decision
ADR 0020: `autonomous-research-loops` survives its merge-back trigger, and the ADR-0018 boundary holds
About this note
| Field | Value |
|---|---|
| Status | accepted |
| Date | 260805 |
| Deciders | chamin |
On this page
ContextDecisionAlternatives consideredConsequencesContext#
Two prior ADRs left explicit conditions attached to this pair of notes, and S22 (Darwin Gödel Machine) is the source that tests both at once.
ADR-0017 created
autonomous-research-loops.md on a single source, S13
(karpathy/autoresearch), and recorded it as a merge-back candidate if no second primary arrives.
The note has carried that warning in its Status line since - one T4 source; merge-back candidate.
ADR-0018 created
self-improvement.md beside it and drew the boundary between them as
a matter of altitude: autonomous-research-loops changes an artifact, self-improvement
changes the model. It recorded a re-test trigger: "revisit on the next source that lands claims in
both."
S22 lands claims in both, and it is unambiguously a second primary. Published at ICLR 2026 (the strongest venue in this brain), open-sourced code, and two ablations that isolate its claimed components. It is a system that iteratively rewrites its own codebase, validates each change against coding benchmarks, and keeps an archive of every agent it has produced - reaching SWE-bench 20.0% to 50.0% over 80 iterations.
And it sits cleanly on the artifact side of ADR-0018's line, by its own admission. The framing in
the abstract suggests a system that could "rewrite their own training scripts (including training a
new foundation model)"; §3 concedes "we do not show that in this paper" and that the work "focuses on
improving the design of coding agents with frozen pretrained FMs" (S22 d3). The model never
changes. The scaffolding around it does.
Decision#
Three things, taken together.
-
autonomous-research-loops.mdis no longer a merge-back candidate. ADR-0017's condition is met: a second primary has arrived, it is independent of S13 in every respect (UBC / Vector / Sakana against a personal repository), and it is far better evidenced. The Status line's merge-back warning is removed and replaced with a record of what resolved it. -
S22's claims are homed in
autonomous-research-loops.md, and cross-referenced fromself-improvement.mdrather than duplicated. This follows ADR-0018's altitude rule applied to a source that states its own altitude. -
ADR-0018's boundary is confirmed rather than merely reasserted. The re-test asked for by ADR-0018 has now been run against the hardest available case, and the line held without needing adjustment. The next re-test trigger is narrower and is recorded here: a source in which the same loop modifies both the artifact and the model weights. S22 explicitly declines to be that source and names it as future work, so it is a real possibility rather than a hypothetical.
This is not a one-way door. Both notes remain mergeable if a later source shows the distinction is not carrying weight.
Alternatives considered#
- Merge the two notes now that both have two sources. Rejected. The altitude distinction is doing
more work after S22, not less:
self-improvementholds coverage/precision, test-time scaling and the generation-verification gap, all of which are properties of sampling from a model, whileautonomous-research-loopsholds freezes, accept rules, archives and rollback, which are properties of an unattended optimisation over an artifact. A merged note would have to explain why an exponentiated power law and a viability gate belong together. - Home S22 in
self-improvement.mdbecause its title says "self-improving agents". Rejected under ADR-0012's test - a source feeds a topic when it teaches within the scope, not when it is named after it. S22 teaches search design over a frozen model. - Home it in both, duplicating the claims. Rejected as the stacking the contract forbids. Cross- referencing costs a reader one click and keeps one home per claim.
- Create a third note for "self-modifying systems". Rejected under the don't-spawn-a-topic-per- source rule. S13 and S22 are the same subject at different levels of rigour, which is what a topic note is for.
Consequences#
Easier. autonomous-research-loops.md can now be cited as a real topic rather than as a provisional
one, and its central claims acquire a second, far stronger source. Specifically, S13's four freezes -
recorded as one practitioner's design - are now readable against a peer-reviewed system that froze two
things for stated reasons, and claim 114 (an accept rule with no variance notion banks noise) gains
a counterpart in S22's staged evaluation with an explicit noise-derived threshold.
Harder. The two notes now have to be read together to get the whole picture on any self-improving system, and a reader arriving at either one alone will get half the frame. Both notes carry a pointer; that is the mitigation and it is not free.
To revisit. The narrower trigger recorded above - a single loop modifying both artifact and weights.
Also worth watching: if autonomous-research-loops acquires a third and fourth source while
self-improvement stays at one independent source, the asymmetry may argue for a different split than
altitude.
Follow-up edits made in the same pass: the Status line and synthesis in
autonomous-research-loops.md, a cross-reference in
self-improvement.md, the Topics rows in the root
INDEX.md, and a dated line in log.md.