The Natural Parallel Test

The analogies that fail are the point

The model that finds an analogy is the wrong one to judge it. Finding parallels is what it is for. So candidates go to a separate agent, analogy-auditor, running on a different model with refuted as its default verdict. It checks the biology before it checks the logic, because attractive analogies travel on appeal rather than evidence. These three still turn up in strategy decks.

• Determination · analogy-auditor

  The alpha wolf

  Refuted

  Taken from captive, unrelated wolves. Wild packs are families, and the researcher who put the word into circulation spent years trying to take it back.

• Determination · analogy-auditor

  The queen who decides

  Refuted

  She lays eggs. She directs nothing. Where the colony moves next is settled by scouts advertising sites and a quorum forming at one of them.

• Determination · analogy-auditor

  Trees sharing food through fungi

  Contested

  Bilateral trade with sanctions against poor partners holds up. The generous-sharing story does not. Library entry E-23 carries that flag instead of the story.

Rejected candidates stay in the report. They are the evidence that a search happened at all.

How a run goes

Eight steps. Point the plugin at a plan, an architecture, an org design or a policy, and it works through them in order.

1. 1State the problem, not the solution, so the proposed answer stops steering the search
2. 2Split the plan into ten mechanisms and mark each one present, absent or not applicable
3. 3Search seven levels of nature in parallel, one scout per level
4. 4Put every candidate through four validity tests
5. 5Rewrite each survivor as a rule in the language of the plan
6. 6Name what the natural system tolerates that this plan cannot
7. 7Score twelve durability principles from cited evidence
8. 8Classify, and list what was rejected

Ten mechanisms

A plan almost never has one natural parallel. It has a strong one for how information moves, a weak one for how decisions get made, and none at all for how old parts are retired. A single global analogy averages those together and hides the mechanism that will fail. An absent mechanism is a finding, not a gap in the search.

• 1

  Resource flow

  How do energy, money, people and materials move?

• 2

  Information flow

  How does the system detect what is happening?

• 3

  Decision-making

  Who decides, and on what information?

• 4

  Coordination

  How do separate parts work together?

• 5

  Adaptation

  How does the plan change when reality changes?

• 6

  Growth

  How does it expand without collapsing?

• 7

  Repair

  How does it recover from damage?

• 8

  Selection

  How are bad ideas removed and good ones kept?

• 9

  Boundaries

  What keeps harmful things outside?

• 10

  Renewal

  How are old parts replaced?

Distance between sensing and acting is the single most predictive measurement in the framework. It gets recorded explicitly.

Seven levels of nature

Searched in parallel so the answer cannot collapse into animals, which is the most common failure of this method. The analogue library holds around thirty-five entries spread across the seven.

• Physical systems 10⁻⁹–10⁶ m

  Gradients, crystal growth, drainage networks, the turbulent transition. Least open to storytelling, so search here first.

• Cells 10⁻⁶–10⁻⁴ m

  Selective membranes, immune memory, programmed death, DNA proofreading, one shared energy currency.

• Organisms 10⁻³–10¹ m

  Reflex arcs beside central integration, homeostasis, wound healing, bone remodelling under load.

• Social species 10⁰–10³ m

  Quorum decisions, pheromone trails, response-threshold division of labour, mound ventilation.

• Ecosystems 10³–10⁷ m

  Succession, keystone effects, the gradient from mutualism to parasitism, response diversity.

• Evolution 10⁶–10⁹ yr

  Variation and selection, duplication then divergence, Red Queen dynamics, modularity as a condition for change.

• Cosmic systems 10⁹–10²⁶ m

  Accretion, hydrostatic equilibrium, self-limiting star formation, orbital resonance.

Six of those are lengths. One is a duration, which is why it is searched on its own axis: a mechanism that pays off over ten generations is not available to a two-quarter plan.

Four validity tests

Each candidate gets all four. One failure rejects it.

1. 1 Same problem

   Compare the function. Not the domain, not the shape, not the vocabulary.

   Automatic fail: both things are called networks

2. 2 Similar constraints

   Scarcity, uncertainty, delay, competition, failure, one at a time. The one people skip is who pays for failure.

   Automatic fail: it works because failed units die

3. 3 Comparable mechanism

   Write the natural mechanism out step by step, then write what the plan would have to do to run it. Most candidates die here.

   Automatic fail: it needs genetic relatedness

4. 4 Relevant scale

   Size, speed and timescale. All three have to hold.

   Automatic fail: colony scale is 10⁴–10⁶ and the plan has twelve people

Twelve durability principles

Each one scored against written anchors, so the number means the same thing on the second run as on the first.

• 1Feedback

  Can the system detect the results of its actions?

• 2Adaptation

  Can it change without being redesigned?

• 3Redundancy

  Can one component fail without taking everything with it?

• 4Diversity

  Are there several ways to solve the same problem?

• 5Modularity

  Can one part change without destabilising the whole?

• 6Local autonomy

  Can local parts act without waiting for permission?

• 7Coordination

  Do local actions still support the larger system?

• 8Resource limits

  Does the plan recognise finite money, energy and attention?

• 9Repair

  Can damaged parts be restored or replaced?

• 10Selection

  Can failing ideas be stopped?

• 11Renewal

  Can old structures be retired safely?

• 12Environmental fit

  Does the plan suit the conditions it will operate in?

Four verdicts, and five things that beat the score

• Established natural pattern

  19–24

  A close analogue for the plan's dominant mechanism, under similar constraints, through something the plan can actually run. Needs an accepted analogue at high or medium confidence, not just a high total.

• Partial natural pattern

  6–18

  Some mechanisms have strong analogues and others have none. The most common result. Its value is the list of unsupported mechanisms rather than the number at the top.

• Novel but compatible

  no band, assessed on principles

  No close analogue was found, but the plan follows principles that last. Not a lesser result. It needs an honest search across at least five levels, shown in the coverage log, to separate it from not having looked hard enough.

• Contrary to natural patterns

  0–5, or any override below

  The plan depends on something nature does not sustain. This verdict gets the strongest challenge and the plainest wording.

Five dependencies that beat the score

Any one of these makes a plan Contrary whatever it totals, and the report names which one fired. Twelve generous readings should not outvote one structural dependency.

• Permanent controlOperation needs one central authority available at all times, with no defined degraded mode when it is not.
• Unlimited growthGrowth is planned against a resource with no stated limit, or the binding constraint does not grow with demand.
• Perfect predictionCorrectness rests on a forecast being right, with no revision path for when it is wrong.
• Zero failureNothing is allowed to fail. No failure analysis, no degraded operation.
• No adaptationThere is no revision path at all.

Five commands

• /npt:test

  The full eight-step assessment: mechanisms, analogue search, adversarial audit, score, verdict. Run it before committing to a plan.

• /npt:worksheet

  One page, one pass, no subagents. A fast read, or a way to decide whether the full test is worth the time.

• /npt:score

  The twelve-principle scorecard on its own, with no analogue search. For comparing two plans, or re-checking one after it changes.

• /npt:challenge

  Adversarial. It starts from the assumption that the plan is Contrary and tries to prove it. Use it when a plan was agreed too easily.

• /npt:analogues

  Search only, across all seven levels, for one problem under stated constraints.

Each takes a file path, a description, or falls back to whatever plan is under discussion. The skill also fires on its own when a nature analogy shows up in conversation.

Install

Two lines in Claude Code.

/plugin marketplace add cocodedk/natural-parallel-test
/plugin install npt@natural-parallel-test

Restart, or run /reload-plugins, and the commands are there. Claude Code v2.1.x or newer. The plugin adds about 530 tokens to a session; everything else loads only when you invoke it.