RESEARCH RECORD / E4 → OTB22
A history of questions and evidence.
Follow the selected research progression. These are different systems and tasks, not a single benchmark leaderboard. Failed conditions and later diagnostics stay visible.
E4
Confirmation · fixed scopeCan equations predict which available preparation will work?
An analytic selector reached 95.02% joint-ready performance versus 95.16% for the 24-candidate oracle. It met the declared non-inferiority criterion in this fixed task.
E5
Confirmation · fixed scopeCan preparation sustain computation when the system is not reset?
SYNC_WARM achieved 79.18% seven-answer success, versus 54.70% for observation-based selection and 55.21% for a shadow that assumed a reset. Tracking the inherited state mattered.
E7D
Confirmation · fixed scopeCan an oscillator readout support closed-loop preparation selection?
The 64-unit readout reached 57.8% seven-answer success versus 48.5% for frozen E7, a gain of 9.3 points. New input cohorts reproduced the exploratory improvement on the same five circuits.
OTB8
Exploratory numerical studyCan one-step experiences support unlabelled multi-step paths while preserving unknowns?
With partial experience, the original condition answered 830/1,024 queries completely correctly. After diagnosis, preparation input plus reading only after stimulation reached 1,024/1,024. The earlier failures remain part of the result.
OTB12
Exploratory numerical studyDoes silence at a fixed deadline mean the relation is absent?
At 15 edges, the 8,000-step period gave 12/24 fully correct queries, including false negative relations. A later 16,000-step diagnostic recovered 24/24 on the same cases.
OTB15
Exploratory numerical studyCan recent activity and a teaching error update a retained relation?
The designed local rule learned the tested binary mappings, retained them for 80,000 steps, and adapted the changed operation. Normal learning and reversal each gave 64/64 correct test responses.
OTB16
Exploratory numerical studyWhat grows besides the number of retained relations?
With 12 features and eight operations, the main condition answered 3,072/3,072 held-out profile queries correctly. Holding the total tag bandwidth to 0.04 produced 441 abstentions, while stored memory bands remained intact.
OTB17
Exploratory numerical studyCan a learned relation be used from different working states?
The pilot-selected 128-step restoration achieved 2,124/2,304 supported successes at the shortest total budget, leaving 180 abstentions. At longer budgets it reached 2,304/2,304.
OTB18
Exploratory numerical studyDoes improving readiness automatically improve learning?
Restoration improved the first stable-world query from 80/128 to 128/128 correct, yet all three normal policies updated the experienced reversed context in 64/64 cases after one experience.
OTB19
Exploratory numerical studyCan preparation prevent an unrelated memory from being overwritten?
The usual update learned the new context in 32/32 cases but retained the unchanged relation in 16/32. Guarded waiting retained 32/32 while learning 32/32. A conservative fixed wait also succeeded.
OTB20
Exploratory numerical studyIs an input bound that protects memory sufficient for useful learning?
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.
OTB21
Public reproducible studyDoes a retention condition transfer to changed phases and query durations?
The guard learned the new relation and retained the unchanged one in 128/128 conditions in every duration group. At the shortest duration, immediate and fixed waiting retained only 64/128.
OTB22
Exploratory numerical studyCan separately held experiences be used together while connected?
Three modules separately retain A→B, B→C and C→D. Shared firing produced a D support response in the recipient in 16/16 conditions; the same states without sharing produced 0/16.
OPEN FRONTIER / NOT ESTABLISHED
The questions beyond this evidence.
- Autonomous concept acquisition from observations without supplied object identities.
- Conflict detection and justified abstention across owners.
- Learning another owner’s experience into the recipient’s own memory.
- Matched-time controls, optimized digital comparisons, and measured hardware efficiency.
Turn a question into a test ↗Textbook revision 1 · Evidence snapshot: 05 Nov 2026 · Changes & corrections ↗