Status: research protocol / architecture hypothesis
CEM is the current IDKMesh hypothesis for building coherent systems from vague or evolving goals without surrendering quality, safety, or auditability.
It combines:
The constitution is defined in ../../CONSTITUTION.md.
At iteration t, maintain:
G_t: current Goal Graph;A_t: archive of viable candidate approaches;E_t: accumulated evidence and failures;R_t: available human/compute resources;L: constitutional laws / hard invariants.A simplified loop is:
1. Observe unresolved/high-value regions of G_t.
2. Allocate bounded exploration/verification budget from R_t.
3. Generate candidate proposals using humans, agents, mutation, search, or recombination.
4. Reject candidates that fail constitutional viability gates L.
5. Place surviving candidates into behavior/architecture niches in A_t.
6. Emit bounded Work Units for experiments, implementation, tests, benchmarks, and attacks.
7. Execute Work Units on capability-matched workers.
8. Independently verify returned Result Manifests.
9. Add positive and negative evidence to E_t and the Goal Graph.
10. Update confidence, risks, and open questions in G_t.
11. Replace a niche incumbent only when the challenger has stronger verified evidence.
12. Reallocate resources using value, information gain, uncertainty, novelty, risk, and redundancy.
13. Integrate only through normal project/release gates.
14. Repeat.
Let candidate x have multi-objective measurements:
M(x) = [Q, S, R, A, N, I, C]
where for example:
Q: verified quality/usefulness;S: safety/security;R: robustness/reproducibility;A: adaptability;N: useful novelty;I: expected information gain;C: human + compute cost.Hard viability is separate:
V(x, L) = 1 if all required constitutional constraints are satisfied
0 otherwise
Only V(x,L)=1 candidates may enter the viable archive.
Do not permanently collapse M(x) into one score. Use Pareto dominance, niches, task-specific scalarizations, or temporary policies that remain inspectable and revisable.
A candidate research-allocation score might be:
Priority(x) = a*ExpectedValue + b*InformationGain + c*Novelty + d*RiskReduction - e*Cost - f*Redundancy
with an explicit protected exploration budget so the currently popular branch cannot monopolize all resources.
A niche is not necessarily a product category. It can be defined by behavior or architecture descriptors such as:
The descriptors themselves are research parameters and may evolve.
The archive should retain the strongest verified candidate per region rather than the most popular candidate.
Direct all-to-all communication does not scale to large populations. Agents should coordinate indirectly through durable traces:
A worker can act on the current environment without knowing every other worker. The environment becomes collective memory.
Positive feedback is useful:
But every positive loop needs damping:
Without negative feedback, the mesh can converge prematurely or amplify a wrong belief.
CEM should evolve itself gradually.
Goals, tasks, priorities, proposals, and experiments can change freely within ordinary project rules.
Scheduler weights, exploration temperature, replication factors, and queue limits may change automatically inside bounded ranges and with metrics/rollback.
Schedulers, aggregation rules, verifier strategies, and reputation formulas may compete experimentally. Promotion requires benchmark evidence and compatibility review.
Changes to the rules that judge future changes require the strongest gates, public rationale, adversarial analysis, and explicit human governance.
The system must never allow a self-improving mechanism to make itself successful merely by weakening its own evaluation criteria.
CEM is not:
Every mechanism remains a falsifiable hypothesis.
See:
../../sim/emergence_sim.py../../tests/test_emergence_sim.py../../experiments/E011-emergence-vague-goals.md../../experiments/results/E011-reference-seed7.jsonThe model is intentionally small. Its value is that IDKMesh now has a concrete mechanism that can be attacked, compared, improved, or disproved.