Du test simple au STL-10 64x64

A signature that learns to abstain.

The public V67F story describes a progressive climb: small tests, vision, RGB, cross-architecture, puis STL-10 64x64 resized.

The thread stays the same: repair less, break less, keep gain when geometry allows it.

View short result View the progression
Idea

Before correcting, verify whether the action is supported.

INSACERMO explores a decision layer: if the geometric signal is too low, the system abstains. The goal is not to correct everything, but to reduce actions qui cassent.

Sobriety

Une piste contre le calcul aveugle.

If an intervention can be refused before it costs computation, there is potential to save GPU time, energy and trial cycles. That potential must remain measured, verifiable and replicated.

Progression publique.

Timeline publique INSACERMO vers V67F
From simple tests to V67F: a controlled progression, not a universal promise.
1

Petits tests

Minimal repair starts with controlled cases: measure before acting.

2

Light vision

Digits, Fashion-MNIST, CIFAR-10 : verify whether the signal survives beyond a trivial toy case.

3

Cross-architecture

Test whether a minimal decision can cross different architectures.

4

SVHN / STL-10

Move toward RGB and natural images, with more demanding controls.

5

64x64 resized

Ne pas dire STL-10 natif 96x96 : le protocole public V67F travaille en 64x64 resized.

6

V67F

HARD keeps 0 harm and removes the dangerous action that controls let through.

Public summary from V64A to V67F
Public summary from V64A to V67F.
Conclusion prudente

A serious signature, not yet a total law.

Ce qui fascine ici, c'est la forme de refus : HARD agit moins brutalement, mais laisse moins passer d'actions dangereuses. C'est exactement l'esprit INSACERMO : retrouver de la marge against black boxes, without promising more than the protocol shows.

Access public files