Develop
Build candidate mechanisms across the declared training worlds.
LOCAL RESEARCH RIG / SYNTHETIC LIFETIMES
Substrate Search
Search the mechanism.
Protect the test.
Explore memory, consolidation, context routing, trust gates, and contradiction repair. Keep the final test worlds outside the decisions that choose a candidate.
Select & freeze
Choose the candidate, baseline, and ablations. Seal the complete selection.
Measure once
Evaluate the frozen comparison on independent worlds. Retain the result.
CURRENT RESULT / 48 SEARCH TRIALS
A better experiment.
No winning candidate.
The selected candidate did not beat the validation-selected baseline on the three final test worlds. No discovery cleared promotion.
| Frozen comparison | Score |
|---|---|
| Selected candidate | 0.944390 |
| Uniform consolidation baseline | 0.975217 |
| Candidate minus baseline | −0.030827 |
A run paused after trial 17 resumed to exactly the same sealed result as the uninterrupted 48-trial control.
The substrate discovery command uses Python 3.11 or newer and needs no GPU or external model. From the extracted folder:
python -m pip install -e .
python -m simagotchi substrate-discover --trials 48 --seed 149 --out runs/demo --stop-after 17
python -m simagotchi substrate-discover --trials 48 --seed 149 --out runs/demo --resumeThe journal records trial state and RNG. Resume rejects changed protocols and complete-record corruption. A completed run returns its sealed result without silently rerunning the test.
EVIDENCE BOUNDARY
Synthetic results stay synthetic.
The older discovery loop used its “hidden” worlds during selection. Those old scores are selection evidence. This corrected protocol governs substrate-discover; the repository’s other historical search methods retain their own protocols. None of these results demonstrates model-scale transfer or a universal architecture.