SIGNIT System Architecture — Autonoma/Intelligence

00

Authority topology

Authority does not follow access. The Human retains consequential decisions. The Orchestration Agent coordinates work and verifies evidence. SIGNIT governs the intelligence pipeline. Bounded Agents execute only the capabilities named in an exact work order.

Human

Defines architecture and policy, authorizes task class and provider use, judges evidence, approves editorial copy, and separately authorizes publication, deployment, and delivery.

BOUNDARY · never inferred from a tool permission

Orchestration Agent

An LLM-backed agent that reconstructs exact state, compiles work orders, invokes permitted specialists, synthesizes results, binds packages, and independently verifies outcomes.

BOUNDARY · coordination is not consequential authority

SIGNIT

Ingests sources, preserves lineage, structures analysis, challenges claims, applies evidence finality, updates route memory, and emits governed artifacts.

BOUNDARY · does not author or publish the final Brief

Bounded Agents

Perform scoped analysis, adversarial review, implementation, rendering, or execution from a capability-limited work order and return artifacts plus proof.

BOUNDARY · output remains evidence until reviewed

01

Runtime planes

The system separates probabilistic reasoning from deterministic enforcement and human authority. Arrows show data and control flow, not automatic permission.

System map · scroll horizontally on smaller screens →

Source plane observation + provenance

Public sources

Source router

Ingest + lineage

Intelligence plane probabilistic analysis under deterministic gates

Analyst agents

Claim verification

Adversarial review

Evidence finality

Route memory

Orchestration plane operator-side coordination + verification

Exact task bind

Work-order compiler

Specialist lanes

Orchestration AgentLLM-backed control loop
scope · coordinate · synthesize · verify

Synthesis + adjudication

Manifest ledger

Publication package

Deterministic control plane non-probabilistic enforcement

Schema validation

Provenance

State precedence

Hash binding

Stale-safe checks

Redaction

02

Agent blueprint

An agent is not just its model. The model is the reasoning engine inside a governed runtime envelope. The complete agent includes identity, scope, inputs, capabilities, working state, output contract, verification, and an authority gate.

Agent anatomy · left to right execution contract →

01

Identity + task class

who acts, exact objective, accountable owner, task boundary

02

Exact source capsule

frozen inputs, source hashes, lineage, freshness, exclusions

03

Capability envelope

allowed actions, prohibited actions, path and time limits

04

Tools + models

authorized model profile, deterministic tools, permissions

05

Working state

ephemeral workspace, checkpoint, manifest, session binding

06

Structured output

typed artifact, citations, uncertainty, complete provenance

07

Independent verification

separate checks re-establish claims, sources, and constraints

08

Authority gate

Human decides whether output may influence action or publication

Agent output is evidence, not authority.

03

Evidence paths

Every load-bearing claim must preserve the path from observation to source, verification, adversarial challenge, current-run finality, and downstream use.

Observationtimestamped source capture

Lineageorigin, retrieval context, transformations

Claimbounded assertion with dependencies

Challengeverification, red team, contradiction

Finalitysupported, contested, caveat-only, non-final

Usebrief readiness or explicit hold

route pressure≠evidence

agent verdict≠proof artifact

controlled draft≠publication

deployment≠subscriber delivery

04

Control gates

Deterministic controls establish eligibility; the Orchestration Agent coordinates and verifies; the Human decides whether the system may cross a consequential boundary.

Source lineage

No claim advances without origin, retrieval context, transformations, and dependent-claim trace.

Current-run finality

A stale pass, older recheck, or route signal cannot override a current conflict.

State precedence

Exact-run and canonical evidence outrank derived state, mirrors, agent reports, and conversational memory.

Capability binding

Every agent operates inside an explicit tool, model, path, mutation, and output envelope.

Exact authorization

Approval binds to one task or one byte-exact package. A changed byte invalidates publication authority.

Independent verification

Execution reports are not accepted as proof; load-bearing results are re-established from exact evidence.

FAIL CLOSED: if identity, authority, evidence freshness, active-writer state, exact package binding, or verification is ambiguous, stop. Bind the missing fact from its authoritative source before proceeding.