Status: experimental algorithm proposal + simulator target
Date: 2026-09-22
Scope: routing, verifier allocation, generation backpressure, and diversity preservation for the Verified Swarm Runner / Connector Control Plane.
IDKMesh already has useful nature-inspired mechanisms:
The missing piece is a single closed-loop trust policy that joins those ideas around the product’s strongest question:
How should IDKMesh allocate workers and verifiers when generation is cheap, verification is scarce, and nominal reviewer count can badly overstate independent evidence?
AVE is a proposed answer. It is deliberately not a biological simulation and not a claim that software projects are organisms or markets. Biology, ecology, economics, and physics supply candidate control mechanisms. The mechanism stays only if it beats simpler baselines under matched budgets.
WorkUnit
|
v
policy/risk feasibility filter
|
v
niche-aware worker routing
| | -- entropy/temperature preserves exploration
| -- family occupancy penalizes monoculture
| -- shadow prices penalize scarce review/compute
v
untrusted candidate
|
v
danger / uncertainty estimate
|
v
immune-style verifier portfolio
| | -- known-bad probes maintain detector memory
| -- correlation penalty favors independent families
| -- risk-adaptive quorum/fan-out
v
VerificationResult + evidence
|
+--> update worker evidence
+--> update verifier evidence
+--> update queue shadow prices
+--> update concentration/diversity state
|
v
human/governance integration remains outside AVE
AVE is a recommendation and resource-allocation layer. It does not grant permissions, decide truth without verification, or merge.
Useful engineering inspirations:
IDKMesh mapping:
| Immune concept | AVE object |
|---|---|
| antigen | candidate artifact / known-bad probe |
| detector | verifier / validator |
| memory | verifier reliability posterior and breach history |
| immune diversity | verifier-family diversity |
| danger signal | risk + uncertainty + provenance anomaly + correlated-worker signal |
| tolerance | low-risk, well-characterized work uses minimal sufficient verification |
| immune overreaction | unnecessary verification cost / false rejection |
| immune escape | defect accepted by all selected verifiers |
Important boundary: “danger theory” is used only as an engineering analogy for combining contextual anomaly signals. It is not required as a biological claim.
Ecological coexistence work emphasizes that stable diversity depends on differences in niches/resource use; biodiversity can sometimes provide resilience when components respond differently to disturbances.
IDKMesh mapping:
The goal is not diversity for its own sake. Diversity is valuable only when it reduces correlated failure or covers a capability niche.
A verifier queue is a congestion problem.
Instead of a token or cryptocurrency, AVE uses an internal shadow price for scarce resources.
For resource r:
lambda_r(t+1) =
clip(
lambda_r(t)
+ eta_r * (utilization_r(t) - target_r),
0,
lambda_max
)
When review demand exceeds sustainable capacity:
lambda_review rises;When capacity recovers, the price falls.
This is inspired by network-utility/congestion-control work using feedback and shadow prices. It is not money and creates no transferable asset.
A fixed exploration rate is brittle. AVE uses a temperature-controlled distribution.
For eligible worker a and task j, define an adjusted utility U(a,j).
P(a -> j) =
exp(U(a,j) / T)
-----------------------------
sum_k exp(U(a,k) / T)
Interpretation:
T: uncertainty is high -> broaden exploration;T: evidence is mature or verification is congested -> exploit stronger routes;T is bounded; it never disables policy constraints.A first feedback rule can be:
T(t) = clip(
T_base
+ k_u * mean_route_uncertainty
+ k_s * stagnation_signal
- k_q * lambda_review,
T_min,
T_max
)
This is a practical use of the “free-energy” idea already discussed in SCIENTIFIC_FOUNDATIONS.md: quality pressure plus a controlled diversity/entropy term.
For each worker family f and task class c, maintain a posterior over verified success.
A simple Beta-Bernoulli starting point:
theta_(f,c) ~ Beta(alpha_(f,c), beta_(f,c))
A router can sample theta (Thompson-style) or use posterior mean + uncertainty.
Updates must use retained verification/outcome evidence, not raw activity volume.
This makes the routing question:
Which worker family should be tried for this task class, given observed verified outcomes and uncertainty?
rather than:
Which provider has the biggest model?
MAP-Elites is useful as a model for an archive indexed by meaningful behavioral descriptors.
A future AVE archive can index candidate worker/verifier strategies by dimensions such as:
Each cell retains one or a few high-performing verified configurations.
This prevents a globally strong provider from erasing specialized combinations that are better for a minority of tasks.
For WorkUnit w:
For worker/agent a:
For verifier v:
After the hard policy/permission filter, a first worker-task utility is:
U_worker(a,w) =
Value(w)
* Capability(a,w)
* SampledReliability(a, class(w))
* IndependenceBonus(a,w)
* NicheBonus(a,w)
/ (
1
+ lambda_review * ExpectedReviewCost(a,w)
+ lambda_compute * ExpectedComputeCost(a,w)
+ RiskFriction(a,w)
)
Terms:
Value: impact + information gain;Capability: declared/observed task fit;SampledReliability: posterior sample/estimate, preserving uncertainty;IndependenceBonus: higher for workers whose failures differ from active attempts;NicheBonus: negative-frequency dependence / anti-monoculture;The final choice is entropy-regularized, not deterministic greedy selection.
Let n_(f,c) be active attempts by worker family f in task class c.
A simple negative-frequency term is:
NicheBonus(f,c) =
1 / (1 + n_(f,c))^gamma
Add a global family-share cap:
share_f <= K_f(risk_class)
unless no feasible alternative exists.
For high-risk work, the allowed concentration can be stricter.
This is not a fairness score or contributor worth score. It is a correlated-failure control.
IDKMesh already has seeded known-bad candidate logic in the gate-audit direction.
AVE makes this an online control input.
For verifier v:
probe_reliability_v ~ Beta(a_v, b_v)
On a known-bad probe:
a_v += 1;b_v += 1.A verifier that repeatedly accepts known-bad probes is not silently trusted because of nominal status.
A first bounded score:
D(w,candidate) =
q_r * Risk(w)
+ q_u * RouteUncertainty(worker, w)
+ q_n * Novelty(w, candidate)
+ q_p * ProvenanceAnomaly(candidate)
+ q_c * CorrelatedWorkerPressure(w)
+ q_h * HistoricalEscapeRate(task_class)
D maps to a minimum verifier budget.
Example policy:
D < 0.30 -> minimum 1 verifier
0.30 <= D<0.65 -> minimum 2 verifier families
D >= 0.65 -> minimum 3 verifier families + human checkpoint if policy requires
The exact thresholds are experimental.
Select verifiers greedily or through a small combinatorial optimizer to maximize:
VerifierValue(S) =
ExpectedDetection(S)
+ Diversity(S)
+ Coverage(S)
- lambda_review * Cost(S)
- CorrelationPenalty(S)
subject to:
This is the component most directly tied to E017: N verifiers are useful only if their joint evidence is stronger than one verifier’s evidence.
Do not use one permanent majority rule.
For a verifier portfolio S, the aggregation rule should depend on:
E017 showed that for the measured one-sided partial-test panel, majority was a poor rule and unanimity-to-accept substantially reduced errors. That result must not be generalized to arbitrary two-sided verifiers.
AVE therefore treats quorum as a policy selected from evidence, not a constant.
Generation should slow when verification cannot keep up.
Let:
rho_verify =
arriving_verification_work /
effective_verification_capacity
When rho_verify > 1 persistently:
lambda_review;This is preferable to producing an ever-growing queue of untrusted candidates.
Add optional policy hooks, not vendor-specific logic:
route(work_unit, eligible_connectors, evidence_state) -> RouteDecision
select_verifiers(candidate, evaluator_requirements, evidence_state) -> VerificationPlan
update(outcome) -> EvidenceStateDelta
The connector kernel still owns:
AVE consumes normalized connector metadata and produces recommendations.
Use existing fields where possible. Do not add new schema fields until an implementation proves they are necessary.
Potential future fields/extensions:
AVE should generate or help select an EvaluatorPlan, but the plan remains verifier-owned and content-bound.
A router must not weaken:
The current gate-audit surface can become the first measurement input for AVE:
gate-audit output
-> measured verifier correlation
-> effective panel size / breach evidence
-> AVE verifier-family state
-> future verifier allocation
This creates a clean product progression from today’s tool to the future control plane.
Show the control state directly:
Compare under equal workloads and seeds:
Sweep:
Report a Pareto vector, not one permanent score:
AVE should be rejected or simplified if, under matched budgets:
A successful result is a reproducible Pareto improvement, not merely a more complex algorithm.
sim/;Only after the cohort shows benefit:
AVE MUST NOT:
Every policy update must be inspectable and replayable.
Primary external inspirations:
Repository evidence and foundations:
SCIENTIFIC_FOUNDATIONS.mddocs/algorithms/ACO_STIGMERGIC_TASK_ROUTING.mddocs/algorithms/HOMEOSTATIC_STIGMERGY_ROUTING.mdexperiments/E017-item-difficulty-and-quorum.mddocs/research/EVALUATOR_PLAN_BINDING.mddocs/architecture/AGENT_MODEL_CONNECTOR_CONTROL_PLANE.mdThe first cumulative matched-budget ablation adds an important caution to the original AVE proposal.
The strongest synthetic safety effect currently comes from verifier-family diversity. In contrast, using known-bad probes as a positive verifier-ranking signal is not yet justified: when probe coverage is unrepresentative, posterior updates can steer selection toward a misleading estimate of real verification quality.
Therefore the next candidate policy for product dry-run is deliberately smaller:
AVE-core =
ecological task anti-monoculture
+ posterior task-family learning
+ bounded entropy exploration
+ verifier-family diversity
+ risk-adaptive verifier floor/quorum
+ verification shadow-price backpressure
+ price-aware routing
Known-bad probes remain valuable, but for now their supported roles are:
gate-audit;They should not yet create positive production trust or automatically promote a verifier.
The targeted experiment is:
sim/adaptive_verification_ecology_targeted.pyIt compares:
ave-core;This is a design correction, not a claim that immune-style memory is useless. The claim under test is narrower: probe-derived routing trust must be earned by representative-outcome evidence rather than assumed from synthetic probe success.