SLATE · Illustrative preview

SLATE: How it works

Spectral-Lyapunov Annealed Training Estimator

A narrated journey through a SLATE training run, from selecting a training profile to a complete stability report with standards relevance and an honest at-risk contrast. It plays like a video, loops automatically, and you can pause, scrub, or use the arrow keys at any time.

Illustrative walkthrough, not the production tool

narrated walkthrough

Scene 1 of 7 · Select Training Profile

Safety-critical AI

Frontier pretraining

Long-context

Regulated deployment

Research baseline

The selected profile determines which stability guarantees apply to the run.

SLATE configuration for this run

Base method

DeepSeek NSA (open source)

INHERITED

Compression operator

Spectrally-constrained, non-amplifying

VALIDATED

Training schedule

Lyapunov-annealed, bifurcation-guarded

DEMONSTRATED

Token selection

Fisher-weighted, utility-based

PROTOTYPE

SLATE wraps your existing training loop. Each guarantee carries its current validation status, so nothing is presented as more settled than it is.

Guarantees enforced

SLATE enforces its constraints throughout training, not only as a check at the end.

Stability Report

SLATE · Illustrative example

STABLE

Stability profile

STRONG

Evidence confidence

HIGH

Error amplification

compression is provably non-amplifying

PREVENTED

Information integrity

no information created by compression

PRESERVED

Guarantee coverage

one of three innovations fully validated

PARTIAL

Validation scale

production-scale validation underway

EARLY

Guarantee ratings

Spectral compression

VALIDATED

Information monotonicity

VALIDATED

Lyapunov schedule

DEMONSTRATED

Fisher selection

PROTOTYPE

Key insight

Strongest guarantee: the compression step cannot amplify errors. Largest opportunity: proving advantage at production scale.

Every run reports what is proven, what is demonstrated, and what is still a prototype.

Standards and assurance relevance, linked to the STABLE verdict

FrameworkFocusRelevance

ISO/IEC 42001:2023

AI management systems

Documented, provable training controls support an AI management system.

NIST AI RMF 1.0

Measure and Manage

Provable stability supports the Measure and Manage functions.

EU AI Act

Article 15 robustness

Training-time robustness supports Article 15 expectations for high-risk AI.

ISO/IEC TR 24029

Neural network robustness

Supports structured assessment of neural network robustness.

UK AI principles

Safety and robustness

Aligns with the safety and robustness principle.

Provable training stability supports the assurance frameworks organisations are measured against.

Stability Report

Research baseline, spectral constraint disabled

UNSTABLE

Stability profile

WEAK

Error amplification

compression can amplify errors without bound

NOT PREVENTED

Guarantee ratings

Spectral compression

DISABLED

Information monotonicity

NOT GUARANTEED

Lyapunov schedule

N/A

Fisher selection

N/A

Key insight

Without the spectral constraint, nothing bounds error growth through the compression step. SLATE reports this plainly rather than hiding it.

ISO/IEC 42001

without provable constraints, training-time controls cannot be evidenced.

SLATE shows the difference its guarantees make, and stays honest when they are switched off.

Provable stability. Not empirical hope.

Guarantees you can inspect, run, and cite.

Explore SLATE on GitHub Read the paper

Safety-critical AI

Frontier pretraining

Long-context

Regulated deployment

Research baseline

The selected profile determines which stability guarantees apply to the run.

Space toggles play/pause · Arrow keys move between scenes

Back to SLATE