Goal
A teleodynamic harness tests whether structural changes pay for themselves. It is not just an accuracy benchmark. It simulates evidence intake, candidate generation, scoring, budget updates, and rollback.
Structural actions
| Action | When considered | Main risk |
|---|---|---|
| split | one specialist fails on separable subdomains | fragmentation |
| add | a missing capability is proven and affordable | bloat |
| merge | two specialists overlap strongly | interference |
| retire | a model is underused or unsafe | loss of rare coverage |
| no-op | no action beats threshold | false stagnation if evidence is poor |
PROCEDURE run_teleodynamic_harness(ecology, workload, policy)
baseline <- REPLAY_WORKLOAD(ecology, workload)
proposals <- PROPOSE_STRUCTURAL_ACTIONS(ecology, baseline.failures)
candidates <- []
FOR each proposal IN proposals
IF POLICY_ALLOWS(proposal)
candidates.ADD(SIMULATE_ACTION(ecology, proposal))
END IF
END FOR
outcomes <- []
FOR each candidate_ecology IN candidates
score <- REPLAY_WORKLOAD(candidate_ecology, workload)
budget <- COMPUTE_RESOURCE_DELTA(candidate_ecology, ecology)
outcomes.ADD(VIABILITY_SCORE(score, budget, policy))
END FOR
best <- SELECT_BEST_OR_NOOP(outcomes, policy.threshold)
RETURN HARNESS_REPORT(baseline, proposals, outcomes, best)
END PROCEDUREHarness report
The report should include baseline behavior, proposed actions, rejected proposals, score deltas, resource deltas, threshold values, selected action, no-op rationale if selected, and rollback feasibility.
Test data discipline
Use separate sets for proposal discovery and final selection. If the same failures generate and judge the candidate, the candidate may overfit the failure cluster without improving the broader ecology.
Source reports used for this guide
These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.