise gets fitted, and a favourable result arriving after enough of it would not deserve belief.
Result: LOSS, both arms. Hypervolume gained over the shared start, objectives normalised, matched greedy acquisition, seeds 800-809:
Date 2026-09-12 · Question A Gaussian process cannot generate - it ranks a list it
is handed. Can a learned generator produce crystals meeting several objectives, including
compositions no list contained? · Provenance PREREGISTER_generative.md,
…/generative.py
Why this differs from E35. That comparison had the circuit selecting from the same 4,000-point pool as the process, which removed the only structural difference between them. Here the circuit generates by run-and-tumble and the process does what it can do, which is rank a pool. The question standing behind it is whether a learned generator is viable at all - which is also the premise of the generative-flow-network work.
Result: LOSS, both arms. Hypervolume gained over the shared start, objectives normalised, matched greedy acquisition, seeds 800-809:
| arm | gained | standard error | left the pool |
|---|---|---|---|
| Gaussian process, selecting | 0.0750 | 0.0117 | 90%* |
| random generation | 0.0495 | 0.0085 | 30%* |
| the fly, generating | 0.0123 | 0.0040 | 100% |
| the fly, generating + the expansion's uncertainty | 0.0079 | 0.0032 | 100% |
-5.09 and -5.56 standard errors. The circuit generates compositions outside any pool on every seed and they are not good ones. Adding the expansion's measured uncertainty as an exploration term made it slightly worse, not better.
A flaw in one column, found after the run and reported rather than dropped. The process was given its own Latin hypercube pool (a different draw from the reference), so "left the pool" measures only that its draw contained points the reference draw did not dominate - not that it generated anything, which it cannot. That column distinguishes nothing as built; to mean what it was meant to, the process's pool must be the reference pool. The hypervolume comparison is unaffected.
Also settled, separately: the promotion signal changes nothing. Teaching the circuit a verdict - was this candidate worth its promotion, +1 or -1 - rather than an objective's value is the better fit for machinery built on appetitive against aversive dopamine, and was argued for on those grounds before testing. Measured: 0.7737 +/- 0.0184 against 0.7783 +/- 0.0142 for the value signal. No difference.
Standing count. Five fair configurations now - single-objective selection (E33), representation substitution (E33), multi-objective selection (E35), generative multi-objective (here), and promotion-based learning (here). The circuit has won none of them. Each loss has had a mechanism found and fixed, and the next has followed. That pattern is itself the finding: repeated repair with the outcome fixed in advance is how noise gets fitted, and a favourable result arriving after enough of it would not deserve belief.
Built with PRISMWebsite and visualizations made using Claude