Vector Pulse Dynamics
KAYA / Overview

KAYA

Kill All Your Assumptions

The scientific-method runtime for AI agents.

KAYA turns a question into a bounded investigation: expose the assumptions, choose a qualified environment, test the hypothesis, and validate the evidence before action.

Watch the CLI Think Inspect the Methods
13
deliberate methods
8
segmented control layers
0
ambient authority paths

System tour

One inquiry. Six controlled stages.

Operating thesis

Models generate.
KAYA determines what earns belief.

Most agent systems collapse planning, execution, judgment, memory, and authority into one conversation. KAYA runs the scientific method as runtime state, then routes each stage into the smallest qualified harness.

Confidence is not evidence.

Every relied-on claim keeps provenance, contradiction, freshness, and purpose.

Activity is not completion.

Executor output remains a candidate until an independent validator accepts it.

Evidence is not authority.

A supported conclusion can propose action; it cannot authorize its own effect.

Expose assumptions

A question is never just a question.

KAYA decomposes an instruction into claims, unknowns, constraints, dependencies, and hidden premises before a model can turn ambiguity into irreversible work.

Incoming inquiry

Assumption register

What must be true before proceeding

    Observed Supported Contested Refuted Unknown Stale Accepted for this decision

    Route the thinking

    Different questions require different ways of knowing.

    KAYA routes by question type, stakes, reversibility, evidence, time, and uncertainty. One orchestrating method owns the lifecycle; techniques assist, gates can stop, and retrospectives can propose change without rewriting history.

    7OrchestratingOwn the root lifecycle
    3Embedded techniquesServe a declared stage
    2Cross-cutting gatesBlock but never authorize
    1RetrospectiveLearns without rewriting

    Power

    Best used for

    Produces

    Cannot establish

    Example operation

    Canonical lifecycle

    Every transition leaves an artifact

      Execution method Build / validate / repair

      Turn thought into a project

      Planning, building, and proof belong to different hands.

      Once a thinking method produces an accepted design, KAYA compiles it into bounded work. An orchestrator coordinates, an executor builds, and an independent validator decides whether the candidate actually satisfies the frozen criteria.

      Input
      Accepted design + objective
      Contract
      8 frozen criteria · 2-attempt ceiling
      Output
      Immutable candidate + evidence receipts
      Effect authority
      Human-owned integration and ship
      00 / 07 Ready to compile

      Protocol demo Predetermined project evidence shows the control flow; this page does not mutate a repository or ship software.

      1. 01 / Coordinate

        Orchestrator

        Decomposes the accepted design, assigns minimal environments, and compiles evidence-driven rework.

        Owns
        Plan · delegation · rework
        Cannot
        Write · validate · ship

        Standing by

      2. 02 / Build

        Executor

        Works inside one scoped environment, runs sealed checks, and freezes an immutable candidate with receipts.

        Owns
        Bounded mutation · candidate
        Cannot
        Change plan · self-accept · integrate

        Waiting for assignment

      3. 03 / Prove

        Validator

        Tests the read-only candidate against frozen criteria and publishes a typed, independent verdict.

        Owns
        Evidence · verdict · findings
        Cannot
        Modify · repair · ship

        Waiting for candidate

      Protocol ready

      Freeze the contract before work begins.

      The objective, non-goals, acceptance criteria, authority, budget, and retry ceiling become immutable inputs to the build cycle.

      Artifact
      Awaiting compiled mission plan
      Boundary
      No role can silently widen the contract.

      Artifact trail

      One repair loop, no self-grading

      7 transitions
      1. 01
        OrchestratorMission plan compiled
      2. 02
        ExecutorCandidate 01 frozen
      3. 03
        ValidatorCriterion refuted
      4. 04
        OrchestratorBounded rework compiled
      5. 05
        ExecutorCandidate 02 frozen
      6. 06
        ValidatorAll criteria confirmed
      7. 07
        Human gateReady for owner decision
      Human effect authority Accepted is not shipped.

      Integration, merge, publish, and deployment stay locked until the owner reviews the validated evidence bundle.

      Locked

      Run the inquiry

      Watch one sentence become an auditable investigation.

      The CLI does not jump from prompt to answer. It exposes what the request assumes, traces what failure would mean, routes distinct ways of thinking, and compiles a minimal harness for every stage before a candidate can reach human review.

      Interactive product simulation The inquiry and evidence are predetermined so you can inspect KAYA’s control model—not a fabricated live claim.

      kaya / inquiry runtime Ready

      When started, the demonstration types the selected KAYA inquiry command and appends its ordered runtime events below.

        Choose an inquiry, then run the investigation. 0 / 7 phases
        Root owns the lifecycle Method routing pending
        Harnesses stay segmented No environment compiled
        Effect authority Locked

        Only a person can accept residual risk and authorize merge, push, or deploy.

        Compile the environment

        More capability outside the window. Less noise inside it.

        The registry can be enormous. The active environment cannot. KAYA resolves the smallest qualified combination of model, tools, skills, context, policy, isolation, and evidence for each scientific-method stage.

        Model×Tools×Skills× Context×Policy×Safety EnvironmentRevision
        01

        CLI and SDK surface

        One control protocol across terminal, TUI, Rust, TypeScript, Python, RPC, daemon, MCP, and editor clients.

        02

        Bounded agent loop

        Streaming turns, cancellation, budgets, compaction, retries, checkpoints, and explicit terminal outcomes.

        03

        Model and environment router

        Eligibility before ranking, exact revisions, health, privacy, residency, fallback legality, cost, latency, and abstention.

        04

        Tool and effect plane

        Typed schemas, capabilities, approvals, safe built-ins, sandbox tiers, receipts, uncertain-effect recovery, and revocation.

        05

        Operation coordination

        Plan compilation, scoped workers, leases, fencing, heartbeats, independent validation, integration, and cancellation.

        06

        Ledger, claims, and sessions

        Append-only events, artifacts, migrations, replay, bitemporal memory, contradiction, provenance, and stale-dependency closure.

        07

        Enterprise control plane

        Identity, tenancy, secrets, information flow, audit, policy, approvals, retention, isolation, and incident response.

        08

        Operations and release

        Doctor, observability, packaging, compatibility, updates, rollback, backups, recovery, support, and supply-chain evidence.

        Capability registry Models / tools / skills / MCP / memory / methods

        Discoverable, qualified, versioned, and normally outside model context.

        Stage working set Only what this decision requires

        Artifact handles replace history; retrieval replaces injection; overflow triggers decomposition.

        Bounded learning

        Failure becomes an experiment, not a memory fragment.

        KAYA joins the decision, task, environment, evidence, and failure that produced an outcome. A proposed repair must survive shadowing, a bounded trial, independent evaluation, monitoring, and automatic rollback before it can influence later work.

        1. 01Detect recurrence

          Cluster failures without treating similarity as causality.

        2. 02Propose repair

          Down-rank a route, inject a guard, or add a targeted check.

        3. 03Shadow

          Compute behavior without changing the live decision.

        4. 04Experiment

          Compare exact treatment and control under a sealed plan.

        5. 05Graduate or revert

          Measured benefit earns bounded use; regression rolls back.

        The actor cannot grade itself into production.

        Memory can inform a decision. It can never become truth, permission, or policy by repetition.

        Vector Pulse Dynamics

        Vector Pulse Dynamics

        Building control systems for machine intelligence.

        Vector Pulse Dynamics designs infrastructure that makes autonomous systems more inspectable, bounded, and accountable. KAYA is the first product in that platform.

        Purpose

        Move AI from plausible activity to defensible outcomes.

        Give operators a chain of custody from question to evidence to authorized action.

        Why now

        Agent horizons are expanding faster than confidence.

        Parallel workers and larger tool registries increase the need for context economy, independent validation, and authority boundaries.

        Unique insight

        The harness must choose the process before the model.

        Reliable behavior comes from separating method, environment, evidence, and authority rather than adding more prompt instructions.

        Business thesis

        Open local runtime. Managed operations. Qualified intelligence.

        Local evidence stays locally authoritative; commercial value comes from operations and consent-governed cross-workload learning.

        Capital purpose

        Turn an engineered system design into measured runtime proof.

        1. 01Foundation and enterprise vertical slice
        2. 02Qualified providers, tools, isolation, and validation
        3. 03Design-partner evidence and measured product pull
        4. 04Controlled repair loop and managed operations