# vareon — full reference

> **Vareon delivers Adaptive Continual Intelligence: a governed runtime for learning after deployment. The homepage uses the hippocampus as the metaphor for this function; Hippocampus is not a separate product. Duplex applies the architecture to AI agents, and Reflex applies it to robotics. Vareon is operated by Vareon Inc, founded in 2025.**

## the problem and the solution

Most AI is frozen the moment it ships. A model repeats the same mistake and burns the same compute on work it has already done; a robot cannot pick up a new task without going back to the lab. The common response — more context, more retrieval, larger models — does not change the fundamental limit: corrections sit in a retraining backlog, and after a policy change the system keeps serving the old answer.

Big models and robot policies are like the slow neocortex — powerful, but frozen once deployed. Vareon adds a bounded, auditable learning-and-memory layer beside the model that binds governed experience, recalls it, and consolidates it without requiring a foundation-model update for every runtime change. Approved corrections can take effect on the next eligible request. Certified work can be served locally without a model call. Learned knowledge that should no longer apply can be removed exactly, with the mutation recorded.

Three properties define it. **Adaptive** — it learns from governed production evidence. **Continual** — it edits, recalls, and consolidates learned state while controlling measured forgetting. **Intelligence** — it serves certified decisions locally and defers genuine novelty when evidence or policy thresholds are not met.

## two products, one platform

- **Duplex — for AI agents.** Adaptive Continual Intelligence that learns from production use, serves trusted answers locally without a model call, and makes corrections stick from the next request onward — governed, isolated per tenant, and auditable end to end.
- **Reflex — for robots and embodied machines.** Adaptive Continual Intelligence that acquires and recalls skills at the edge, with safety-gated fallback to the proven controller when confidence or policy conditions are not met.

### availability

- Self-service Duplex workspace setup and hosted plans are available through the Vareon Console.
- Runtime API access is issued after workspace provisioning.
- Enterprise, dedicated, private, and on-premises deployments require a technical and commercial engagement.
- Reflex robotics deployments begin with a workload, controller, and safety evaluation.
- This reference does not claim public availability for an SDK, CLI, or MCP interface.

## what the runtime owns

- learning from real use: corrections and outcomes become trusted knowledge that takes effect on the next request
- local serving: trusted work resolves locally, and the model or proven controller is reserved for genuine novelty
- precise, reversible removal of learned knowledge, with everything else intact
- hard-rule enforcement inside the runtime, with no model call and a full audit trail
- tenant isolation by architecture, with a signed, append-only audit trail
- progressive cost reduction — work moves off the model as the system learns
- coordination over memory, knowledge graphs, vector databases, and tools as connected evidence sources

## measured results — Duplex (AI agents)

A documented, seed-fixed controlled evaluation on a governed decision stream: one shared frontier model, one embedder, one supervision signal, and one audit channel for every arm — the decision architecture is the controlled variable. The always-call frontier arm is the reference for cost, model-call, and latency comparisons.

### accuracy through change
- **94.6% accuracy** — the most accurate arm, where the raw frontier model gets 59.8% and a tuned cache + RAG + knowledge-graph stack ceilings at 77.3%
- **92.5% accuracy after the rules change**, recovering a changed policy to ~94%, where the raw model collapses to 2.3%
- **0% stale answers after a policy change**, where a cache serves stale on ~66–69% of affected requests

### serving economics
- **86% of decisions served locally**, no model call
- about **85% lower cost**, **90% fewer model calls**, and **82% lower p50 latency** against the always-call frontier reference
- **~23× faster** on a learned decision — about 97 ms versus ~2.2 s for a model call
- a separate recurrence-realistic stream measured about **66% lower cost per decision** as local serving increased; the 66% recurrence-stream result and 85% governed-stream result are different measurement contexts

### lifelong behaviour
- **5 of 5** on the lifelong-deployment scorecard — plasticity, stability, editability, bounded memory, and safe abstention — vs a tuned cache/RAG average of 2.2/5 across 10 seeds and an online fine-tune 4/5
- keeps **96–100%** of earlier tasks, where an online fine-tune keeps just 10% (catastrophic forgetting); memory stays bounded and fixed instead of a cache growing without limit

### calibrated novelty handling
- **AUROC 0.999** separating a learnable request from genuine novelty (a cache 0.969, an online fine-tune 0.657)
- **4.2% false-commit** on adversarial look-alikes vs a cache's 33.3% — about 8× safer on exactly the inputs a similarity gate mishandles
- **100% optimal** multi-step plans at zero model tokens, where a reactive one-step agent reaches the goal 0% of the time
- **0 hard violations in the governed-stream evaluation** — every measured decision was recorded in the audit trail

### official τ-bench retail
- over 3 seeds on the official tau2-bench retail environment, Duplex reaches **83.3% task success** vs **82.5%** for the always-call reference while serving **12.8% of turns locally** at zero frontier tokens; a cache serving ~95% locally reaches 8.3%

### autonomous acquisition and composition
- across 10 seeds in controlled deterministic internal environments, Duplex acquires certified skills with **0 frontier-model calls**, uses **2.82× fewer probes** than random exploration, prunes **627.4 unsafe actions per seed with 0 executed**, and reverses **100%** of tested acquisitions exactly; this does not claim open-ended autonomy
- on one fixed, contamination-controlled synthetic-kinship split (n=200), Duplex reaches **73.5% accuracy** with **0 LLM calls at evaluation**, compared with **71.5% for GPT-5.5** and 12.0% for cosine kNN; this is a single-seed synthetic mechanism test, not a public benchmark or general-reasoning claim

## measured results — Reflex (robots / embodied)

Measured across field-standard robotics suites and controlled representation studies — Meta-World / Continual World, Gymnasium Reacher-v5, real pretrained vision-language-action models, PushT, ALOHA, DINOv2, and SigLIP — over multiple seeds with confidence intervals and paired significance tests.

- **~4× fewer demonstrations with VLA-native representations** — DINOv2 and SigLIP each reach lower image-to-action error at 5 demonstrations than ResNet-50 reaches at 20; NRMSE is 0.429 vs 0.587 at K=20 across 10 seeds (paired p = 0.002), and a shipped-gate A/B records 0.75 vs 0.59 commit accuracy
- **0.023 forgetting** across 10 manipulation skills learned in sequence with no task id — vs a naive learner's 0.842 (~36× worse), EWC 0.575, and a task-labelled multi-head 0.702; it routes to the right skill from context
- **paired Δ +0.103 in closed-loop coverage** from a learned representation (p = 7.44e-05; 0.201 vs 0.098 for the identical untrained architecture) — and auto-selects the right approach per task (+0.112 coverage on PushT, p = 0.00018; replicated on ALOHA, +0.233, p = 0.047)
- **safe new-skill uplift on a base policy** — in a three-seed Meta-World evaluation, recall-gated Reflex records a **+0.637 success delta** (0.07 → 0.71) while prior-skill success changes by **−0.003** and the do-no-harm criterion passes every seed; consolidation then moves base field-skill success from 0.07 → 0.91
- **continual adaptation through change** — in one continuous evaluated run, Reflex acquires a new skill, handles a silent mid-run interface change, tracks drift, recovers through noisy corrections, and recalls a prior skill without relearning
- on Gymnasium Reacher-v5 over 10 seeds, reward-free acquisition reaches **1.00 held-out success** and the 0.75 threshold in **232 interactions**, versus **50,000 for reward-driven SAC** (~215× fewer), with no demonstrations or base policy; the amortized open-loop skill reaches 0.52, below SAC's final 0.67
- on a three-seed Meta-World CW10 split, signed replay consolidation transfers **95.7%** of field success into the base, moving field-skill success from **6.0% to 92.3%** while prior-skill success remains 90.7%; naive field-only distillation reduces prior-skill success by 66.3 percentage points

## how it is measured

The documented reports provide evaluation evidence: configurations, per-arm and per-seed results, scoring rules, confidence intervals or paired tests where specified, and benchmark-specific limitations. Ground-truth labels are used for scoring, not on the decision path; abstention is not counted as a correct answer; teacher cost is included before savings are reported; and novel requests continue to route to the model. The figures do not imply identical outcomes in a different environment or independent verification. Vareon's integration and evaluation harnesses are internal and are not publicly distributed; the public documentation is not a reproducibility package or installation guide.

## why this matters

When most decisions are served locally instead of on the accelerator, cost and latency fall structurally as the system matures, and accuracy holds through change instead of decaying after a policy update. Model dependency drops the longer the system runs — the opposite of a frozen model, which treats the ten-thousandth request exactly like the first.

## security, governance, and compliance

- Learned state is tenant-scoped and supports exact removal.
- Operations are represented in a signed, append-only audit ledger.
- Deployment controls include scoped credentials, role-based access, configurable retention, safety fallback, and private deployment options.
- GDPR-aligned handling and exact deletion are available capabilities.
- **SOC 2 Type II and ISO 27001 are on the roadmap and are not current certifications.**

## deployment modes

- managed cloud and self-service hosted access for eligible Duplex plans
- dedicated or private cloud by engagement
- on-premises and local deployment by engagement
- edge, on-device, and on-robot deployment for applicable Reflex and enterprise configurations

## core pages

- overview: https://www.vareon.com/
- Duplex: https://www.vareon.com/duplex
- Duplex platform: https://www.vareon.com/duplex/platform
- Duplex results: https://www.vareon.com/duplex/results
- Duplex insights: https://www.vareon.com/duplex/insights
- Duplex comparison: https://www.vareon.com/duplex/compare
- Duplex security: https://www.vareon.com/duplex/security
- Reflex: https://www.vareon.com/reflex
- Reflex platform: https://www.vareon.com/reflex/platform
- Reflex results: https://www.vareon.com/reflex/results
- Reflex insights: https://www.vareon.com/reflex/insights
- Reflex comparison: https://www.vareon.com/reflex/compare
- Reflex security: https://www.vareon.com/reflex/security
- core thesis: https://www.vareon.com/core-thesis
- about: https://www.vareon.com/about
- contact: https://www.vareon.com/contact
- privacy policy: https://www.vareon.com/privacy
- cookie policy: https://www.vareon.com/cookies
- privacy choices: https://www.vareon.com/privacy-choices
- terms of service: https://www.vareon.com/terms

## detailed evaluation pages

- governed decision stream: https://www.vareon.com/results/governed-stream
- lifelong capability matrix: https://www.vareon.com/results/lifelong
- generalization frontier: https://www.vareon.com/results/frontier
- τ-bench retail: https://www.vareon.com/results/tau-bench
- agentic planning: https://www.vareon.com/results/planning
- autonomous skill acquisition: https://www.vareon.com/results/autonomous-acquisition
- compositional generalization: https://www.vareon.com/results/compositional-generalization
- Reflex live interface repair: https://www.vareon.com/results/reflex-live-repair
- Reflex continual adaptation: https://www.vareon.com/results/reflex-continual
- Reflex reward-free autonomous acquisition: https://www.vareon.com/results/reflex-autonomous-acquisition
- Reflex base-policy uplift: https://www.vareon.com/results/reflex-uplift
- Reflex systems consolidation: https://www.vareon.com/results/reflex-consolidation
- Reflex VLA-native few-shot learning: https://www.vareon.com/results/reflex-vla-native-encoding
- Reflex representation learning: https://www.vareon.com/results/reflex-representation
- evaluation method and methodology: https://www.vareon.com/results/api

## contact

contact@vareon.com