idkmesh

ACO Stigmergic Task Routing for IDKMesh

Status: Experimental algorithm proposal.

Why this biological algorithm

If IDKMesh adds one biology-inspired algorithm as an executable coordination primitive, a strong candidate is Ant Colony Optimization (ACO).

Ant colonies coordinate without a central planner by leaving local environmental traces (pheromones). Useful routes become easier to rediscover; unused traces decay. IDKMesh has a natural equivalent: issues, Work Units, evidence, pull requests, reproductions, reviews, benchmarks, and verified descendants are public traces that future humans/agents can observe.

The engineering hypothesis is:

Verified useful work should leave a decaying stigmergic signal that increases the probability that compatible future contributors discover and select related high-value work, while congestion, review load, risk, and herding reduce that probability.

This is an experiment, not a claim that open-source communities literally behave like ants.


1. Core state

For each currently selectable task or Work Unit j, maintain a pheromone-like scalar:

tau_j(t) >= 0

tau_j is not reputation and not priority by fiat. It is a decaying public signal representing accumulated evidence that useful descendants have recently emerged from that work path.

For each worker/agent a and task j, compute a local heuristic desirability:

eta_(a,j)(t).

The worker can then probabilistically select among feasible tasks rather than always choosing the globally highest score.


2. Data required

Task / Work Unit data

Each candidate task j should expose or estimate:

Worker / agent data

For candidate worker a:

Worker identity is not required for the mathematics; privacy-preserving or pseudonymous capability descriptors can be used where appropriate.

Evidence data

After work is attempted:

Only evidence-backed outcomes should create strong positive pheromone deposits.


3. Local desirability formula

For worker a and feasible task j, define a bounded heuristic:

eta_(a,j) = (I_j * G_j * S_(a,j) * D_(a,j) * F_j * A_j) / ((1 + H_j) * (1 + L_j) * (1 + X_j))

where:

All positive components should be normalized to stable ranges (for example [epsilon, 1]) before multiplication.

The important IDKMesh-specific term is D_(a,j): ten nearly identical agents should not look ten times more desirable than one independent approach.


4. Stigmergic task-selection probability

For a worker a, let F_a(t) be the set of tasks currently feasible under dependencies, permissions, resource limits, and policy.

Choose task j with probability:

P(a -> j | t) = [tau_j(t)^alpha * eta_(a,j)(t)^beta] / sum_(k in F_a(t)) [tau_k(t)^alpha * eta_(a,k)(t)^beta]

Parameters:

For IDKMesh bootstrap experiments, prefer moderate beta and conservative alpha to avoid early lock-in from tiny samples.

A minimum exploration floor should ensure that low-pheromone tasks remain discoverable:

tau_j := max(tau_min, tau_j).


5. Pheromone update

The basic biological ACO update becomes:

tau_j(t+1) = clip((1-rho) * tau_j(t) + Deposit_j(t) - Penalty_j(t), tau_min, tau_max)

where:

Verified-evidence deposit

For each verified descendant e of task j:

DeltaTau_(e,j) = (Q_e * V_e * D_e * U_e) / (1 + H_e + C_e)

where:

Then:

Deposit_j(t) = sum_e DeltaTau_(e,j).

This means an inexpensive, independently verified contribution that creates useful follow-up work deposits more signal than a high-volume but weakly verified artifact.

Congestion / failure penalty

A simple bounded penalty can be:

Penalty_j = kappa_1 * Overload_j + kappa_2 * Defect_j + kappa_3 * Correlation_j

where:

A failed experiment should not necessarily receive a large penalty if it produced valuable information. Information gain belongs in the evidence evaluation.


6. Anti-herding rule

Classic ACO can converge too strongly on one path. IDKMesh needs persistent diversity.

Use at least three controls:

  1. pheromone evaporation (rho);
  2. lower/upper pheromone bounds (tau_min, tau_max);
  3. a diversity/congestion term in eta_(a,j).

A useful congestion multiplier is:

C_j = 1 / (1 + n_j)^gamma

where n_j is the number of currently active correlated attempts.

Then replace:

eta_(a,j) <- eta_(a,j) * C_j.

This naturally pushes later workers toward neglected tasks or independent approaches when one path becomes crowded.


7. GitHub-native mapping

A first GitHub implementation does not need a separate distributed database.

Biological ACO concept IDKMesh/GitHub equivalent
ant human or AI worker
path task / Work Unit / evidence route
pheromone decaying evidence-backed task signal
food found verified useful artifact/descendant
pheromone deposit verified outcome score
evaporation time decay
crowded trail too many active/correlated attempts
colony exploration probabilistic assignment / discovery
environmental constraint permissions, dependencies, review capacity, risk

Possible storage for an early prototype:

The pheromone value should be reproducible from public evidence wherever possible.


8. Minimal algorithm

For each routing epoch:

1. Read feasible tasks and current evidence state.
2. Evaporate all pheromone values.
3. For newly verified descendants, deposit evidence-weighted pheromone.
4. Apply bounded penalties for overload, regressions, and correlated duplication.
5. For each available worker:
      a. filter tasks by dependency, permission, risk, and resource constraints;
      b. compute worker-task heuristic eta;
      c. compute probabilistic ACO selection P(a -> j);
      d. sample or recommend a task;
      e. preserve a non-zero exploration probability.
6. Do not treat selection as authorization to merge or bypass verification.
7. Record outcome/evidence for the next epoch.

The algorithm can recommend work; canonical acceptance remains governed by IDKMesh verification and human/policy gates.


9. First falsifiable experiment

Compare four task-routing strategies on the same bounded task pool:

  1. random feasible routing;
  2. greedy highest estimated value;
  3. capability-only matching;
  4. ACO stigmergic routing with evaporation + diversity/congestion penalty.

Use a mixture of humans/agents or a simulator with heterogeneous workers.

Measure:

Falsification criterion

Do not adopt ACO merely because it sounds biologically elegant.

ACO should be considered useful only if, across repeated runs, it improves a predeclared multi-objective outcome such as:

VerifiedUtility / (HumanAttention + ComputeCost)

without materially worsening defects, concentration, accessibility, or review backlog compared with simpler baselines.


10. Safety and governance constraints


11. Why this is particularly suitable for IDKMesh

ACO operationalizes several existing IDKMesh principles in one small mechanism:

The most important idea is not “copy ants.” It is:

Convert verified outcomes into decaying public coordination signals, then let many independent actors probabilistically follow or challenge those signals under explicit capacity and diversity constraints.