Status: current research registry
Date: 2026-09-22
Purpose: keep cross-disciplinary mechanisms attached to concrete IDKMesh problems, evidence, and promotion gates.
IDKMesh borrows mechanisms, not prestige or metaphors.
A proposal belongs in the architecture only when it has:
Nature-inspired policy never overrides constitutional constraints such as:
| Level | Meaning |
|---|---|
| N0 — inspiration | idea documented; no executable mechanism |
| N1 — executable synthetic | deterministic simulator/tests exist |
| N2 — retained synthetic evidence | comparative result retained with limitations |
| N3 — real dry-run | observes or recommends against real project data but cannot actuate |
| N4 — bounded live cohort | limited actuation behind existing policy/permission gates |
| N5 — production-eligible | repeated evidence, rollback, monitoring, and governance support promotion |
No mechanism should skip levels merely because its scientific source is well established. A mechanism can be valid in biology/economics/physics and still be wrong for IDKMesh.
| Mechanism | Inspiration | IDKMesh subsystem | Maturity | Current evidence / artifact | Promotion question |
|---|---|---|---|---|---|
| Verified stigmergy / ACO | ant-colony pheromone trails | task / WorkUnit routing | N2 | docs/algorithms/ACO_STIGMERGIC_TASK_ROUTING.md, E014 |
does verified trace memory beat simpler routing after charging for diversity and review? |
| Homeostatic stigmergy | density-dependent biological regulation + feedback control | task routing / duplication control | N2 | docs/algorithms/HOMEOSTATIC_STIGMERGY_ROUTING.md + simulator/tests |
does adaptive diversity pressure improve the Pareto frontier over fixed ACO/capability routing? |
| Quality-Diversity / MAP-Elites | evolution/ecological niches | architecture / worker-policy archive | N0/N1 | issue #22 and research design | can multiple verified specialists be preserved without keeping inferior variants for diversity alone? |
| Adaptive Verification Ecology (AVE) | immunity + ecology + congestion economics + entropy | worker/verifier allocation and generation backpressure | N2; N3 shadow adapter ready | PR #622/#645, issues #621/#644, AVE-0/1/2 | does AVE-core make useful pre-outcome recommendations on real canonical WorkUnits without weakening EvaluatorPlan gates? |
| Verifier-family diversity | ecological niche separation / portfolio diversification | verifier selection | N1/N2 research branch | AVE ablations + E017 motivation | does family diversity add independent evidence rather than nominal variety? |
| Known-bad verifier probes | artificial immunity / adversarial testing | gate audit and verifier diagnostics | N2 diagnostic; not trust authority | gate-audit direction, AVE probe tests | do probe results predict live verifier failures on representative tasks? |
| Review shadow price | congestion pricing / network utility | verification queue and optional fan-out | N1/N2 | AVE + verification-debt/backpressure work | does a scarcity signal control verification debt without starving important work? |
| Entropy / temperature exploration | statistical mechanics / entropy regularization | routing exploration | N1 | AVE and scientific foundations | does adaptive exploration improve recovery from shift enough to pay its cost? |
| Physarum conductance routing | slime-mold adaptive transport networks | admitted multi-node compute/federation paths | N2; N3 blocked on topology evidence | PR #631, issues #630/#649, PHY-0/1 | is multi-hop/federated path routing a real IDKMesh need, and can real admitted topology/telemetry exist without inventing edges? |
| Replicator-mutator policy weights | evolutionary dynamics | ACE community-growth strategy controller | N1 offline | issue #57 / ACE controller design | do strategy weights improve verified descendants per reviewer/maintainer attention? |
| Carrying-capacity governor | population ecology / logistic regulation | community/reviewer growth | N1 | ACE design | does growth stop before reviewer load becomes the bottleneck? |
| Verification backpressure | queueing/control theory | generation vs verification capacity | N2 | roadmap, ADR-0007 and verification-scaling research | can verification debt be bounded while preserving high-value throughput? |
| Mechanism | Potential use | Current decision |
|---|---|---|
| Percolation thresholds | predict fragmentation under node/provider loss | N0 — simulation/observability candidate for multi-machine stage |
| Spectral connectivity / graph Laplacian | detect weak links, partitions, slow information mixing | N0 — use as diagnostics before allowing it to influence scheduling |
| Epidemic/contagion models | malware/bad-evidence propagation and containment | N0 — security simulation only |
| Simulated annealing | task partitioning, test portfolio, architecture search | N0 — compare to MILP/bandit/evolutionary baselines first |
| Free-energy objective | explicit quality/cost vs useful-diversity trade-off | N0/N1 research concept — never replace the underlying metric vector with one permanent scalar |
| Reaction-diffusion / morphogenesis | decentralized specialization of cells/agents | HOLD — no demonstrated gap that requires it yet |
| Predator-prey dynamics | generation-verification balance | HOLD — current queue/backpressure/AVE mechanisms already address this; add only if they fail in a measurable regime |
| Neural/homeostatic plasticity | adaptive routing weights | HOLD — likely overlaps bandit/controller machinery |
| Genetic/evolutionary mutation | generate architectures/policies | proposal-only — randomness may propose; independent verification still decides acceptance |
| Blockchain/token economics | cross-org settlement / incentives | not a trust mechanism — existing project decisions require a concrete demonstrated settlement need first |
| Quantum-inspired optimization | constrained routing/partitioning | optional research — conventional baselines remain mandatory |
Use different mechanisms for different layers.
goal / task selection
-> Quality-Diversity, bandits, stigmergy, homeostasis
worker + verifier allocation
-> AVE, measured error correlation, backpressure
compute admission
-> hard repository policy, authorization, capability/trust filters
compute path routing after admission
-> Physarum candidate, conventional network-routing baselines
community strategy
-> ACE carrying capacity, replicator-mutator controller
integration
-> evidence + independent verification + human/governance authority
Do not combine these into one “bio-inspired score.”
A single opaque score would make attribution, debugging, and falsification harder.
Every new mechanism should add a record containing:
problem:
mechanism:
source_domain:
state_variables:
hard_constraints:
baseline_1:
baseline_2:
strong_baseline:
hypothesis:
matched_budget:
metrics:
known_failure_modes:
falsification_rule:
authority_boundary:
real_data_gate:
rollback_or_disable_path:
Requires:
Requires:
Requires real project data or a controlled real node/candidate corpus.
The policy remains advisory/dry-run.
Use the common shadow-evidence protocol introduced by PR #637 / issue #636:
adaptive-policy-plan-v0.1 against an exact revision/input digest;adaptive-policy-outcome-v0.1;shadow_counterfactual_observed=false and causal_effect_estimate=null.N3 should answer whether a policy makes useful, measurable, materially different recommendations on real state. It does not establish that an unexecuted recommendation would have caused a better result.
Requires:
Requires repeated operation across relevant workload regimes, observed failure handling, and a governance decision.
Reject these patterns:
“Ants/slime molds/immune systems are robust, therefore IDKMesh should copy them.”
Invalid. The IDKMesh mechanism must independently earn its place.
“ACO + immune memory + markets + thermodynamics must be stronger together.”
Not necessarily. AVE ablation exists specifically to remove components that do not contribute enough value.
High use, many comments, model fame, or route frequency do not prove correctness or independence.
Different model/provider/agent names are only hypotheses about independence. Measured error behavior is stronger evidence.
A “price”, “budget”, or “market” inside an algorithm does not grant permission to spend real money.
No learning/routing algorithm may let a worker satisfy the verifier independence and EvaluatorPlan ownership boundaries for its own result.