An open, continuously-published log of small, pre-registered continual-learning and credit-assignment experiments. Negative results included by design.
By Macheng Shen (independent researcher), working with a self-built AI system.
A companion log to the broader open research program (the credit-transport /
General-Learning-Machine thread) — see the sibling repositories
living-information-system
and reversible-layer-aging.
A working lab notebook, published as it goes. Each entry is a small experiment on how a learner decides how to learn — meta-control, credit assignment, replay, structural allocation — run at smoke scale (single GPU / CPU, a handful of seeds), pre-registered, and reported honestly whether the answer is yes, no, or "the question was malformed."
The premise: most of the value in this kind of work is in the intermediate results and the diagnoses — the under-trained null that looks like a finding, the bug that silently invalidated three weeks of comparisons, the lever that turned out to have no leverage. Those normally die in a private notebook. Here they are the product. If a result is useful to you, take it.
- Negative + diagnostic results are first-class. A pre-registered "no" that rules out a lever is a real contribution; we publish it instead of burying it.
- Reproducibility over polish. Each entry states the exact pre-registered question, the falsification criteria committed before running, and the decision rule. No moving the goalposts after seeing the data.
- Continuous, event-driven. New intermediate results get appended as they
land (see
CADENCE.md) — not batched for a paper, not on a fixed clock.
Every entry commits, before running, to:
- A single primary metric (named, not chosen post-hoc from several).
- Falsification criteria — what result would kill the hypothesis.
- Confound-controlling ablations that any positive result must survive (e.g. compute-matched control, random-signal control, frozen-signal control).
- A stop rule — we do not add seeds to chase significance, do not prune seeds, and do not tune the setup until a marginal effect crosses a pre-registered line.
These are not aspirational: the first entry below exists because this discipline caught two artifacts a mechanical pipeline would have shipped as real findings.
| Date | Title | Verdict |
|---|---|---|
| 2026-05-31 | Does vector-valued meta-control help continual learning? | Negative + diagnosis |
| 2026-06-18 | Can a model learn its own "intrinsic time"? | Negative + surviving isomorphism |
| 2026-05-31 | Do brain-plausible / richer-than-scalar credit mechanisms beat backprop? | Negative + reframe |
| (pre-reg) | Capacity vs. credit (2×2) — pre-registration | Pre-registered design |
These are smoke-scale studies: a few seeds, modest compute, designed so the shape of the result is visible without a cluster. Each entry's directory carries its own pre-registration, decision rule, and the numbers behind the conclusion. Where a benchmark or runner is releasable it will live alongside the entry; until then, each write-up is self-contained enough to re-derive the result from the stated design.
This is a synthetic-benchmark research log. It contains only the scientific content: designs, metrics, pre-registrations, and findings on engineered / benchmark data. It deliberately contains no private operational detail, no infrastructure internals, and no third-party personal information. Every entry passes a mechanical content review before it is added.
Code and benchmark artifacts are released under the MIT License. The written research log (prose, write-ups, and figures) is licensed under CC-BY-4.0.