🧬 @smartplant/migration

Inheritance between plants. A mature plant leaves what it learned to a new plant of the same species.

Federated learning pools anonymous statistics from many homes into a species profile. This is the other shape of it: a directed transfer, one known source to one known receiver, carrying the context needed to check itself on arrival.

npm install @smartplant/migration

import { createPlant } from 'smartplant'
import migration from '@smartplant/migration'

await mature.use( migration )
const { bundle, shipped, withheld } = await mature.plugin( 'migration' ).bequeath()

await seedling.use( migration, { inherit : bundle } )
seedling.plugin( 'migration' ).status()

Evidence is not transferability

The obvious way to rank a learned policy for export is by how much evidence stands behind it. That ranking is close to backwards.

A policy with four hundred outcomes, every one recorded on the same windowsill, has enormous evidential weight and almost no transferable content — what it encodes is that windowsill. A policy tried thirty times across cold mornings and warm afternoons, damp substrate and dry, has less evidence and far more of what you want: a regularity that survived a change of conditions.

Context diversity is therefore the gate, and no amount of evidence buys past it:

✓ shift_toward_window
    Transferable (0.931): held across 4 distinct situations over 32 outcomes.

✗ water_at_low_soil
    Not transferable: learned in a single set of conditions — this describes
    that spot, not the plant.

The withheld policy had twelve times more evidence than the one that shipped.

Inherited pathology is caught on arrival

The bundle carries the conditions its priors were learned under, so the receiver can compare them against where it has landed.

A plant that learned "water at 15% soil" did so in a pot that drained badly. Ship that to a pot that holds water and you have shipped a fix for a problem it does not have. Waiting for the prior to dilute away is not enough — the plant suffers for the whole dilution, which is when it is most fragile.

await seedling.plugin( 'migration' ).wouldSuit( bundle )
// { compatible: false,
//   verdict: 'The two spots differ where it matters: soil (40–50 vs 63–73)…' }

wouldSuit() answers before anything is changed.

The new body always wins

An inherited prior is a starting guess with a finite weight. It is advisory until the plant has outcomes of its own, and fades by Bayesian shrinkage as those accumulate.

const mig = seedling.plugin( 'migration' )
mig.outcome( 'move_to_light', 0 )   // this plant found it useless
// → { expected, weight, why }  — weight falls with every local outcome

API

| Method | What it does | | --- | --- | | bequeath( opts ) | Package this plant's inheritance. Returns {bundle, shipped, withheld, summary} | | receive( bundle, opts ) | Graft one on. Returns {compatibility, admitted, held, report} | | wouldSuit( bundle ) | Would it fit here? Changes nothing | | status() | What is inherited and how much still applies | | outcome( action, reward ) | Record a local result so the prior dilutes |

Install with { inherit: bundle } to make a plant the heir at setup time; it raises migration:inherited.

What travels: comfort ranges the source thrived in, policies that passed the gate, care cadence, the electrome baseline shape, and the origin conditions.

What does not: the name, current wellbeing, raw readings, notes, timestamps of when a home was occupied, open wounds, keys. The new plant is born with an inheritance, not a borrowed biography.

License

MIT © Alejo Malia · PigeonPosse