Snath AI · Open Source Research Initiative

We're not a normal company.
We publish the proofs.

Most companies ship features and hide the research. We ship the research. Every claim is a preprint with a DOI, runnable code, and an honest account of what we haven't proven yet. Six papers, one architecture — annotation-free continual learning without catastrophic forgetting, traced from a single equation to a robot getting back up on ice.

D = ‖softmax(zA) − softmax(zB)‖1 / √C
One divergence score. It detects the hard case, labels the curriculum, survives the encoder swap, and verifies the fix.
6 published · permanent DOIs Annotation-free · no forgetting by design Apache 2.0 code · CC BY 4.0 papers HMAC-audited, seed-published · independent

One Architecture, Six Movements

Each paper opens the gap the next one closes.

  1. 1
    Detect
    DAS
  2. 2
    Universalise
    UCR
  3. 3
    Learn
    LTL
  4. 4
    Remember
    EIM
  5. 5
    Embody
    PAV
  6. 6
    Persist
    PERSIST

Detect the disagreement · universalise it across any model · learn from it without labels · remember it across encoder upgrades · embody it in a deployed robot · persist until the problem is resolved.

The Papers

Six preprints. Every claim falsifiable.

01 DAS Detect

Divergence Is Not Noise

Multi-Stream Routing Without Modal Fusion and the Safety–Learning Equivalence

Read paper
doi:10.5281/zenodo.20278781 Published · v6

Don't fuse perception streams — route on their disagreement. The invariants that make routing safe are the same ones that turn disagreement into a self-labelling training curriculum.

  • 3.11× divergence lift on contradictory X-ray / report pairs · routing AUROC 0.87
  • 6.5× higher replan rate on real contradictions · Fisher exact p = 3.1×10⁻¹²
  • Routing 0.717 vs. fusion 0.508 on MS-COCO, where global fusion sits at chance
Theorem 1 · Safety–Learning Equivalence
02 UCR Universalise

Universal Cognitive Routing

A Forward-Compatible Architecture for Heterogeneous AI Systems

Read paper
doi:10.5281/zenodo.20278775 Published

Any model — LLM, JEPA, SSM, GNN, diffusion — that implements one abstract interface becomes a first-class, equally-routable node. The routing spine never branches on model type, so architectures that don't exist yet are already supported.

  • A ten-ABC cognitive contract pinned down by 33 named invariants
  • Forward-compatibility verified by construction: zero model_type branches in the routing layer
  • 8 executable instantiations across 7 verticals — finance to power-grid to biomedical
Design Property 1 · Interface Sufficiency
03 LTL Learn

The Lár Training Loop

Routing Flags as Gradient Signals

Read paper
doi:10.5281/zenodo.20581128 Published · v2

A routing flag — the moment two independent streams disagree — is a free, automatically-labelled training example. Act on those flags and the system learns faster, with no human annotation.

  • Self-curated curriculum beats random by +0.089 AUROC across 5 seeds · paired p < 0.001
  • Annotation-free difficulty oracle reaches AUROC 0.736 ± 0.008
  • 6.5× faster policy learning from routing priors in robotics
Honest scope Honest negative: the pre-registered isotropy target (ρ > 1.15) was not met — the gain is bounded by the encoder's baseline isotropy. We report it anyway.
04 EIM Remember

The Encoder Is Not the Memory

World-Grounded Difficulty Representations for Encoder-Invariant Continual Learning

Read paper
doi:10.5281/zenodo.20583318 Published · v1

What's hard is a fact about the world, not about your current encoder. Store the hard cases as raw pairs and the memory survives a model upgrade — swap the encoder, keep the knowledge.

  • Difficulty identity is stable across encoders · Spearman ρ = 0.672
  • A canonical projection head recovers 90% after a ViT-B/32 → ViT-L/14 upgrade (direct retrain: 104%)
Honest scope Honest negative: image alone can't predict difficulty (AUROC 0.560). You must see both modalities — reactive routing is a necessity, not a design choice.
Invariant V7 · Difficulty Invariance
05 PAV Embody

Physics Assumption Violations

Label-Free Detection via Concept-Space Routing in Deployed Robotic Systems

Read paper
doi:10.5281/zenodo.20682615 Published · v1

When the floor turns to ice, a robot's live embedding drifts away from its reference. That drift is detectable from geometry alone — no labels, no rules, no reward signal.

  • 95% violation detection the moment friction changes · ΔD = +0.50
  • 83% commit on normal terrain, 95% replan on ice (CLS-GRU encoder)
  • 65% divergence reduction after consolidating just 19 unlabelled events
Honest scope Scope, stated up front: a single simulation, a single seed. A proof of mechanism, not a benchmark.
06 PERSIST Persist

PERSIST

Proprioceptive Error Resolution with Scope-bounded Invariant Signal Tracking

Read paper
doi:10.5281/zenodo.20820042 Published · Zenodo · June 2026

Detecting a violation isn't adapting — it's logging. The signal that detected the problem is the same signal that verifies whether the fix worked. One stream, two jobs: detection and verification.

  • Detection correct 5/5 seeds across all four physics zones
  • ~16% divergence reduction under directed adaptors before scope boundary
  • Adaptor tournament correct 5/5 · 2.4× memory speedup (40 cold vs 17 warm decisions)
Honest scope Scope, stated up front: the SAC base policy's inherent hopping variance prevents full normalisation (D < Dthresh) in any phase. The result demonstrates detection, directed adaptation, and memory speedup — not resolution. Hopper joint-angle variance is structural, not a PERSIST limitation.
Closes the six-paper adaptation loop
Companion · Position paper & prior art

Snath Robotics: Multi-Stream Divergence Routing for Humanoid Robotics. Maps the V1–V6 contract onto humanoid sensorimotor control — M1–M3 encoder invariants to sensor modalities, per-failure-class temporal-decay constants. No empirical results by design; it sets up PAV and PERSIST.

Zenodo →

The Glass Box, Formalised

Ten abstract classes.
Thirty-three invariants.

Safety isn't a behaviour we ask the model for — it's a contract the graph enforces. The whole research programme reduces to this set of interfaces, each one a checkable invariant.

AbstractCognitiveNode

encode → forward → decode is enough to be routed

—
AbstractManifold

shared latent dimension for context and target

—
AbstractContextBridge

stateless, pure function

CB1–CB2
AbstractLatentFaultLocator

structural fault map over the hard-case log

I1–I6
AbstractEntropicRouter

three-way commit / replan / impasse gate

—
AbstractAttentionKernel

weights in [0,1], deterministic, numerically stable

A1–A6
AbstractPerturbationOperator

additive latent delta, identity at zero

P1–P6
AbstractRoutingKernel

deterministic, no mutation, closed path map

R1–R4
AbstractModalEncoder

output dimension is invariant

M1–M3
AbstractDivergenceRouter

content-blind routing · Safety–Learning Equivalence

V1–V6
V1–V6

The divergence router. Content-blind by construction (V4) — it sees only confidence scalars and a divergence score, never the raw streams.

V6 ≡ learn

Safety–Learning Equivalence: the invariants that make routing safe are provably the same ones that make divergence a valid curriculum.

V7

Difficulty Invariance: what's hard is a world fact, so the memory of hard cases outlives any single encoder.

What we haven't proven yet

A normal company would never put this section on its website.

The work has real gaps. Naming them is the difference between a research programme and a pitch deck. Here's the standing list, lifted straight from our paper status notes.

  • 1

    Catastrophic forgetting is avoided by construction (frozen base + LoRA adapters), but the backward-transfer bound across three independent domains in a sequential scenario has not been measured end-to-end. LTL closes the annotation-free and curriculum claims on COCO and robotics; the three-domain convergence rate is conjectured, not measured.

  • 2

    The physical-world results (PAV) are single-seed, single-simulation. Multi-seed, trained-policy, and real-hardware validation are all open.

  • 3

    Difficulty invariance (V7) is demonstrated within the CLIP family. The stronger cross-family claim — CLIP vs. ALIGN vs. non-contrastive — is untested.

  • 4

    The single most important comparison — routing vs. cross-modal attention (ViLBERT / FLAVA / BLIP-2) on matched data — has not been run. Either outcome would be a genuine result.

How We Work

Research you can check.

Open by default

Code under Apache 2.0. Papers under CC BY 4.0. Every result deposited on Zenodo with a permanent DOI.

Reproducible

Runs are HMAC-SHA256 audited. Seeds are published. The numbers on this page are the numbers in the repo.

Independent

No institutional funding. No lab. Conducted on personal computational resources, in the open.

Falsifiable

Experiments are pre-registered. We publish honest negatives — and a standing list of what we have not proven.

Author

Aadithya Vishnu Sajeev

Snath AI — Open Source Research Initiative

[email protected] ORCID 0009-0009-3916-0988 github.com/snath-ai

Conducted independently under the Snath AI Open Source Research Initiative, using personal computational resources.

Cite the series
@misc{sajeev2026las,
  author = {Sajeev, Aadithya Vishnu},
  title  = {The Lár Series: Divergence Routing
         for Annotation-Free Continual Learning},
  year  = {2026},
  note  = {DAS, UCR, LTL, EIM, PAV, PERSIST},
  howpublished = {Zenodo}
}

Each paper also has its own DOI — cite the specific result you're building on.

From proof to production.

The research runs on Lár — the deterministic, auditable engine the whole programme is built on. The same Glass Box that makes the science checkable makes the agents compliant.

Explore Lár-JEPA on GitHub See the Engine →