IDKMesh

ADR-0013 — Cross-disciplinary algorithms remain optional policies behind hard gates

Status: Proposed
Date: 2026-09-22

Context

IDKMesh intentionally studies mechanisms from biology, ecology, economics, physics, control theory, and complex systems.

Examples already include:

• ACO/stigmergic routing;
• homeostatic regulation;
• Quality-Diversity;
• verification backpressure;
• Adaptive Verification Ecology;
• Physarum-inspired adaptive compute routing;
• ACE evolutionary/community controllers.

These mechanisms can improve exploration, adaptation, resource allocation, resilience, or diversity. They can also create herding, reward hacking, over-exploration, excess review cost, unstable feedback, or opaque policy behavior.

The scientific origin of a mechanism is not evidence that it is correct for software-development coordination.

IDKMesh also has hard authority and safety boundaries that must not become tunable outputs of an adaptive controller.

Decision

Cross-disciplinary algorithms are optional policy layers operating inside a feasible set defined by hard gates.

Conceptually:

constitutional / repository constraints
        |
        v
hard feasibility and authority gates
        |
        v
eligible action set
        |
        v
optional learned / nature-inspired policy
        |
        v
recommendation or bounded action
        |
        v
verification / evidence
        |
        v
explicit integration authority

An adaptive mechanism may rank, sample, allocate, or recommend among already eligible choices.

It may not create eligibility.

Hard constraints outside adaptive authority

At minimum, adaptive algorithms cannot relax:

• WorkUnit scope, permissions, or source binding;
• repository compute-spend policy;
• resource/compute admission;
• secret boundaries;
• minimum trust/security requirements;
• required validator coverage;
• EvaluatorPlan ownership and content binding;
• worker/verifier identity-separation requirements;
• branch protection;
• explicit human/governance merge/integration authority.

If the eligible set is empty, the correct outcome is no route/no action, not policy relaxation.

Evidence rule

Every adaptive mechanism must progress through explicit maturity stages documented in:

docs/algorithms/NATURE_INSPIRED_ALGORITHM_REGISTRY.md

Promotion requires stronger evidence at each stage.

For the N2 -> N3 transition, adaptive mechanisms should use the common shadow-evidence contract from PR #637 / issue #636 so recommendations are frozen before outcomes, exact-revision/input-bound, and retrospectively joined without inventing the unexecuted counterfactual.

A mechanism must be removable if a simpler baseline matches its result.

Interpretability rule

Each decision-support mechanism must expose enough state to answer:

• what hard constraints filtered the action set?
• what policy selected among the remaining choices?
• which evidence/state influenced the policy?
• what uncertainty or exploration was present?
• what resource cost did the policy trade for its benefit?
• what would have happened under a named baseline where feasible?

Opaque composite “swarm intelligence” scores are discouraged.

No trust-by-identity rule

Provider, model, agent, organization, popularity, or family labels may be routing features, but they are not correctness evidence.

In particular:

different family label != independent evidence
many reviewers          != many independent reviewers
popular route           != correct route
high probe score         != safe live verifier

Observed verified outcomes and measured failure dependence are preferred where available.

Economic terminology rule

Internal shadow prices, utility, budgets, auctions, or market-like mechanisms do not grant financial authority.

Repository monetary policy remains a hard external constraint.

Consequences

Positive:

• research can remain innovative without weakening safety/governance boundaries;
• algorithms can be compared and removed independently;
• negative results become actionable simplifications;
• subsystem ownership remains understandable;
• real deployment can begin in dry-run/explain mode.

Costs:

• some adaptive mechanisms will appear less powerful because they cannot alter hard constraints;
• experiments need stronger baselines and cost accounting;
• several policies may coexist rather than one global optimizer;
• promotion to live use is slower.

These costs are intentional.

Related

• SCIENTIFIC_FOUNDATIONS.md
• docs/algorithms/NATURE_INSPIRED_ALGORITHM_REGISTRY.md
• ADR-0006 zero-project-spend compute
• ADR-0007 verification-debt backpressure
• ADR-0009 evaluator sovereignty
• PR #622 Adaptive Verification Ecology
• issue #621 AVE ablation
• PR #631 Physarum compute routing
• issue #630 Physarum stress testing
• PR #637 adaptive policy shadow contract
• issue #636 N3 real shadow cohort

This site is open source. Improve this page.