R-Think
GitHub

Concepts

R-Think is built on a fixed set of concepts that govern how a reasoning process moves from observation to accepted decision. These concepts are non-negotiable and apply to every mission executed by the runtime.

What is R-Think

R-Think is a Cognitive Reasoning Runtime for observable, evidence-governed execution. It is not an AI model, not an LLM, and not an agent framework. It is a framework, protocol, and runtime that governs how a reasoning process observes real conditions, understands context, questions assumptions, validates claims, connects evidence, challenges decisions, discovers new understanding, and evolves in a controlled manner.

Core Principles

R-Think is governed by nine non-negotiable principles:

  1. Observability — Every step in the reasoning process must produce state, artifact, or evidence that can be inspected.
  2. Evidence Before Completion — A mission is not complete until the required artifact and evidence have been produced and accepted.
  3. Challenge Authority — Decisions can be challenged within the mission. Authority boundaries are explicit and enforced.
  4. Explicit State — Cognitive and operational states are modeled explicitly. Implicit state is not permitted.
  5. Replayability — Every mission produces an append-only event history that can be replayed in full.
  6. Model Independence — The runtime does not depend on a single model or provider. Consumers and providers are decoupled.
  7. Controlled Evolution — Contradictions and new understanding surface as explicit nodes that may trigger retry, challenge, or evolution.
  8. Authority Boundaries — Only authorized actors or governance layers may issue decisions, accept outcomes, or override authority gates.
  9. Completion Criteria — Completion is validated against the mission's risk level and adaptive depth input, not asserted.

Canonical Algorithm

The canonical algorithm is the fixed sequence of steps that every governed reasoning process must pass through. The algorithm is not always linear; contradictions, insufficient evidence, and method failures return the runtime to earlier steps through loop rules.

canonical-algorithm.txt text 
Observe
→ Understand
→ Question
→ Validate
→ Connect
→ Challenge
→ Discover
→ Evolve

Each step produces a cognitive state transition, an artifact, and optionally an evidence graph update. The Mission Runtime Coordinator ensures no step is skipped and no transition occurs without the required artifact or evidence.

Cognitive States

The runtime tracks eight cognitive states:

IDLE — Awaiting observation input.
OBSERVING — Ingesting and normalizing source data.
UNDERSTANDING — Building context from observations.
QUESTIONING — Identifying assumptions and constraints.
VALIDATING — Checking claims and evidence compliance.
CONNECTING — Mapping evidence to claims and decisions.
CHALLENGING — Challenging decision quality before acceptance.
ACCEPTED — Mission complete; result is accepted.

Operational States

The runtime tracks twelve operational states:

PENDING
RUNNING
PAUSED
RETRYING
AWAITING_EVIDENCE
AWAITING_ACCEPTANCE
CHALLENGED
EVOLVING
COMPLETED
REJECTED
FAILED
CANCELLED

Adaptive Depth

The runtime adjusts verification and reasoning depth based on mission risk:

Level Risk Expectation
L0 Minimal Basic artifact only; minimal evidence required.
L1 Low Trace plus evidence for each claim.
L2 Medium Full evidence graph, authority gate, and completeness check.
L3 High Multi-party challenge, formal acceptance, and independent replay.

Loop Rules

The runtime may return to earlier steps under five loop rules:

  1. Insufficient Evidence — Return to Observation or Validate until evidence meets the required depth.
  2. Contradiction Detected — Return to Connect or Challenge whenever evidence conflicts with an existing claim.
  3. Method Failure — Return to Observation when a provider or method cannot satisfy the requirement.
  4. Challenge Issued — Return to Question or Validate when an authorized challenge is raised.
  5. Evolution Required — Return to Discover when the current process cannot resolve the contradiction within allowed attempts.

Loop Limit

Each mission has a maximum loop count per step. Exceeding the loop limit transitions the mission to FAILED with a reason code. The limit is determined by the mission's risk level and adaptive depth.

Authority Boundaries

Authority boundaries define who or what may issue decisions, accept outcomes, or override governance:

  • Mission Operator — Authorized to issue decisions for L0 and L1 missions.
  • Domain Authority — Required for L2 acceptance when the mission touches regulated or safety-critical domains.
  • Independent Reviewer — Required for L3 high-stakes missions to provide challenge and acceptance.
  • Runtime — May reject any decision that violates the evidence or authority boundary rules. The runtime is the enforcer, not the override.

The Inspector exposes authority records for every decision so auditors can verify that the correct authority was consulted.