PCDVR research · 01 / 03

Same letters. Different paths.

We found a sequence effect. Then we tried to understand what it preserves.

An illustration, not a statistical test

Same nodes. Different journeys.

LISTEN · SILENT · ENLIST · TINSEL · INLETS

These five words contain exactly the same letters and the same PCDVR state inventory. Their occupancy representations are identical; their ordered trajectories are not.

Experiment 21I: five colored anagram trajectories connect the same six letter positions in a three-dimensional coordinate plot.
Original Experiment 21I illustration. The axes describe the experimental representation, not physical speech or personality.

LISTEN / ENLIST: identical state occupancy, 80% directed-edge overlap, but full trajectory distance 0.667.

0LISTEN / ENLIST occupancy distance
80.0%directed adjacency overlap
0.667full trajectory distance

These are representation-specific measurements, not compatibility scores. Different paths can visit exactly the same states.

Compare LISTEN + ENLIST →

Experiment 21P · 96 reference / 48 challenge words

What happens when order disappears?

Letters are deterministically converted into PCDVR states. This test preserves those states and their frequencies while changing only their order.

View the sequence test

Sequence-destruction control

Sequence carries measurable information

On an independent 48-word challenge across four word-form families, the complete ordered PCDVR trajectory classified 33 of 48 words correctly (68.8%).

State occupancy31.2%31.2%
Directed transitions50.0%50.0%
Complete ordered trajectory68.8%68.8%
Accuracy after destroying within-word sequence 2,000 randomized-order controls
Distribution of 2,000 sequence-destruction controls The shuffled distribution centers at 43.9 percent accuracy. One shuffled control matched the ordered accuracy of 68.8 percent. 25% chance43.9% mean68.8% ordered

Each shuffled control preserved the same PCDVR states, state counts, word lengths, and family labels; only within-word order was changed.

2,000shuffled controls
43.9%mean shuffled accuracy
1 / 2,000matched or exceeded ordered
p = 0.0010plus-one empirical estimate

This result is conditional on the current curated reference and challenge vocabularies. It supports a sequence-order effect in this experiment; it does not establish universal language-origin prediction or superiority over ordinary spelling.

Experiment 21P · 2,000/2,000 replicates retained · Result fingerprint 13b5f4470b92…ccbb

Experiment 21G · 93 reference / 47 challenge words

How much of the path do we need?

Accuracy increased as the tested local windows grew from one state to four. None of those tested representations reproduced the full-path result.

1-state local windows29.8%29.8%
2-state local windows40.4%40.4%
3-state local windows44.7%44.7%
4-state local windows51.1%51.1%
Full ordered trajectory70.2%70.2%

The experiment compared increasingly broad local sequence context with the complete ordered trajectory.

Four-family equal-choice chance: 25%. The complete trajectory classified 33 of 47 words correctly.

Why 70.2% here, but 68.8% above?

21G uses a 47-word challenge subset; 21P uses all 48 challenge words. This is a different evaluation set, not evidence that the model improved.

The one-state local-window method is also not the same measure as 21P’s occupancy baseline. The upward sequence of accuracies does not establish that every adjacent increase is statistically significant.

This experiment isolates local versus complete sequence context inside PCDVR. It does not test whether PCDVR outperforms raw spelling.

Next: testing word boundaries →

Help chart what comes next

Questions, thoughtful criticism, and support keep this independent research moving.

Support the project → Contact Bryan →