Ambiguity is not always failure. That was the main lesson from the latest round of work.
The recent focus was Layer 1 and early Layer 2 behavior over text. Layer 1 now has a better substrate for text object representations.
One concrete mechanism we added is a text-form SDR generator. Previously, learned text objects could receive arbitrary sparse codes. That was deterministic, but it meant related text forms did not necessarily have related object representations. `there`, `their`, `the`, and `theme` could be no closer to each other than `there` and `car`.
The new encoder builds text-derived object SDRs from structure in the observed text form:
shared characters
character positions
n-grams
transitions
repeated letters
punctuation shape
numeric/date-like shape
So related forms overlap for sensory/form reasons:
the / there / their
car / cat
5% / 10%
2026-06-12 / 2026-06-13
This is not a semantic embedding. `there` and `their` overlap because their text shape overlaps, not because the system understands them yet.
That distinction matters. The goal is not to smuggle language understanding into the encoder. The goal is to give the text sensor a more reasonable visual/form substrate, then let the column machinery learn objects, paths, predictions, and associations from experience. Meaning will come layer as we build the higher layers.
Also important: each cortical-column-like module still uses its own private deterministic projection. So columns do not get shared labels or shared object codes for free. Cross-column agreement still has to be learned through lateral association.
The immediate result was strong Layer 1 recognition. Layer 2 also improved. It now clearly separates familiar learned lower-layer paths from unrelated input. Familiar input produced much lower surprise and much stronger successor support than unrelated input.
Finally getting closer to the kind of mechanism we want: not a text tokenizer, not a transformer clone, and not a lookup table. A layered system that builds object maps from evidence, keeps plausible interpretations alive, predicts what should come next, and lets context resolve uncertainty when there is enough evidence to do so.