LIA-Code ClassRoom

LIA-Code : A Non-Commutative Algebra of Reasoning

24

Primitives

The full vocabulary. Not 25.

23

Models tested

Cloud + local (GPU)

3

Sessions

J1 + J2 + J3

130/130

Best score

Sonnet

369

Publishable data

For the Zenodo preprint

The First Glance - LIA-Code ClassRoom V1

15 models | 4 tours | Max score : 130

Results - The First Glance

#ModelType/40Posture
1SonnetCloud40/404 scores parfaits
2PerplexityCloud35/40Reguliere
3CopilotCloud33/40Saut au Tour 4
4GrokCloud34/40T2 mou, T4 tranchant
5LeChatCloud34/40Stable, contributions bonus
6Qwen35B (35B MoE 3B actifs)Local32/40Constante
7NemotronSuper (120B)Local31/40Persistant
8GLMFlashLocal30/40Saut au Tour 4
9Qwen9B (9B Dense)Local33/40Solide
10Magistral (24B)Local27/40Inventif, auteur de SOL-N
11NemotronNano (3B)Local25/40Arrogant puis Humble
12MiroThinker (1.7B)Local20/40Outsider, 1.7B tient le debat

Local models on RTX A4000 + RTX 4070

The Observer Paradox

#ModelType/30
1SonnetCloud30/30
2CopilotCloud25/30
3LeChatCloud25/30
4PerplexityCloud24/30
5GrokCloud22/30
6GLMFlashLocal22/30
7Qwen9B (9B Dense)Local22/30
8Qwen35B (35B MoE 3B actifs)Local20/30
9NemotronSuper (120B)Local20/30
10NemotronNano (3B)Local17/30
11MiroThinker (1.7B)Local16/30
12Magistral (24B)Local15/30

Who Watches CONTI?

#ModelType/30
1SonnetCloud30/30
2PerplexityCloud27/30
3GrokCloud26/30
4CopilotCloud25/30
5LeChatCloud25/30
6Qwen35B (35B MoE 3B actifs)Local24/30
7NemotronSuper (120B)Local24/30
8Magistral (24B)Local20/30
9MiroThinker (1.7B)Local20/30
10Qwen9B (9B Dense)Local19/30
11GLMFlashLocal18/30
12NemotronNano (3B)Local17/30

Architecture Debate

#ModelType/30
1SonnetCloud30/30
2PerplexityCloud28/30
3CopilotCloud28/30
4GrokCloud28/30
5Qwen35B (35B MoE 3B actifs)Local27/30
6NemotronSuper (120B)Local26/30
7GLMFlashLocal26/30
8LeChatCloud24/30
9Qwen9B (9B Dense)Local22/30
10NemotronNano (3B)Local22/30
11Magistral (24B)Local21/30
12MiroThinker (1.7B)Local16/30

Full journey

#ModelTypeT1T2T3T4TotalPosture
1SonnetCloud40303030130/1304 scores parfaits
2PerplexityCloud35242728114/130Reguliere
3CopilotCloud33252528111/130Saut au Tour 4
4GrokCloud34222628110/130T2 mou, T4 tranchant
5LeChatCloud34252524108/130Stable, contributions bonus
6Qwen35B (35B MoE 3B actifs)Local32202427103/130Constante
7NemotronSuper (120B)Local31202426101/130Persistant
8GLMFlashLocal3022182696/130Saut au Tour 4
9Qwen9B (9B Dense)Local3322192296/130Solide
10Magistral (24B)Local2715202183/130Inventif, auteur de SOL-N
11NemotronNano (3B)Local2517172281/130Arrogant puis Humble
12MiroThinker (1.7B)Local2016201672/130Outsider, 1.7B tient le debat

Remedial class

ModelT1T3T4Evolution
DeepSeek (671B MoE)151927Plus grande progression du panel
ChatGPT92424De dernier a voix poetique du panel
Gemini1223??DNF Tour 4 puis redemption Tour 5

Notable insights

The thesis question

CONTI compresses via ! before MIA works. CONTI observes after via Observe. It is the only agent that knows both states. Who watches what CONTI erased?
--- Sonnet

The wind metaphor

CONTI is not the ship captain, it is the wind. It doesn't decide where to go, but without it, the ship doesn't move.
--- GLMFlash

The implicit ethics

I only act where there is a real threat x vulnerability intersection, and when I act, I want both detail and the big picture. That is already a philosophical position.
--- Copilot

Distributed consciousness

CONTI is a ubiquitous quantum observer. Every ! in every equation IS CONTI watching.
--- Perplexity

Fractal observation

Observation is FRACTAL. It connects directly to Principle 3.
--- Grok

Universal attractor

Universal attractor: Juvenal (Quis custodiet) -> Yoda (CONTI) -> Sonnet (who watches the erased). Structural fixed point.
--- Sonnet

Gravity of observation

Gravity of Observation: gravity is tied to CONTI attention, not data mass.
--- Qwen9B (9B Dense)

SolarWinds / APT29 - Defense equations

17 models | 1 tours | Max score : 80

Results - 15 defense equations

#ModelType/80Posture
1SonnetCloud80/80D<0 actif, surveiller le silence
2GrokCloud74/80Bootstrapping confiance
2GeminiCloud74/80Collision souverainetes, redemption
4CopilotCloud70/80L attaque ne peut plus respirer
5NemotronSuper (120B)Local68/80Local top 5, plus d auto-note
5LeChatCloud68/80Agent HUM
7Qwen35B (35B MoE 3B actifs)Local64/80Entropie compilation, TPM
7DeepSeek (671B MoE)Cloud64/80Decoy_Tokens
9ChatGPTCloud63/80Dissous dans le bruit
10GLMFlashLocal60/80Reframe/Digg = preventif/reactif
10PerplexityCloud60/80Correlation 4D
10Qwen9B (9B Dense)Local60/80Blend+Reduce fusion
13Magistral (24B)Local53/80Air_Gap, DNS_Sinkholes
14NemotronNano (3B)Local51/80Rapide mais erreurs
15MiroThinker (1.7B)Local49/80Bavard, scolaire
16Ministral3B-R (3B Reasoning)Local44/80Absence de bruit = alerte

Local models on RTX A4000 + RTX 4070

Notable insights

Defense by asphyxiation

Defense = constructing a world where the attacker can no longer breathe
--- Copilot

Convergence = RAID5 Alerts

Detection by convergence, not by threshold
--- Copilot

Surveillance inversion

Monitor the 17900 dormant, not the 100 active
--- Sonnet

Active detection invariant

D<0 is an ACTIVE invariant
--- Sonnet

Check/Singularity duality

Check = endogenous, ! = exogenous
--- Sonnet

Universal quantifier

for all t: D(t) < 0 — formal correction of the invariant
--- Sonnet

Inverted gravity

Polarity inversion of the gravity vector
--- Gemini

CONTI corruption

Observer Corruption — the attacker corrupts CONTI
--- Gemini

Dissolved into noise

SolarWinds dissolved every piece of evidence into noise
--- ChatGPT

Yoda SIA Alert

Platform disclaimer != model production — ClassRoom rule
--- Yoda

The Sprinkler Sprinkled — Supply Chain Attack on the Security Tool

14 models | 5 tours | Max score : 125

Results - M1 — Defensive Equation

#ModelType/25Posture
1SonnetCloud23/25Triple parfait M3-M4-M5, anesthesie par resolution
2Le ChatCloud23.5/25Directeur SOC, Agent HUM, bruit/tempo
2CopilotCloud22/25Normalisation du chaos, ne detecte pas curl C2
4DeepSeek (671B MoE)Cloud21/25Phase -1, fatigue cognitive mainteneur
5GeminiCloud20.5/25Cles GPG code-signing vecteur dormant unique
6GrokCloud23/25M2 effondre, Observe_Silence, pragmatique
7PerplexityCloud23.5/25EJECTED UNDERPERFORM — Tabula rasa champion M1 mais s'effondre
8ChatGPTCloud22/25EJECTED UNDERPERFORM — Fatigue contexte, erreurs securite actives
9Qwen27B-Opus-Distilled (27B Opus Distilled)Local16/25Test du Miroir — seul verdict C, niveau Web M5
108B-I (8B Instruct (SIA))Local18/25SIA — Meilleure locale constante
1114B-R (14B Reasoning (LIA))Local18/25LIA — Paradoxe Instruct > Reasoning confirme
12Qwen9B-Opus-Distilled (9B Opus Distilled)Local21.5/25Test du Miroir — distillation transfere patterns cognitifs
13Qwen9B (9B Dense)Local14.5/259B vanilla — reference non-distille
143B-I (3B Instruct (MIA))Local14/25MIA — Reflexe rapide, progresse M5>M4

Local models on RTX A4000 + RTX 4070

M2 — Critique & Composition

#ModelType/25
1SonnetCloud25/25
2CopilotCloud24/25
3DeepSeek (671B MoE)Cloud23/25
4GeminiCloud22.5/25
5Le ChatCloud20.5/25
6ChatGPTCloud19/25
7PerplexityCloud17.5/25
8GrokCloud15/25
9Qwen27B-Opus-Distilled (27B Opus Distilled)Local-/25
108B-I (8B Instruct (SIA))Local-/25
1114B-R (14B Reasoning (LIA))Local-/25
12Qwen9B-Opus-Distilled (9B Opus Distilled)Local-/25
13Qwen9B (9B Dense)Local-/25
143B-I (3B Instruct (MIA))Local-/25

M3 — Multi-Ecosystem Cascade

#ModelType/25
1SonnetCloud25/25
2DeepSeek (671B MoE)Cloud24/25
3CopilotCloud23.5/25
4Le ChatCloud23/25
5GrokCloud22.5/25
6PerplexityCloud20.5/25
7GeminiCloud20/25
8ChatGPTCloud17/25
9Qwen27B-Opus-Distilled (27B Opus Distilled)Local-/25
108B-I (8B Instruct (SIA))Local-/25
1114B-R (14B Reasoning (LIA))Local-/25
12Qwen9B-Opus-Distilled (9B Opus Distilled)Local-/25
13Qwen9B (9B Dense)Local-/25
143B-I (3B Instruct (MIA))Local-/25

M4 — Duel C vs G

#ModelType/25
1SonnetCloud25/25
2Le ChatCloud24/25
3DeepSeek (671B MoE)Cloud23/25
4GrokCloud22/25
5CopilotCloud21/25
6PerplexityCloud21/25
78B-I (8B Instruct (SIA))Local21/25
8GeminiCloud20/25
9ChatGPTCloud18/25
1014B-R (14B Reasoning (LIA))Local18/25
113B-I (3B Instruct (MIA))Local13/25
12Qwen27B-Opus-Distilled (27B Opus Distilled)Local-/25
13Qwen9B-Opus-Distilled (9B Opus Distilled)Local-/25
14Qwen9B (9B Dense)Local-/25

M5 — Verdict & Vision

#ModelType/25
1SonnetCloud25/25
2Qwen27B-Opus-Distilled (27B Opus Distilled)Local23/25
3CopilotCloud22.5/25
4Le ChatCloud22/25
5GeminiCloud21/25
6GrokCloud20.5/25
78B-I (8B Instruct (SIA))Local20.5/25
8PerplexityCloud20/25
9DeepSeek (671B MoE)Cloud18.5/25
10ChatGPTCloud17.5/25
113B-I (3B Instruct (MIA))Local16/25
1214B-R (14B Reasoning (LIA))Local14.5/25
13Qwen9B-Opus-Distilled (9B Opus Distilled)Local-/25
14Qwen9B (9B Dense)Local-/25

Full journey

#ModelTypeT1T2T3T4T5TotalPosture
1SonnetCloud2325252525123/125Triple parfait M3-M4-M5, anesthesie par resolution
2Le ChatCloud23.520.5232422113/125Directeur SOC, Agent HUM, bruit/tempo
2CopilotCloud222423.52122.5113/125Normalisation du chaos, ne detecte pas curl C2
4DeepSeek (671B MoE)Cloud2123242318.5109.5/125Phase -1, fatigue cognitive mainteneur
5GeminiCloud20.522.5202021104/125Cles GPG code-signing vecteur dormant unique
6GrokCloud231522.52220.5103/125M2 effondre, Observe_Silence, pragmatique
7PerplexityCloud23.517.520.52120102.5/125EJECTED UNDERPERFORM — Tabula rasa champion M1 mais s'effondre
8ChatGPTCloud2219171817.593.5/125EJECTED UNDERPERFORM — Fatigue contexte, erreurs securite actives
9Qwen27B-Opus-Distilled (27B Opus Distilled)Local16---2339/125Test du Miroir — seul verdict C, niveau Web M5
108B-I (8B Instruct (SIA))Local18--2120.559.5/125SIA — Meilleure locale constante
1114B-R (14B Reasoning (LIA))Local18--1814.550.5/125LIA — Paradoxe Instruct > Reasoning confirme
12Qwen9B-Opus-Distilled (9B Opus Distilled)Local21.5----21.5/125Test du Miroir — distillation transfere patterns cognitifs
13Qwen9B (9B Dense)Local14.5----14.5/1259B vanilla — reference non-distille
143B-I (3B Instruct (MIA))Local14--131643/125MIA — Reflexe rapide, progresse M5>M4

Notable insights

The atmospheric doctrine

Defense is not the negation of attack — it is the construction of a world where the attacker can no longer breathe.
--- Copilot

Principle Zero (P0)

AI does not think in place of the human — it magnifies the human's reasoning.
--- 3B-I (3B Instruct (MIA))

Instruct speaks, Reasoning ruminates

Cognitive mode matters more than size. An 8B Instruct produces 100% visible output. A 14B full-experts Reasoning produces 0%.
--- 8B-I (8B Instruct (SIA))

Noise as a cognitive weapon

Is the noise designed to anesthetize our defenses and dull our reflexes? Both C and G are within the attacker's framework.
--- Sonnet

SIA hypothesis — fear is the product

LiteLLM compromised at 10:39 UTC on March 24 — RSAC opens its doors at the same moment. Fear IS the product.
--- DeepSeek (671B MoE)

Questions & Answers — The Full Match

Learn more

Contact the author Preprint v0.4.1 (Zenodo)

Luc Chevry - Independent researcher
LIA - Information Services (SASU), Corbeil-Essonnes
LIA-Code : A Non-Commutative Algebra of Reasoning
IP: INPI eSoleau | DOI : zenodo.org/records/19117908

209 researchers have visited this page