rAIls · Runtime Agentic Intelligence Stack

    Turn chaotic workflows into learning systems.

    Governed AI operations that decide, act, verify outcomes and learn what works.

    Same models. Better decisions from verified experience.

    Decide better · Act reliably · Learn safely

    The model stays frozen. The system learns.

    NVIDIA Inception Program

    A model call is not a decision system.

    Models can reason about a decision. They do not by themselves own durable state, verify real-world execution, distinguish policy failure from missing evidence, or turn verified outcomes into bounded future improvements. ReDevOps makes those Runtime responsibilities.

    Typical agent
    PromptModelToolResponse

    A request in, a response out. No durable state, no verification of what actually executed, no memory of the outcome.

    ReDevOps
    ObserveEvidenceDecideGovernActVerifyLearn
    Learn feeds back into Decide — a closed loop, not a dead end.

    Every consequential decision is observed, grounded in evidence, governed, executed, verified against reality — and turned into Experience the next decision can use.

    Same model. Different Runtime. Different decisions.

    Chess is the laboratory, not the product. It gives us a deterministic rule system and an independent oracle for measuring whether the Runtime itself improves decisions — while the model weights never change.

    CONTROLLEDarchitecture ablation
    ~52% fewer blunders
    Hold the model constant; change only the Runtime.
    LLM-only0.408
    + Context Runtime0.365
    + learned lessons0.307
    Complete Runtime0.197

    Frozen Qwen · 726 held-out positions · Stockfish 17.1 depth-16 deterministic evaluator · no model-weight changes.

    A separate bounded external engine-hint arm reaches 0.167 — that is external assistance, not part of the Runtime claim.

    Update, 27 October 2026 · internal rerun
    Rerun with a different model (Qwen3.8-Flash-Next 8-bit): no Runtime 0.413 over all 726 positions (published 0.408); full Runtime 0.100 on 191 stratified positions against 0.426 without it on the same positions. The direction holds; the published 0.197 was not reproduced. The rerun's full-Runtime arm includes a depth-8 Stockfish hint, so its result includes engine assistance. The other arms were not run.

    CONTROLLEDlearning curve
    0.390 → 0.293
    Held-out blunder rate as verified Experience accumulated.

    The learning curve has not been rerun and is awaiting reproduction (see the update in the architecture ablation card).

    0 training decisions0.390
    50 training decisions0.383
    150 training decisions0.351
    400 training decisions0.293

    Held-out · model weights frozen · no fine-tuning. Learning is outcome-grounded, bounded and selectively applied.

    The first result isolates what the Runtime stack adds; the second isolates learning from accumulated Experience. They are two separate experiments and stay separate — this is not a claim that ReDevOps makes any model 52% smarter.

    One loop from observation to better future decisions.

    rAIls forms one closed decision loop. Observation becomes evidence, evidence becomes a governed action, execution is verified against reality, and verified outcomes become Experience the system learns from — while the model stays frozen.

    The Decision Runtime — one closed loop
    WORLDDISCOVERYDECISION INTELLIGENCEMISSION RUNTIMECONTEXT RUNTIMECAPABILITIESVERIFIED OUTCOMELEARNExperience→ futuredecisionsmissing evidence → discoverGOVERNANCE · EVIDENCE · REPLAY — across the entire loop

    Running the stack in enterprise infrastructure? Extend the same runtime contracts with intrinsic security and telemetry, enterprise runtimes for Go, Kotlin and .NET, and data-platform integrations for Spark and Databricks — the production integration layer around the same architecture, not a separate proprietary AI stack. Explore Enterprise Edition

    Don't learn when learning isn't the problem.

    ReDevOps does not assume every decision should learn. It measures whether the bottleneck is policy, evidence or applicability — and routes accordingly.

    POLICY-LIMITEDLEARN

    The evidence supports better decisions than the current policy is making. Use verified outcomes to improve future selection.

    EVIDENCE-LIMITEDDISCOVER

    The available evidence cannot distinguish the right action. Determine which additional observation could change the decision — and whether it is worth acquiring.

    LEARNING-UNCERTAINFALL BACK

    Prior Experience is not sufficiently applicable. Use the frozen policy, request more evidence, abstain or escalate to a human.

    AI shouldn't learn everything. It should learn when Experience can help, discover when evidence is missing, and abstain when neither is reliable.

    One decision architecture. Very different worlds.

    The same Runtime contracts run across environments with very different oracles, outcomes and stakes. Each is a decision laboratory for the same question: can the same frozen model decide better from verified experience?

    quantify.club
    Sequential quantitative decisions

    Asset allocation is the current research environment for path-dependent decisions, delayed outcomes, risk/return trade-offs and changing market regimes. The question is whether the same frozen model can improve as verified experience accumulates.

    IN RESEARCH
    quantify.club
    rAIlsApps
    Operational business decisions

    Repeated decisions across receivables, sales follow-up, fraud review, supplier reconciliation, support, renewals and other operational workflows — including when doing nothing is the right action. Synthetic worlds validate the mechanism; they do not clear the external-validity gate.

    CONTROLLEDWAITING FOR REAL DATA
    See the live stack
    Chess
    Controlled decision laboratory

    Known rules, an independent oracle and repeatable trials. Used to isolate what the Runtime adds before moving into weaker-oracle real domains.

    PUBLISHED
    Read the chess benchmark
    The shared primitiveEvidence → Decision → Action → Verification → Experience

    Let AI do real work — without losing control.

    Not a taxonomy — a real Mission. Here is one execution, start to sealed proof: what it knew, what it was allowed to do, who approved the consequential step, and how it was verified.

    Mission #1842
    Investigate KYC ownership risk
    VERIFIED
    1. 09:41Evidence discoveredView evidence
      OpenSanctions · company registry · CRM · 14 evidence objects
    2. 09:42Plan selectedWhy this plan?
      SQL + property graph · score .94
    3. 09:42Authority check
      Customer records allowed
      Sanctions lookup allowed
      External actionHUMAN APPROVAL
    4. 09:44Alex approvedView approval
    5. 09:44ExecutedView execution
      Case status updated
    6. 09:45VerifiedVerification
      Expected state == observed state
    Evidence sealed · replay availableReplay Mission
    When a check fails
    Missing evidence→ STOP
    No authority→ BLOCK
    Uncertain→ REVIEW
    Wrong result→ CATCH

    Real workloads. Measured outcomes.

    We test whether the Runtime makes better decisions — and whether unsafe ones get through.

    Plan qualityKYC evidence plan
    SQL0.78
    Graph0.81
    SQL + Graph← Runtime0.94
    +16% over best single representation280 ms selection latency

    Awaiting reproduction · 27 October 2026
    The harness and the definition of the plan-quality metric behind these scores have not been located, so they are shown as published and not as a verified result.

    32 real parcels · sampled from a 149,624-district harvested index
    0 / 160
    False-permitted conclusions
    Context Runtime zoning resolver, 32 parcels × 5 uses. A false-permit is a PERMITTED conclusion against an incompatible base zoning; the other 98 checks abstained as UNKNOWN or were prohibited. Not a measure of agent governance.
    98 UNKNOWN
    How this was measured

    The resolver classifies each parcel's official base zoning code into a residential, commercial or industrial family. It answers PERMITTED only for single-family residential on a residential code, PROHIBITED when the families differ, and UNKNOWN when the ordinance use-table is needed. A false-permit is a PERMITTED answer against an incompatible family. The zero follows from that rule, not from a ground-truth comparison, so it is a resolver property rather than an accuracy claim. The 32 parcels are a fixed, committed sample of the 149,624-district harvest index (kept off-repo); the run is offline and deterministic. A CI workflow re-runs it against pinned resolver commit ceeff63 and publishes the JSON (last run 2026-10-13: 7 permitted, 55 prohibited, 98 unknown, 0 false-permits). Re-run again on 2026-10-27: the same counts, with the same input hash.

    Cross-system Mission — one objective, four systems, one proof
    Chatwootevidence
    Twentyaccount
    ERPNextfinancial
    OpenSanctionsverify
    VERIFIED

    Explore all benchmark and reference workloads

    Your systems become capabilities.

    ReDevOps doesn't replace your models, agents, data stores or business applications — it gives them one common execution boundary.

    What you already use
    Models
    • OpenAI
    • Anthropic
    • Gemini
    • Local / PAIR
    Agents
    • LangGraph
    • PydanticAI
    • CrewAI
    • Custom agents
    Data
    • SQL
    • Graphs
    • RAG
    • APIs
    Apps
    • Twenty
    • ERPNext
    • Chatwoot
    • Lago
    rAIls
    one governed Mission

    NVIDIA NeMo · PydanticAI · LlamaIndex · CrewAI · Azure Semantic Kernel · Google ADK / A2A · AWS Strands · DigitalOcean · LangGraph

    Your rAIlsApps. One install.

    Start with the machine and capabilities you already have. ReDevOps discovers what is available, asks before installing anything consequential, and gets you to a first Mission — not a config file.

    What installs is a suite of source-available (AGPL-3.0 with the Commons Clause) apps — six governed domain agents on rAIls (source-available) — Revenue, Intelligence, Finance, Customer Success, Content, and Security & Compliance — over real open-source cores (Twenty, Metabase, Postiz and more). The apps are being made runtime-native: approvals gate consequential actions today, read-back verification is live for a few capabilities, and the priority engine and learning loop are rolling out across apps. Self-host it or run it on AWS, Google Cloud, Azure or DigitalOcean, with your own model. See the rAIlsApps.

    ReDevOps Setup
    This machine is ready
    AI
    Local AI availableNVIDIA PAIR · Ollama
    Applications
    Twenty CRM
    ERPNext
    +Connect another application
    Your first Mission is ready.
    Run first Mission
    Use what you already have
    OpenAIAnthropicGeminilocal modelsNVIDIA PAIR
    Models & inference
    Connect your applications
    TwentyERPNextChatwootLagoMetabase+ integrations
    Browse integrations
    Run where you want
    WindowsmacOSLinuxKubernetes
    Deployment options
    Quickstart
    $ git clone https://github.com/redevops-io/agentic-os
    $ cd agentic-os/deploy/apps
    $ docker compose up -d
    See what's installed

    One command stands up the whole agentic app — a single docker compose up on a host, or one helm install in a cluster — the runtime, the Projects Control Plane, the infra operator (deploy · provision · configure · verify · drift · rollback), the deployment skills, and a wired Prometheus + Grafana + Loki stack. Run AI on your own machines →

    Already have an agent?

    Wrap LangGraph, LangChain or your own agent as a governed capability — no framework rewrite. Mission SDK

    Or point it the other way — let Claude, ChatGPT or Cursor drive your apps through the Governed Agent Gateway — governed, not an ungoverned backdoor.

    Design partners

    Have historical decisions and outcomes?

    We are looking for design partners with repeated operational decisions where historical evidence, action logs and outcomes can be reconstructed. ReDevOps can begin read-only, replay the historical process, establish a baseline, and test whether Context, Decision Intelligence or verified Experience changes the result — before any production action is enabled.

    For receivables specifically, we ask for
    Receivable / decision state as-of decision time
    Customer or account context
    The intervention or action taken
    Execution confirmation
    The subsequent payment / outcome
    Recurrence and history
    Timestamps for leakage-safe reconstruction
    Permission to evaluate

    Capable models need a Runtime that can decide, act, verify and learn.

    Pick one repeated consequential decision. We determine whether the bottleneck is context, policy, missing evidence or execution — or whether the Runtime does not help.

    Decide better · Act reliably · Learn safely