BASIN MV RESEARCH JOURNAL
← Research record

OTB20 / EVIDENCE SNAPSHOT 2026-11-05

OTB20 · Protection can block learning

Exploratory numerical study

Is an input bound that protects memory sufficient for useful learning?

What was found

The bounded-error conditions protected the non-target memories in 32/32 cases but updated the target in 0/32. Ordinary error with a 4,096-step wait achieved both in 32/32.

32/32Bounded-error protection
0/32Bounded-error target update

Task, conditions and denominator

Fixed T7, companion and e/s rules. Only the external teaching-error amplitude is controlled. The exploratory grid contains 576 trials.

Controls and failures

Normal error at waits 0 or 2,048 updated the target in 32/32 but preserved non-target bands in 16/32. The bounded conditions protected bands at all three tested waits.

External dependencies and limits

Protecting an unchanged state does not prove the ability to write a new one. Both a protection upper bound and a finite-time writing requirement matter.

Read the primary report

The source snapshot is in Japanese; this page is a bilingual editorial summary. A report snapshot is not an independently reproduced result. Original relative links may require files from the source repository.

Open the preserved report (Japanese) ↗
Source identity and verification scope

Source commit and path:

https://github.com/type37c/oscillator-arithmetic/blob/e3a5e1e66eae794a7fd56bae5f4d5054a74528c2/results/otb20_bounded_error_amplification/REPORT.md

The original development repository may require access. The report snapshot above is public. Only OTB21 has a full runnable public package; other entries are report-level evidence.

SHA-256: 537a2a677a7733ab2184d559dc03dcc389e6cb511c8064af15f9e2d95af2d831

Source manifest ↗

Concepts to learn first

WHAT THIS RESULT MAKES US ASK

Does the successful short fixed wait survive changed phase and query timing?

Discuss a test or counterexample ↗

Textbook revision 1 · Evidence snapshot: 05 Nov 2026 · Changes & corrections ↗