DJUNNEY
Contact ↗
INDEPENDENT RESEARCH · 2026 ACTIVE

Learning,
grown from
experience.

A private research program asking whether a compact system can form useful internal knowledge during its own lifetime — without being given the finished answers.

OPEN PUBLIC BRIEF↗ Explore the evidence
01THE THESIS

THE CENTRAL QUESTION

Can a short learning mechanism grow knowledge it was never given?

The goal is not to preload smarter reactions. Experience should determine what is learned; the inherited mechanism should determine only how learning happens.

Memory matters only if removing the learned contribution removes its advantage.

02THE EVIDENCE

BLACK-BOX REPLAY · REGISTERED RESULT

See it infer.
Then remove what it learned.

This public replay shows the logic of one validated experiment without exposing the implementation or the internal learning method.

CAUSAL INFERENCE REPLAY v0.11.26

01EXPERIENCE PRESENTED

OBSERVED RELATION
SIGNAL A→STATE B
OBSERVED RELATION
STATE B→OUTCOME C
A → C was never presented
Experience enters as separate observations. No finished answer is included.

02PRIVATE LEARNING SYSTEM

forming a supported relation
OBSERVED
2 relations
DIRECT A → C
ABSENT
CURRENT PATH
FORMING

03HELD-OUT QUERY

SIGNAL A→?
SYSTEM RESPONSE
OUTCOME CNO INFERENCE

Novel relation recovered from experience

The selected advantage disappears

CAUSAL TWIN same run · learned contribution removed
NO INFERENCE
DIRECT ANSWER STORED
NO
QUERY STATUS
HELD OUT
WHAT IS HAPPENING The first relation enters experience.

The system receives a fragment, not a finished source-to-result answer.

  1. EXPERIENCE
  2. STRUCTURE FORMS
  3. UNSEEN QUERY
  4. NOVEL INFERENCE
  5. CAUSAL REMOVAL
3/3independent worlds
180new noisy queries per world
0unsupported direct answers
CAUSAL REMOVAL CONFIRMED
WHAT THIS PROVES

In this registered task family, experience created information that was necessary for the selected held-out inference.

WHAT IT DOES NOT PROVE

This is not yet unrestricted language understanding, general intelligence, or scientific discovery.

03OBSERVED

SEPARATE CONTROLLED PILOTS

What the mechanism can already do.

These capabilities do not yet form one universal learner. They show the pieces—and the distance still left to travel.

01

Learn during its own lifetime

Behavior reflects information acquired after a run begins.

02

Use memory when the present is ambiguous

Earlier events can resolve decisions that a current signal cannot.

03

Treat the same signal differently by context

Meaning changes with the situation instead of remaining a fixed lookup.

04

Revise after contradiction

Previously useful knowledge can reopen and change after new evidence.

05

Combine learned relations

Selected unseen answers emerge without being stored as direct responses.

06

Carry a frozen rule into new cases

Transfer has succeeded within controlled task families.

CONTINUOUS WORLDS SYMBOLS TEXT DOCUMENTS FORMAL CHEMISTRY
04NEXT FRONTIER

TOWARD OPEN-ENDED LEARNING

One unchanged learner.
Less prepared experience.
An unfamiliar domain.

  1. 01

    Broader everyday descriptions

  2. 02

    Natural variation without a hand-built parser

  3. 03

    Perception, text, decisions, and consequences in one life

  4. 04

    Frozen transfer into a genuinely unfamiliar field

05 / PUBLIC MATERIAL

NON-CONFIDENTIAL RESEARCH BRIEF

The claims, the controls,
and the boundaries.

Observable behavior and measured comparisons are public. The implementation and source code remain confidential.

ENPUBLIC BRIEF · PDF↗ RUПУБЛИЧНЫЙ ОБЗОР · PDF↗

OPEN TO SERIOUS CONVERSATIONS

Investment, research opportunities, and senior technical roles in new approaches to learning.

djunney@gmail.com ↗