Minimal substrate
Start deliberately under-structured. Run the fast loop only and log loss, uncertainty, utilization, and synthetic R.
Build roadmap
A staged path from deliberately under-structured learning to auditable systems that can grow, halt, prune, and restabilize under their own resource economy.
Do not begin with the biggest model. Begin with a minimal substrate whose failures are easy to see, then add viability state, slow-loop structure, and audit traces one layer at a time.
A prototype is not teleodynamic until no-op, split, merge, add, and retire decisions are resource-gated, locally justified, logged, and visible in phase plots.
Start deliberately under-structured. Run the fast loop only and log loss, uncertainty, utilization, and synthetic R.
Add gain, decay, viability floor, per-action costs, maintenance cost, and blocked-action telemetry.
Add split and no-op. Generate candidates from persistent confusion pairs, high entropy, or uncertainty spikes.
Add merge, retire, and one domain-specific operator such as a glyph variation distinction or tool-use module.
Track error-complexity trajectories, structural action rate, R utilization, and no-op dominance.
Export slow-loop traces with triggers, candidates, R before/after, alternatives, cost, expected gain, and justification.
| Milestone | Build artifact | Passing evidence | Do not claim yet |
|---|---|---|---|
| Under-structured baseline | Small learner, fixed structure, telemetry log. | Clear confusion clusters and uncertainty spikes. | Teleodynamic structure. |
| Internal R(t) | Mutable resource state with gain, decay, action cost, and maintenance burden. | Expensive actions are blocked when R is low and re-enabled after success. | Intrinsic purpose. |
| First structural edits | Split/no-op candidates selected by local objective. | No-op wins when no affordable split improves local cost. | Open-ended intelligence. |
| Reversible library | Split, merge, add, retire, no-op, and one domain operator. | Complexity plateaus after utility drops and novelty can reopen growth. | Production safety. |
| Constraint closure graph | Dependency registry of structures and maintenance cycles. | Prune candidates are explainable from dependency and utility traces. | Biological equivalence. |
| Audit package | Trace export, phase plots, local objective records, and evidence bundle. | A third party can reconstruct why a distinction was added or retired. | Certification or conformance. |
The most concrete near-term domain is semantic glyph interpretation. It naturally exposes representation growth, resource cost, public-symbol boundaries, and human comprehension testing.
A first domain-specific operator can add a variation-sensitive glyph distinction only when the distinction improves interpretation enough to repay complexity, compute, and review cost.
Structural actions per 1,000 inputs should rise during discovery, then plateau as no-op becomes dominant.
Plot accuracy against complexity and energy consumed. A useful model sits on the frontier, not at maximal size.
Measure how often R stays above the viability floor during distribution shift, novelty, or adversarial stress.
A reviewer should reconstruct why a distinction was added from the trace alone.
Inject novelty. The system should re-enter growth, pay cost, restabilize, and return to no-op dominance.
For communication systems, measure whether target users understand new glyph distinctions above threshold.
The target is a research prototype that can explain its own structural history: what it added, what it refused, what it could not afford, what it retired, and why the current organization remains viable enough to use.