PCDVR research · 02 / 03
Testing word boundaries.
Comparing complete trajectories with word-boundary baselines in the current dataset.
From Experiment 21H to 21I
Testing word boundaries.
In the two-block condition, mean accuracy was 45.9%, with 78.4% of directed adjacency retained on average. Rearranging blocks also changes boundary positions. Experiment 21I compared word-boundary representations and controls to examine this factor.
The block-rearrangement test · 21H
1,000 controls per condition.
Randomly cut every >=4-state word into 2, 3, or 4 contiguous blocks, preserve exact order within each block, forbid the original block ordering, and independently permute block order in both reference and challenge sets. Compare with an individual-state shuffle.
93 reference words, 47 challenge words, four families. The two-block shuffle’s central 95% accuracy interval was 34.0%–57.4%; this is a distribution of controls, not a confidence interval for generalization.
Experiment 21I · 93 reference / 47 challenge words
Comparing final characters, final states, and full paths.
On this challenge, tie-aware accuracy was 66.0% for the raw final character, 66.1% for the PCDVR final state, and 70.2% for the full trajectory. These measurements do not establish an advantage beyond ordinary spelling or isolate the cause of the difference.
These bars use tie-aware fractional accuracy: when multiple families tie for best, credit is shared among the tied families if the true family is included. Four-family equal-choice chance is 25%.
Why the scoring detail matters
The raw-final-character and PCDVR-final-state diagnostics each had 19 tied queries out of 47. Deterministic tie-breaking scored both at 72.3%; tie-aware scores were 66.0% and 66.1%. The full trajectory had no ties and scored 33/47 (70.2%) either way.
Prefix- and suffix-preserving controls
Means across 500 controls per condition. The last-two-state condition had higher mean accuracy than the first-two-state condition in this experiment. Whether boundary placement accounts for the earlier performance change remains a question for further controlled tests.
Boundary effects must be compared with raw first/last-character baselines. A strong suffix/end effect can be ordinary orthography rather than a special PCDVR property. These results do not quantify how much independent signal remains after all ordinary spelling effects are controlled.
Experiment 22 · Now reported
Testing the remaining signal.
The orthographic residual challenge did not detect residual word-family signal under its declared spelling controls in the fixed dataset. The classification phase is now in intermission.
Experiment 22 and the Star Branch →Help chart what comes next
Questions, thoughtful criticism, and support keep this independent research moving.
Support the project → Contact Bryan →