MemGuard / Overfitting
Quand apprendre commence à mémoriser
Reading the progressive closure of learning: the model appears to keep learning, but begins to contract toward memorization.
Maquette de validation
Foundations INSACERMO
INSACERMO does not only look at what happens. It tries to read the state that makes what happens possible.
Before the visible rupture, a system may already begin to change regime.
Le noyau
Mémoire — cohérence — état complet
INSACERMO starts from a simple idea: an isolated value is not always enough to understand a system.
What matters is the state that carries that value: its memory, coherence, tensions, constraints and still-open possibilities.
A data point is therefore not just a number. An image is not just an image. A text is not just a sequence of words. An AI system is not just an output.
Each medium carries a regime. INSACERMO tries to read that regime.
INSACERMO does not first look for an anomaly. It looks for the state that makes the anomaly possible.
Read before concluding
Before deciding, read the state.
- Ce qui tient.
- What changes.
- Ce qui se ferme.
- Ce qui reste possible.
Mémoire et futur possible
La mémoire ne garde pas seulement le passé.
Dans INSACERMO, la mémoire agit aussi sur ce qui reste possible.
Quand la mémoire augmente, un système peut se stabiliser. Mais il peut aussi se fermer, garder sa cohérence ou perdre sa capacité de variation.
Le futur n’est pas seulement ce qui arrive. C’est aussi ce qui reste possible compte tenu de l’état actuel du système.
Bascule
Read the regime before the shift
Un système ne bascule pas always d’un coup.
Before the peak, before the failure, before the break, before dead repetition, there may be a more discreet transformation.
Le signal se tend. La cohérence se déplace. The regime change.
INSACERMO looks for the moment when the present begins to stop behaving as before.
Before the visible peak, INSACERMO cherche la bascule invisible.
Contraction / openness
Ce qui se ferme, ce qui reste ouvert
Un système peut être cohérent sans être ouvert.
Il peut tenir, mais se contracter. Il peut répéter, mais ne plus transformer. Il peut produire des réponses, mais fermer trop vite ses futurs possibles.
INSACERMO observes this tension between closure and openness: what closes, what holds, what repeats, what remains capable of variation.
Instruments
La lecture devient manipulable.
The demonstrators are the practical instruments of the framework: signal, image, text, learning and regime change.
To test them without weighing down this page, the next step is organized on the Instruments page.
Traces visuelles
Cartes et traces expérimentales
These visuals are not decorative. They show how INSACERMO makes regimes, tensions, closures or shifts visible across different media.
They do not replace the interactive tools: they provide a visual memory of experiments and working outputs.
AI and learning
Closure, drift, loss of openness
INSACERMO reads signs of closure, drift or loss of openness in a training dynamic.
Image |K-G|
Image comme champ de cohérence
Morphological reading: the image is not just a set of pixels, but a field of coherence, rupture and local tension.
Sandbox image
Local ruptures and morphology
The image sandbox makes this reading manipulable: load an image, observe hot zones, compare texture, rupture and visual regime.
AI learning
Variation du régime d’entraînement
Experimental trace around AI dynamics: the curve is not only a score, it indicates a shift in the learning regime.
Cyber / structures informationnelles
Complete state and openness
Explore coherences, ruptures and tensions in digital or informational structures, without reducing openness to a simple key.
Résultats / limites
See what holds, see what remains partial
INSACERMO must also show its zones of validity and its limits: a partial result remains named as partial.
Cohérence / classes
Cartographier un régime
A graphical output shows how a structure is distributed, where it holds and where it becomes fragile.
Branches
Branches de recherche
INSACERMO n’est pas limité à un seul type de donnée.
The same question can be asked across several media: what state is the system in, and is it beginning to change
Images et morphologie
Read local ruptures, visual tensions and regime changes inside an image.
Signals and time series
Observe transitions, peaks, regimes and shifts in time-series data.
AI and learning
Read the signs of closure, overlearning, drift or loss of openness in a training dynamic.
MemGuard / MemAct
Explore how learning can be monitored, slowed, stopped or reopened before it closes into memorization.
Language
Observe whether a dialogue closes into dead repetition or keeps structured openness: reprise, variation, coherence, progression.
Cyber et structures informationnelles
Explore ruptures, coherences and tensions in digital or informational structures.
Environnement
Test INSACERMO on real signals: rain, pollution, space weather, public series and regime ruptures.
Attention collective
Observe how public attention can enter a regime before certain visible peaks, for example in Wikipedia series.
Plusieurs supports
Une même lecture, plusieurs supports
- Does the system hold
- Is it tightening
- Is it closing
- Is it beginning to shift
- Does it still keep an opening
A time series, image, CSV, model, conversation or signal can all carry a state. The media change. The question remains the same: read the regime.
Results and evidence
Résultats forts, résultats partiels
INSACERMO does not try to validate everything everywhere. Some signals produce strong results. Others remain partial. Some domains do not respond clearly.
Cette distinction est essentielle : identifier les cas où une structure devient lisible, et ceux où le signal reste insuffisant.
Limites
Une lecture, pas une vérité automatique
INSACERMO ne remplace pas l’expertise humaine. Il ajoute une couche de lecture.
An alert is not a certainty. A score is not truth. A strong result must remain reproducible. A partial result must remain named as such.
INSACERMO doit rester testable, critiquable et améliorable.