idkmesh

Conversation record — executable ACO continuation

Date: 2026-08-28

Trigger

After adding the biology-inspired ACO stigmergic task-routing proposal, the project owner asked to continue.

Initial decision

Do not connect ACO directly to live GitHub task assignment. First convert the biological analogy into deterministic, falsifiable experiments with simple baselines and CI.

Phase 1 — fixed ACO implementation

Added:

• sim/aco_stigmergy_sim.py
  • synthetic heterogeneous tasks and workers;
  • random, greedy, capability-only, and ACO routing;
  • evidence-backed pheromone deposits;
  • evaporation and bounded pheromone state;
  • congestion and correlated-attempt penalties;
  • explicit exploration floor;
  • JSON output and reproducible seeds.
• sim/run_aco_sweep.py
  • repeated-seed comparison using the same generated worker population per strategy/replicate.
• tests/test_aco_stigmergy_sim.py
  • mechanics, determinism, probability normalization, evaporation, bounds, reinforcement, and anti-congestion tests.
• experiments/E014-aco-stigmergic-task-routing.md
  • predeclared question, baselines, metrics, rejection conditions, and evidence ladder.

Phase 2 — first synthetic result

A 40-seed reference comparison produced a mixed result:

strategy        utility/cost   duplicate   coverage   concentration
random             0.3997        0.2266      8.000       0.1391
greedy             1.1322        0.8333      1.000       1.0000
capability         0.8690        0.4281      4.825       0.3833
fixed ACO          0.5584        0.1551      8.000       0.1944

ACO reduced duplication and concentration and preserved full task coverage, but it did not beat capability-only or greedy routing on immediate utility per cost.

Decision: preserve the mixed result rather than tuning it away.

Results:

• experiments/results/E014-reference-sweep.json
• experiments/results/E014-reference-sweep.md

Phase 3 — parameter/Pareto sensitivity

Added sim/run_aco_parameter_sweep.py and tests.

A sweep of 24 ACO parameter configurations over 12 seeds produced 11 non-dominated Pareto points. The best-efficiency tested ACO point reached approximately:

utility/cost       0.6071
duplicate rate     0.1833
coverage           8/8
concentration      0.2108

The result confirmed a stable tradeoff rather than one universally best setting: stronger exploitation improved efficiency while consuming diversity/congestion margin.

Results:

• experiments/results/E014-parameter-pareto.json
• experiments/results/E014-parameter-pareto.md

Phase 4 — Homeostatic Stigmergic Routing

The mixed ACO evidence motivated a synthesis rather than blind parameter optimization.

Homeostatic Stigmergic Routing (HSR) combines:

1. capability matching for exploitation;
2. evidence-backed stigmergic memory;
3. density-dependent negative feedback that increases diversity/congestion pressure only when crowding exceeds healthy targets.

Core form:

Score(a,j,t) = CapabilityValue(a,j)
               * tau_j(t)^alpha
               * [Diversity(a,j,t) * Congestion(j,t)]^lambda(t)

with feedback:

lambda(t+1) = clip(
    lambda(t)
    + k_d * (DuplicateRate(t) - D_target)
    + k_s * (Concentration(t) - S_target)
    - r * (lambda(t) - lambda_min),
    lambda_min,
    lambda_max
)

Added:

• sim/homeostatic_stigmergy_sim.py
• tests/test_homeostatic_stigmergy_sim.py
• docs/algorithms/HOMEOSTATIC_STIGMERGY_ROUTING.md

A 40-seed comparison produced:

strategy             utility/cost   duplicate   coverage   concentration
fixed ACO               0.5584        0.1551      8.000       0.1944
capability              0.8690        0.4281      4.825       0.3833
homeostatic hybrid      0.6754        0.2796      7.850       0.2515

HSR is a promising intermediate Pareto operating point, not a universal winner. It moves toward capability efficiency while retaining much lower duplication/concentration and near-full coverage.

Results:

• experiments/results/E014-homeostatic-hybrid.json
• experiments/results/E014-homeostatic-hybrid.md

Phase 5 — historical replay interface

Synthetic evidence is not enough for live use. The next gate is read-only historical/shadow replay.

Added:

• schemas/routing-replay-v0.schema.json
  • separates routing-time inputs from retrospective outcomes;
  • records feature/profile provenance and annotation policy;
  • makes hindsight leakage an explicit dataset concern.
• sim/replay_task_routing.py
  • performs deterministic capability, fixed-ACO, and HSR recommendations from frozen snapshots;
  • never calls GitHub or changes repository state;
  • applies only evidence known at the snapshot time;
  • evaluates held-out retrospective outcomes only after recommendations are fixed.
• tests/test_replay_task_routing.py
  • includes an anti-hindsight test: arbitrarily changing future outcomes must not change recommendations.
• examples/routing-replay.example.json
  • illustrative IDKMesh-shaped example only; all annotations/outcomes are clearly marked synthetic and are not Level-2 evidence.

The legitimate next dataset must freeze real source timestamps and annotation rules before evaluation.

Evidence ladder

The research must advance only through explicit stages:

1. unit/mechanics tests — implemented and passing;
2. repeated synthetic simulation — active; mixed results preserved;
3. historical repository replay — interface implemented; real frozen dataset still required;
4. advisory shadow mode — not authorized yet;
5. bounded live experiment — not authorized yet.

A favorable synthetic result is not permission for autonomous production routing.

Safety / anti-Goodhart constraints

• Pheromone is routing memory, not truth, reputation, or authority.
• Unverified popularity/activity must not create strong reinforcement.
• Routing never grants permissions, approves work, merges work, or bypasses verification.
• Evaporation, lower/upper bounds, explicit exploration, congestion penalties, and correlation discounts preserve alternatives and reduce herding.
• Homeostatic regulation is bounded and explicit rather than hidden tuning.
• Retrospective outcomes must remain held out from historical routing inputs.
• Negative and mixed experimental results are durable project evidence and may justify rejecting ACO/HSR.

Current research conclusion

The most useful biological insight from this iteration is broader than Ant Colony Optimization itself:

Exploit strongly while capacity is healthy; when density, duplication, correlation, or reviewer pressure rises, increase negative feedback and diversify effort; relax the inhibition again when the system returns to a healthy region.

This principle may later apply to task routing, Growth Seed generation, verifier allocation, compute scheduling, experiment portfolios, and repository self-evolution, but each application requires its own evidence and safety gates.

This site is open source. Improve this page.