E1P in AI (II)

Two AI systems can revise and retain answers in exactly equal proportions, yet the same feedback can amplify a bias in one system and damp it in the other. This paper shows how to predict which response will occur in a binary energy-based model. The predictor is compositional dispersion, a quantity from Energetic First Principles (E1P) that measures how the balance between revision and retention varies across individual inputs. For a trained visual reader, calibration images gave a critical feedback gain of 1.939 before any answers were fed back. Held-out images gave 1.949, and controlled perturbations confirmed the predicted shift from damping to amplification. A second result establishes a limit on diagnosing model collapse from outputs alone. Mirroring the inputs and reversing the reader’s interpretation produced identical answer sequences with accuracies of 82.36% and 17.64%. Every repetition or diversity score therefore agreed across the two runs. Testing whether a model remains grounded requires independently changing what the evidence supports and scoring whether the answer follows. A five-version route model makes E1P’s Connective and Active dyads testable at the level of meaning. At zero temperature, the selected route preserved every direct link held in memory. At unit temperature, four ways to integrate a new stop together received 59.5% of the probability. Changing one remembered requirement from “immediately precedes” to “precedes” made an integrated route the unique zero-temperature answer. This yields a concrete test for language models: hold the previous answer fixed, change one relation in memory, and measure whether the model selects the newly admissible answer.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-18
DOI
https://doi.org/10.5281/zenodo.24339759
Primary Topic
Explainable Artificial Intelligence (XAI)
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

E1P in AI (II)

Resonant Institute
5 links
Zenodo (CERN European Organization for Nuclear Research)
Explainable Artificial Intelligence (XAI)
preprint

E1P in AI (II)

Resonant Institute
preprint en

Abstract

Two AI systems can revise and retain answers in exactly equal proportions, yet the same feedback can amplify a bias in one system and damp it in the other. This paper shows how to predict which response will occur in a binary energy-based model. The predictor is compositional dispersion, a quantity from Energetic First Principles (E1P) that measures how the balance between revision and retention varies across individual inputs. For a trained visual reader, calibration images gave a critical feedback gain of 1.939 before any answers were fed back. Held-out images gave 1.949, and controlled perturbations confirmed the predicted shift from damping to amplification. A second result establishes a limit on diagnosing model collapse from outputs alone. Mirroring the inputs and reversing the reader’s interpretation produced identical answer sequences with accuracies of 82.36% and 17.64%. Every repetition or diversity score therefore agreed across the two runs. Testing whether a model remains grounded requires independently changing what the evidence supports and scoring whether the answer follows. A five-version route model makes E1P’s Connective and Active dyads testable at the level of meaning. At zero temperature, the selected route preserved every direct link held in memory. At unit temperature, four ways to integrate a new stop together received 59.5% of the probability. Changing one remembered requirement from “immediately precedes” to “precedes” made an integrated route the unique zero-temperature answer. This yields a concrete test for language models: hold the previous answer fixed, change one relation in memory, and measure whether the model selects the newly admissible answer.

Zenodo (CERN European Organization for Nuclear Research)
Explainable Artificial Intelligence (XAI)
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon