An operator that deliberates across several cheap passes, stays silent while unsure, and decides for itself when to answer. Adaptive compute worked but is essentially PonderNet. The orthogonal-synthesis geometry is the one result both new and better, extrapolating well outside its training range.

Normally a model reads something and immediately produces one answer. This one can hold two opposing ideas at once, think quietly without emitting anything, and speak only once it has worked out a third answer that combines them. The harder the question, the longer it sits with it.

The repository is a research log. Each folder under src/ is one iteration, and each fixes a specific limitation of the one before it.

What held up, and what did not

The README states this plainly, which is the point of keeping it as a log:

• Useful and measured: spending more compute only when the problem is harder. Across four benchmarks it matched a full model at much lower cost, with thinking time tracking difficulty.
• Useful but not new: that result is essentially PonderNet and Adaptive Computation Time. Done carefully and validated, but not invented here.
• Genuinely new and better: the orthogonal-synthesis geometry, which builds the answer as a third thing perpendicular to the disagreement. It extrapolates to situations three to four times outside its training range, where ordinary models fall apart.
• Tested and found to do nothing: the project’s own named signal, settling speed. Kept in the log rather than quietly dropped.

What was interesting

Keeping the negative results in the record, and separating “this works” from “this is ours” from “this is new”. Three different claims that are easy to let blur together.