Evaluation evolution and memory

Evaluation, evolution & memory

GenOS treats alternative agent trajectories as candidates that can be measured, selected, inherited, or rejected. This is where the project goes beyond snapshot tooling and becomes a laboratory for agent behavior.

Evaluation as a first-class phase

Branch execution and branch selection are separate operations. A model or policy may generate a candidate; evaluation decides whether the candidate deserves more budget or promotion.

hypotheses → branch outcomes → evidence vectors → selection policy → promotion

An outcome can include:

• invariant and test results;
• task-specific correctness;
• runtime, token, or model-cost observations;
• security or compliance signals;
• semantic/structural divergence;
• novelty, risk, or confidence;
• replay equivalence;
• reviewer approval.

Multi-objective selection

Real decisions rarely optimize one number. A minimal patch may be cheaper while a refactor is more maintainable; one agent may be accurate but slow, another fast but risky.

GenOS includes Pareto selection experiments. A candidate remains non-dominated when no other candidate is at least as good on all declared objectives and strictly better on one.

                 higher correctness
                        ▲
              C ●       │       ● B  ← non-dominated
                        │
           A ●          │
                        └────────────────► lower cost

The frontier preserves meaningful trade-offs instead of hiding them inside an arbitrary weighted average.

Branch evolution

The branch-evolution runtime models a bounded search process:

1. Allocate a global compute budget.
2. Spawn branches from a common capsule.
3. Run and evaluate each branch.
4. Terminate failed or dominated branches.
5. Reallocate budget toward promising survivors.
6. Recursively split a survivor when additional exploration is justified.
7. Preserve lineage for every allocation and death decision.

Core mechanics are implemented. Integration with a live evaluator across every runtime surface remains incomplete.

Genome, phenotype, and controlled variation

The genome/phenotype split supports questions that ordinary prompt orchestration cannot represent cleanly:

• Did behavior change because of inherited configuration or experience?
• Is a measured trait stable across environments?
• Does a mutation improve one objective while harming another?
• Can two parent genomes produce a valid child with traceable ancestry?
• Is a claimed trait reproducible enough to promote?

Repository experiments cover mutation, reversible mutation, breeding, phenotype records, controlled cohorts, factorial interactions, trait claims, heritability, and promotion.

These features are research primitives, not a license for agents to evolve without governance. Mutation and trait promotion remain explicit operations with validation and lineage.

Cognitive merge

When branches contain useful but different information, “take the winner” may be too crude. Cognitive merge uses a manifest to reconcile selected evidence and compatible state.

branch A: passing patch + test evidence ─┐
                                        ├─ reviewed synthesis → successor S1
branch B: root-cause insight ────────────┘

Good merge behavior should:

• retain source branch and evidence links;
• reject contradictions that lack resolution;
• avoid copying complete private context by default;
• validate the successor state;
• record the merge decision as a new lineage event.

Experience packets and memory

An experience packet packages reusable knowledge from a trajectory: context, observation, evidence, applicability, and provenance. It can enter a typed knowledge graph and later be synthesized into another capsule after review.

The synaptic research surface explores associative memory, Hebbian-style strengthening, spike-timing-dependent plasticity (STDP), pruning, and consolidation. These mechanisms are experimental models for retrieval and adaptation, not biological claims about LLM cognition.

Beliefs, contradictions, and claims

GenOS examples model beliefs as structured records rather than undifferentiated prose. This enables:

• confidence updates;
• evidence attachment;
• contradiction detection;
• source and lineage queries;
• claim replication across cohorts;
• trait or knowledge promotion only after a policy gate.

Biomimetic resilience and swarms

The repository also explores mechanisms inspired by biological systems: apoptosis, cryptobiosis, hypermutation, pheromone-like coordination, quorum and huddle consensus, horizontal transfer, and energy-aware behavior.

Some primitives and simulations exist, but the integrated self-healing swarm runtime is a later roadmap milestone. Read the examples as a research program, not as a claim of production autonomy.

The Examples & evidence page maps these ideas to runnable demonstrations.