Minimal substrate
Start deliberately under-structured. Run the fast loop only and log loss, uncertainty, utilization, and synthetic R.
Build roadmap
Roadmap for creating Teleodynamic AI
A staged path from deliberately under-structured learning to auditable systems that can grow, halt, prune, and restabilize under their own resource economy.
Roadmap principle.
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.
Definition of done
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.
Six-phase build sequence.
Resource economy
Add gain, decay, viability floor, per-action costs, maintenance cost, and blocked-action telemetry.
Slow loop
Add split and no-op. Generate candidates from persistent confusion pairs, high entropy, or uncertainty spikes.
Operator expansion
Add merge, retire, and one domain-specific operator such as a glyph variation distinction or tool-use module.
Phase detection
Track error-complexity trajectories, structural action rate, R utilization, and no-op dominance.
Auditability
Export slow-loop traces with triggers, candidates, R before/after, alternatives, cost, expected gain, and justification.
Milestones and artifacts.
| 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. |
Domain track: Teleodynamic AI Communication.
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.
Input pressurenew glyphs, ambiguous sequences, unknown relations
Candidate editsplit class, add relation, retire weak mapping, no-op
Resource gateR.energy, R.complexity, R.compute, R.uncertainty
Evidence resultloss delta, comprehension change, trace completeness
What to plot instead of a static leaderboard.
Stability plot
Structural actions per 1,000 inputs should rise during discovery, then plateau as no-op becomes dominant.
Pareto front
Plot accuracy against complexity and energy consumed. A useful model sits on the frontier, not at maximal size.
Viability retention
Measure how often R stays above the viability floor during distribution shift, novelty, or adversarial stress.
Audit test
A reviewer should reconstruct why a distinction was added from the trace alone.
Reactivation test
Inject novelty. The system should re-enter growth, pay cost, restabilize, and return to no-op dominance.
Comprehension test
For communication systems, measure whether target users understand new glyph distinctions above threshold.
Risks to design against.
- Externalized resource state: if R is only a scheduler or human-authored quota, the learner is not maintaining itself.
- Growth-only operators: without merge and retire, the system becomes complexity accumulation with a new vocabulary.
- Global structural loss: using one global objective erases the local cost-benefit property that makes edits interpretable.
- Black-box structural policy: hiding operator choices destroys the interpretability-by-construction claim.
- Overclaimed communication: glyph interpretation remains approximate unless public evidence, source routing, and human comprehension support the claim.
Roadmap outcome.
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.