Work

Grading a Self-Learning Research Assistant: An Interim Report on CubeMind's First Night

Year: 2026
Type: preprint
Abstract: Status update, 20 November 2026. This version adds a one-page addendum (24346035_addendum.pdf) after an independent referee check: verified errata, what has changed since publication, and links to the newest records. The original files are unchanged. Newest records: SERAPIS v3.4 status update (doi:10.5281/zenodo.24396665); The Quark Thread (doi:10.5281/zenodo.24399408); Weave Control and the Evidence Atlas (doi:10.5281/zenodo.24399447). v1.1 (18 Nov 2026, refereed revision): Adds the whole-night result (0 of 16 proposals promoted), a limitations section (Clopper–Pearson bounds, single AI grader, no baseline) and related work; "self-learning" defined as memory growth; wording fixes. License (from this version): CC BY-NC-SA 4.0 under the AEGIS Community License — free for individuals, scientists and students; any business or for-profit use needs the AEGIS Commercial License (https://github.com/dannyeldridge54/AEGIS/blob/master/licensing/COMMERCIAL_LICENSE.md). Built by one independent researcher with the help of AI. Interim report (one night). CubeMind is a small self-learning research assistant on a home PC: a local open-weight language model proposes falsifiable hypotheses about a string-cosmology model (SERAPIS-E), stores them in an append-only memory and builds a web of linked results. A second AI (Claude) grades each proposal with real computation or literature. Results, 17–18 Nov 2026: 10 proposals graded, 0 promoted — 4 tautologies, 3 falsified by exact calculation, 2 category errors, 1 ill-posed. Three engineering fixes changed proposal quality: anti-tautology rules; repairing memory ingest (a header-format mismatch and a stale copy missing 51 sections); and making correction notes supersede what they retract. Main failure mode found: confident category errors. Lessons: memory quality bounds proposal quality; retractions must be first-class records; the deploy path is part of the experiment (a stale-cache transfer bug, fixed by content-hashed names and md5 checks); a grader stays in the loop until proposals survive. Status: not yet self-sufficient. Full grading log: docs/web/CHECK_INDEX.md in github.com/dannyeldridge54/aegis. Prepared with an AI assistant (Claude, Anthropic).
Language: English
Open Access status: green
OpenAlex ID: W7235914515
Cites: 2
Cited by: 1
Related to: 10