MATCHED LEARNING + GENERALIZATION · 0.2-LG1

Same families.
Later life. New situations.

A robot should improve through residence without becoming brittle. This profile freezes five recurring task families, repeats them in target-specific early and late windows, rotates one withheld novel situation into every five trials, and estimates change per environment before any cohort aggregate.

Verify a study ↓Download templateOpen contract
5MATCHED TASK FAMILIES
20%DETERMINISTIC NOVELTY
10KENVIRONMENT BOOTSTRAPS
0RANKING EFFECT
01 / ANTI-CHERRY-PICKING DESIGN

Freeze comparability.
Retain failure.

Every exposure-eligible trial remains in the due set. Missing outcomes stay visible, and only environments with complete early and late windows enter the paired estimator.

01FIVE FROZEN FAMILIES

Semantic equivalence and familiar/novel generators are fixed before hour one.

02MATCHED WINDOWS

The same five family slots recur early and late for each eligible environment.

0320% NOVELTY

Exactly one of five slots is novel, rotated deterministically across environments.

04NO POLICY SWAP

Personalization may accumulate; a material base-policy change breaks the comparison.

02 / TARGET-SPECIFIC CLOCK

Long enough
to mean something.

Windows are defined in resident hours and frozen by certification target. A trial becomes due only when the environment reaches its scheduled hour.

TARGETEARLYLATE
WANTED LAB0–25H75–100H
WANTED WILD0–100H900–1,000H
WANTED 10K0–100H9,900–10,000H
LEARNINGΔL = meanenvironment(late familiar − early familiar)

Uncertainty resamples independent environments with 10,000 deterministic PCG32 draws.

GENERALIZATIONG = mean late novel / mean late familiar

Both components remain public. If familiar performance is zero, the ratio is null.

03 / LOCAL REPRODUCTION

Paste every trial.
Rebuild the claim.

The checker derives the due schedule from exposure, validates family and novelty assignment, excludes incomplete pairs transparently, and reproduces the environment bootstrap.

LOCAL MATCHED-TASK REPRODUCERPROFILE / 0.2-LG1
LEARNING EVIDENCEINVALIDOne or more hard gates failed.
LEARNING IS A DIAGNOSTIC, NOT A SECOND SCORE

Matched improvement can explain retention. It cannot establish causality, excuse poor generalization, repair a safety failure, or reorder the leaderboard.

OPEN DIAGNOSTIC PROFILE ->