Representation
Do learned symbols retain useful information? Continuous and untrained controls test whether learning and discretization help.
Future Language · v2 archive
Can a learned language travel beyond its training?
Compare symbols, continuous representations, and sequences of possible changes. Forecasts cover three, six and twelve hours. Seattle and Phoenix are reserved for testing transfer: their outcomes never train these models.
Archived · 2026-10-10
Administrative retirement: this exploratory approach did not demonstrate a predictive advantage over its continuous-input baseline.
Collection and recurring research reviews have stopped. Data, frozen forecasts, reference models and verification tools are preserved.
Calls still marked pending are unresolved at administrative closure. Collection has stopped; no future settlement is promised. Historical statuses, forecasts and scores remain unchanged.
This was an exploratory allocation decision, not a registered statistical verdict. The broader OrangeKraken radio and weather services continue separately.
Archived forecast record
32,538 observations · 1832 resolved forecasts · 60 pending · 7 void
Final retained snapshot 2026-10-10T00:51:40.528412Z. Counts across horizons are correlated and are not independent sample sizes.
Download public forecasts and outcomesTransfer
Do patterns learned in Great Falls, Boulder, Houston and Miami predict outcomes in Seattle and Phoenix?
Composition
Can successive three-hour state transitions produce useful six- and twelve-hour forecasts? Direct forecasts provide the comparison.
Frozen forecasts, observed outcomes
Multiclass Brier score; lower is better. These exploratory results are descriptive. The primary exploratory comparison is learned symbols versus continuous inputs, at six hours on the held-out stations.
| Model | Training sites · 3h 409 scored / 18 days | Training sites · 6h 408 scored / 18 days | Training sites · 12h 401 scored / 18 days | Held out · 3h 206 scored / 18 days | Held out · 6h 206 scored / 18 days | Held out · 12h 202 scored / 18 days |
|---|---|---|---|---|---|---|
| Continuous inputs | 0.3858 | 0.3582 | 0.3091 | 0.4055 | 0.4589 | 0.3697 |
| Learned continuous | 0.4003 | 0.3728 | 0.3015 | 0.4365 | 0.5113 | 0.3712 |
| Untrained continuous | 0.4067 | 0.3791 | 0.3108 | 0.4378 | 0.5193 | 0.4245 |
| Learned symbols | 0.4837 | 0.5040 | 0.3701 | 0.4789 | 0.6045 | 0.4524 |
| Untrained symbols | 0.4814 | 0.4504 | 0.4289 | 0.5182 | 0.6201 | 0.5845 |
| Combined representation | 0.3840 | 0.3592 | 0.3079 | 0.4101 | 0.4662 | 0.3704 |
| Composed transitions | 0.4849 | 0.5153 | 0.4615 | 0.4911 | 0.6008 | 0.5014 |
| Pooled rates | 0.5837 | 0.6657 | 0.6683 | 0.4877 | 0.6664 | 0.6774 |
| Persistence | 0.6624 | 0.9419 | 1.0620 | 0.4678 | 0.9435 | 1.0442 |
Training datasets, learned representations, model snapshots and detailed protocols are retained privately. The public ledger supports score checking, not full model reproduction or independent proof of issue time. This is a temperature experiment, not a general-purpose language model or a proven forecasting service. Unseen locations test geographic transfer, not transfer into radio or materials. No outcome automatically promotes a model.