Automatic AI morphology

INSACERMO Morphology V28.1 - Reading AI training logs

Moving is not becoming.

V28 is an automatic morphology detector for artificial intelligence model training logs.

Read the scientific document View the software on Zenodo View code on GitHub Download V28
Scope V28

What is V28 for

V28 is currently dedicated to artificial intelligence model training logs.

It analyzes CSV files containing recognized metrics such as:

train_loss val_loss eval_loss accuracy epoch step

It tries to determine which morphology the training is currently concentrating into:

FAVORABLE FAVORABLE_GAP_WARNING OVERFIT_DRIFT PLATEAU DEGRADATION MIXED UNKNOWN_NOVEL

INSACERMO is a broader framework than V28.
The scientific core of INSACERMO can be applied to different systems: artificial intelligence, EEG, climate, time series, continuous systems or automata.
V28 is the first public automatic implementation of this core, specialized in AI training logs.

V28

Read la morphologie actuelle

V28 observes training metrics, identifies the dominant regime, measures hidden closure and produces a first alert.

MemGuard

Decide whether to intervene

MemGuard evaluates whether it is better to act, monitor or abstain, then verifies that an intervention does not degrade the system further.

Specialized CSV tools

Deepen the analysis

Specialized CSV tools use domain-specific protocols to produce detailed analyses, comparisons and reports.

V28 observes and classifies. MemGuard protects and decides. Specialized tools deepen the analysis.

V28 is not currently a universal analyzer for all CSV files. EEG, climate, biological or physical data require adapters and morphological grammars specific to their domain.

Future section

Test V28 with an AI training log

V28 is currently specialized in AI training logs containing recognized metrics such as train_loss, val_loss, eval_loss, accuracy, epoch or step. It is not a universal analyzer for all CSV files.

Your file is analyzed locally in your browser. It is neither uploaded nor stored by INSACERMO.

Fast web demonstration. The full scientific analysis with more controls is available in the public Python software.

Drop a CSV here

Or choose a recognized AI training log. No real file is embedded in this page.

No file selected.

Log chart

Normalized curves for local visual reading.

Result principal

Waiting

Chronologie

Touch a segment to read its period, state and reason.

Simple indicators

Technical details 

              

            

          

        

      

    


    

      

        

          Automatic reading
          
What V28 detects.

          
From a recognized CSV file, it automatically detects training and validation metrics, their direction of improvement, the dominant morphology, hidden closure, locking level, regime transitions, classification confidence, an alert level and behaviors not covered by the current grammar.

        

        

          Hidden closure
          
The public formula.

          
D_hidden = 1 - H_morph / (H_micro + epsilon)

          
Hidden closure measures how strongly the dynamics concentrate into a reduced number of becomings. The dominant morphology indicates the direction in which this dynamic is locking.

        

      

    


    

      

        

          V28 grammar
          
Main morphologies.

          

            FAVORABLE
            FAVORABLE_GAP_WARNING
            OVERFIT_DRIFT
            PLATEAU
            DEGRADATION
            IMPROVING_TRAIN_ONLY
            WORSENING_TRAIN_ONLY
            MIXED
            UNKNOWN_NOVEL
          

        

      

    


    

      

        

          Command
          
Local use.

          
python insacermo_morphology_auto_v28.py mon_log.csv

          
The public detector runs outside the site as local software. The site publishes links, citation and public archives without exposing sensitive real logs.

        

        

          Public downloads
          
Verify and reproduce.

          
The files below come from the public V28 pack. Private archives, real BERT, Vision or LLaMA logs, master pack and precedence archive are not published.

          

            Download V28
            Download the replication pack
          

        

      

    


    

      

        

          Citation
          
Licenses and references.

          
Author: Benjamin Lenoir

          Program: INSACERMO

          ORCID : 0009-0006-1201-7127

          Document : CC BY 4.0

          Code : Apache-2.0

          
Document citation:

          Benjamin Lenoir (2026). INSACERMO - Moving is not becoming: an experimental measure of the real freedom of systems. Zenodo.

          https://doi.org/10.5281/zenodo.20851280

          
Software citation:

          Benjamin Lenoir (2026). INSACERMO Morphology Auto Detector, Version 28.0.0. Zenodo.

          https://doi.org/10.5281/zenodo.20852518