Glossary

Key terms used throughout the Ejentum documentation. For the theory behind these concepts, see The Method. For practical usage, see the API Reference.

API Response Components

Negative Gate The failure pattern the agent must avoid. A specific, named cognitive error that the ability is designed to prevent. Example: "Treats the first visible symptom as the root cause without tracing the causal chain deeper."

Target Pattern What correct reasoning looks like for this task type. A concrete description of the reasoning behavior the ability is designed to produce.

Reasoning Topology The execution structure of the ability. Defines the steps, decision gates, loops, and exit conditions the agent should follow. Example: S1:identify → S2:trace → G1{found?} --yes→ OUT --no→ S2[LOOP]

Falsification Test A verification criterion the agent checks its output against. If the output fails this test, the reasoning procedure was not followed correctly. Example: "If the proposed fix addresses a surface symptom without tracing the causal chain, root cause analysis was skipped."

Cognitive Payload The combined set of suppression vectors, amplification vectors, cognitive style, and reasoning elasticity delivered with each ability.

Cognitive Payload Fields

Suppression Vectors Failure modes to actively block during generation. These are specific reasoning shortcuts that the ability is designed to prevent. Examples: symptom_treatment_bias, surface_level_stop, optimism_bias, correlation_as_causation. In our benchmarks, suppression signals produce larger behavioral improvements than amplification alone. The effect is multiplicative: each named failure mode eliminates an entire branch of incorrect outputs. Self-monitoring improved +132% primarily through suppression, not amplification.

Amplification Vectors Reasoning patterns to prioritize during generation. These signals pull the model's output toward specific cognitive operations.

Cognitive Style A single semantic anchor that sets the reasoning methodology for the entire generation. Examples: root cause isolation, bayesian inference, counterfactual simulation.

Reasoning Elasticity Controls how tightly the agent must converge on a specific answer versus how freely it can explore. Two components:

• Coherence target: What the output should converge toward (e.g., calibrated probability, audit completeness)
• Expansion factor: How much creative latitude is allowed (conservative, adaptive, high variance, max entropy)

Modes

Dynamic Modes One operation per call: the highest-scoring match for your query, returned as authored. Modes: reasoning, code, anti-deception, memory. Available on all tiers, including the free trial. Best for routine tasks where the general operation fits.

Adaptive Modes The same matched operation, with its procedure and reasoning topology rewritten by an adapter model to name your task's specifics. Modes: adaptive-reasoning, adaptive-code, adaptive-anti-deception, adaptive-memory. The cognitive payload, verification check, failure pattern, and correct-shape example are returned identical to the dynamic version: adaptation changes the approach, never the safety guards. Requires the Go or Super tier.

Reasoning Dimensions

Causal Addresses direction-of-causation errors. Prevents the agent from treating correlation as causation.

Temporal Addresses sequence and duration errors. Prevents the agent from confusing past and future or confabulating timelines.

Spatial Addresses boundary and topology errors. Prevents the agent from violating physical or structural constraints.

Simulation Addresses counterfactual and consequence modeling errors. Prevents the agent from ignoring downstream effects.

Abstraction Addresses category and generalization errors. Prevents the agent from conflating metaphor with mechanism.

Metacognition Addresses self-awareness failures. Prevents the agent from continuing to reason incorrectly without recognizing its own degradation.

Other Terms

RA2R (Reasoning Ability-Augmented Retrieval) The core methodology. Unlike RAG (Retrieval-Augmented Generation) which retrieves information, RA2R retrieves reasoning abilities. The agent receives structured cognitive injections that govern HOW it thinks, not WHAT it knows.

Cognitive Scaffold The complete reasoning structure delivered per ability. Includes negative gate, reasoning procedure, topology, target pattern, suppression/amplification vectors, and falsification test.

Adapter Model The model that rewrites a matched operation's procedure and topology to a specific task in adaptive modes. It changes only those two sections; the failure guard, suppression signals, verification check, and correct-shape example are passed through unchanged.

Delivery & Integration

MCP Server Model Context Protocol server. A small program (or hosted endpoint) that exposes the harnesses as eight MCP tools (four dynamic, four adaptive) any MCP-compatible client (Claude Desktop, Cursor, Windsurf, Claude Code, n8n MCP Client node) can call. Same harness API underneath; different delivery channel. The official Ejentum MCP server is ejentum-mcp. Two install paths: hosted HTTPS at https://api.ejentum.com/mcp for HTTP-MCP clients, or stdio via npx -y ejentum-mcp for subprocess-spawning clients. See the MCP integration guide.

MCPB Bundle A packaging format for stdio MCP servers, built with Anthropic's @anthropic-ai/mcpb tool. A .mcpb file is a zip containing the server's compiled code plus a manifest declaring its tools, user-config schema (the API key prompt), and runtime metadata. Some MCP marketplaces consume this bundle format for one-click installs; the Ejentum recommended install paths are the hosted endpoint (https://api.ejentum.com/mcp) or direct npm (npx -y ejentum-mcp).

Ability A single engineered cognitive operation in the 679-ability catalog. Each ability is a 20-field protocol including suppression vectors, reasoning topology, falsification test, and synergy edges. See Concepts for the full anatomy.