This supplement supports The AI Meeting Nobody Needed. Experiments completed October 11, 2026; article published October 13, 2026. These are existing experimental results, not new runs for publication.
Qwen2.5-0.5B and Falcon-H1-0.5B-Base ran locally with frozen model weights. Qwen's relevant hidden-state width is 896; Falcon's is 1024. A controller trained learned requests, readers, a shared transformation and native-input writers. Each specialist had a different private input: one source record with a color, and one receiver table mapping four colors to cities. The controller received neither the private receiver table nor the target city as an input. Labels entered supervised training losses only.
The English route executed the source model and supplied a learned text handoff to the receiver. The latent route carried selected actual source hidden states through learned vector interfaces. Both directions were tested, with one unrestricted native-vocabulary token as the final city answer. The route was fixed; these results do not demonstrate a learned executive policy. Original records, tables and instructions still contained text.
A vague source request had produced too many unparseable replies. Another 28,000 updates per controller did not improve selected development results. A clearer color-only request and fresh controller fits preceded the reserved original-city comparison. English scored 232/256 (90.625%); original latent scored 224/256 (87.5%). All eight additional English wins belonged to one table mapping. The subsequent transfer cohort is different and must not be pooled with that test.
The latent fine-tune started from the selected original latent controller, with both specialists still frozen. Training added final-city negative log probability through the sender and receiver, plus 0.1 normalized query-embedding and 0.1 fact-embedding imitation anchors. AdamW was reset, with learning rate 0.0001, batch 4 in both directions, and seed 2026101117. Selection used existing exposed development data every 25 updates, ranked by minimum-direction exact accuracy, then total accuracy and earliest tie. The stopping gate was at least 95% in each direction. Step 50 reached 256/256 development accuracy and was sealed before fresh evaluation worlds were constructed. No transfer-test feedback entered selection or further training.
This is a latent engineering follow-up: English received no comparable extra training. The latent controller had 927,232 trainable parameters, compared with approximately 464,000 for English. Additional updates, objective, anchors, optimizer reset, learning rate and capacity prevent attribution to a communication modality alone. Matched optimization and comparable English outcome supervision remain future controls.
The transfer cohort contained 192 worlds: two predefined city palettes, all 24 table permutations in each palette, and all four colors for each table. Both directions produced 384 attempts per controller. City labels and table mappings were held out from this paired-controller training lineage; the wording and four-color task were unchanged. City names were selected strictly for both tokenizers' single-token compatibility, not observed accuracy. The first incompatible palette proposal failed before training or test predictions.
| System | Correct / attempts | Accuracy | Mean live latency | Logical operations / attempt |
|---|---|---|---|---|
| Original latent | 292/384 | 76.04% | 86.70ms | 62.545billion |
| Improved latent | 384/384 | 100% | 86.74ms | 62.545billion |
| Fixed English | 282/384 | 73.44% | 264.03ms | 63.803billion |
Improved latent fixed 92 errors with no regressions against the original latent controller. Each model direction scored 192/192. Original latent scored 170/192 Qwen-to-Falcon and 122/192 Falcon-to-Qwen; English scored 155/192 and 127/192 respectively. English extracted and handed off the correct color on 384/384 attempts, but made 102 final receiver errors. Those errors do not demonstrate loss of the color in translation.
Palettes were Boston/Chicago/Seattle/Toronto and Oxford/Cambridge/York/Lincoln. Improved latent scored 192/192 on each. Four labels (Boston, Chicago, Oxford and York) appeared in older separate transport pilots, without reusing their weights in the current controllers. The labels are not guaranteed novel to the project or pretrained specialists.
Logical arithmetic counts use the project's dense operator census, with two operations per multiply-accumulate and separately recorded nonlinear evaluations/comparisons. They are not hardware instruction counts. All 1,152 live attempts had separate fully instrumented repeats with matching predictions and usage. Models executed FP32 with TF32 off on an RTX 5090. The serial, uncached timing ratio is 3.04 times; latent uses only 1.97% less arithmetic per attempt. Logical arithmetic per correct answer is 62.545 billion for improved latent, 82.251 billion for original latent and 86.881 billion for English on this cohort. Complete training backward/optimizer arithmetic was not measured.
On 32 balanced exposed development controls, full latent scored 32/32, zeroed source returns 6/32, and changing the actual source color produced the corresponding changed answer 32/32. Gradients reached the learned requests, source states and handoff components, while specialist parameters remained frozen. Source, comparator and specialist fingerprints, saved predictions and profiles were independently checked. Controls were exposed-development diagnostics, not an independent transfer confirmation.
One fit seed, two selected palettes, fixed wording and a two-bit fact cannot establish general task transfer, arbitrary model interoperability, richer information preservation or universal efficiency. A table-clustered bootstrap describes variation within the selected palettes, not uncertainty over fit seeds or new tasks. Further tuning needs a new untouched test because the current cohort is now exposed. English reply-length/content controls remain incomplete.
Separate learned-executive experiments failed goal-wording/generalization gates. Initial faithful development performance was 75/192; rewriting those exposed goals into familiar wording raised it to 190/192. The latter is a diagnostic, not a new held-out success, and neither score is part of this fixed-route handoff result.
This public supplement contains aggregate evidence and scope. Full model checkpoints, optimizer state and private operational run archives are retained separately. Related work includes Interlat and Cache-to-Cache. Model cards: Qwen, Falcon.