Three architectures are plausible. A fixed Small Language Model can act as a predictive component beneath a teleodynamic controller. A system can adapt routing, memory, rules, retrieval, or bounded adapters around a fixed base. The strongest proposal would change both language-model parameters and structure under endogenous resource coupling.
Start with the smallest falsifiable claim
For an initial GGUF experiment, keep model weights immutable. Let an interpretable controller choose among explicit structural actions such as adding or retiring a route, memory item, rule, or task adapter. Score noop beside every alternative and record the complete candidate set.
Proposed state mapping
| Teleodynamic field | Candidate implementation | Evidence needed |
|---|---|---|
| H | Routes, rules, retrieval entries, memories, or adapters | Versioned lineage and an actual change in the available hypothesis class |
| θ | Scores, thresholds, calibration values, or bounded adapter parameters | Revision trace with structure held fixed during inner updates |
| E | Normalized deterministic resource ledger | Published update rule, units, costs, bounds, and effect on action viability |
| τ | Append-only decisions and observations | Replayable event order, seeds, candidate scores, state revisions, and external guards |
Controller boundary
The predictive model proposes scores or representations. The controller owns candidate generation, objective calculation, resource feasibility, selection, persistence, and phase diagnostics. Blurring those responsibilities makes the causal claim difficult to audit.
Ordinary runtime controls remain external
Context truncation, KV-cache eviction, quantization, worker cancellation, device fallback, and memory ceilings remain visible safety or performance controls. They do not become emergent dynamics merely because the controller observes them.
Preserve terminology
MiRust should state which component is teleodynamic, which component is a language model, which state is persistent, and which behaviors are ordinary runtime management. A failed or inconclusive experiment is a valid result.