From Chatbot to Digital Colleague: The Paradigm Shift Toward Persistent Autonomous AI

Introduction

• Thinking LLMs use inference-time computation, Chain-of-Thought, reflection, process supervision, and reinforcement learning to support more deliberate reasoning than fast Chatbot-era generators.Chatbot-era systems can struggle with deep reasoning, verification, and long-horizon consistency.Source: §1 · p. 5
• OpenClaw-style systems embed tool use in persistent workspaces with reusable skills and verification, addressing fragile Agent-era tool calls and enabling context maintenance, recovery, and final-state checks.Agent-era systems can be derailed by incorrect action formats, missing observations, failed calls, or unrecovered errors.Source: §1 · p. 5
• Workspace + Skill turns episodic responses and atomic tool calls into durable work by combining persistent state, memory, and evidence with reusable procedures for planning, checks, recovery, and validation.Workspaces can include files, terminals, browsers, repositories, calendars, and other applications; skills are parameterizable procedures.Source: §1 · p. 5
• The learning and evaluation unit shifts from instruction-response pairs and answer correctness toward state–action–observation trajectories and task closure under reproducible, auditable, and safe conditions.Reasoning data and evaluation also expand to include Chain-of-Thought traces, process supervision, verifiable rewards, and reasoning-process judgment.Source: §1 · pp. 5–6
• The transition faces bottlenecks in factual grounding, long-horizon execution, memory and state management, safety, governance, skills, and system maintenance.The survey argues that progress requires stronger foundation models as well as improved execution substrates, skills, evaluation environments, and governance mechanisms.Source: §1 · p. 6
• The survey covers cognition, tool-augmented execution, Workspace + Skill, and data and evaluation, then discusses open challenges and self-evolving AI ecosystems.Its proposed ecosystem continuously turns operational experience into reusable skills, updated memories, stronger verification signals, safer policies, and more reliable outcomes.Source: §1 · p. 6
• The survey’s two-dimensional framework tracks cognitive-core evolution from Chatbot to Thinking LLM and task execution from Agent to OpenClaw, with Workspace + Skill connecting chatbots to digital colleagues.This framing shifts the central question from generating better answers to reliably transforming user intent into completed work.Source: §1 · pp. 5–6

2.1.1 Scaling-Driven Language Generation and Parametric Knowledge Compression

• Next-token pretraining and scale made language a universal task interface, compressing linguistic, factual, and commonsense regularities into parameters for fluent, low-latency responses and prompt-based adaptation.This parametric knowledge is flexible and quickly invoked, but lossy, static after training, and difficult to audit or update precisely.Source: §2.1.1 · pp. 8–9

2.1.2 Alignment, Multimodal Expansion, and the Limits of Fast-Response Cognition

• Fast autoregressive generation lacks intrinsic verification, lookahead, and backtracking, leaving complex mathematics, coding, and long-horizon reasoning fragile.Alignment improved instruction following, and multimodal models expanded perception; however, neither removed the fast-response ceiling. A 1.3B InstructGPT model outperformed 175B GPT-3 in human evaluations.Source: §2.1.2 · pp. 12–13

2.2 The Thinking LLM Era: Reasoning and Reinforcement-Learning-Driven Cognition

• Thinking LLMs generate extended deliberation to decompose problems, explore alternatives, verify intermediate results, and correct errors before answering.Long CoT combines sustained reasoning, multi-path exploration, and reflection that earlier systems needed external structures or iterative multi-call frameworks to provide.Source: §2.2 · p. 13, §2.2.1 · p. 14

2.2.1 Long Chain-of-Thought and Inference-Time Scaling

• Long CoT enables deep derivations, branching among approaches, and error correction within one output, while inference-time scaling can improve capability but excessive chains can degrade accuracy.With sufficient inference-time compute, a 1B model can surpass a 405B model on mathematical benchmarks; DeepSeek-R1’s reported “Aha moment” remains contested.Source: §2.2.1 · pp. 14–15

2.2.2 Reinforcement-Learning-Driven Reasoning and Unified Cognitive Systems

• Reasoning training progressed from imitation of CoT traces and external search scaffolds toward self-evolution through verifiable rewards and multi-stage SFT-RL workflows.LIMO elicited strong mathematical reasoning from 817 curated examples, and s1 reported comparable results with 1,000 samples; RLVR uses verifiable answer and format checks, while GRPO removes PPO’s value model.Source: §2.2.2 · pp. 15–16
• Reinforcement learning internalized deep, branching, self-correcting reasoning in o1 and R1, and later systems enabled fast/slow reasoning modes that adjust effort to task complexity.Qwen3 introduced seamless thinking/non-thinking mode switching.Source: §2.2.2 · p. 15
• Reasoning has shifted from standalone model products toward a tunable capability in general-purpose systems, with tool use and environmental interaction providing rudiments of agentic behavior.The progression includes tool invocation during reasoning, hybrid thinking modes, and unified reasoning-dialogue systems.Source: §2.2.2 · pp. 16, 18
• Real-world autonomy still requires persistent state, environmental feedback, and multi-step execution beyond single-turn reasoning.Reinforcement learning internalizes search and reflection into the model, but these capabilities alone do not supply those requirements.Source: §2.2.2 · p. 18

3.1 The Agent Era: Environment–Action–Feedback Loops

• The Agent era extends single-turn answering into observe–reason–act–feedback loops, organized around perception, planning, memory, and tool use.ReAct [7] established a canonical Thought–Action–Observation pattern, while agent frameworks formalize these capabilities through modular or cognitive architectures.Source: §3.1 · p. 18, §3.1.1 · p. 19

3.1.1 Tool Invocation and Core Agent Capabilities

• Agent perception has expanded from textual APIs to screenshot-based GUI interaction, but remains fragmented across snapshots; planning now spans reasoning search, decomposition, reflection, and reinforcement learning.Early multimodal pipelines delegated perception to specialist models, whereas vision-language and GUI agents can interpret screenshots directly; persistent visual working memory remains limited.Source: §3.1.1 · p. 20
• Agent memory distinguishes semantic, episodic, and procedural functions and is shifting toward adaptive experience management; MEM1 reports 3.5× performance improvement and 3.7× lower memory usage.Mem0 reports 91% lower p95 latency and over 90% token cost savings compared with full-context baselines; the passage describes a broader move toward adaptive long-horizon memory.Source: §3.1.1 · p. 21
• Tool use progresses from executable calls to API grounding and trajectory-level control, exposing the need to move beyond isolated calls toward policies that manage tool use over time.Tool use also progresses toward standardized infrastructure; the Agent-era boundary is that actions leave effects fragmented across transient responses rather than persistent workspaces.Source: §3.1.1 · pp. 21–22

3.1.2 Initial Autonomy: Evaluation and Structural Bottlenecks

• Across agent benchmarks, success decays super-linearly with complexity and horizon length, while WebArena reports 14% success for GPT-4; structural bottlenecks include fragmented perception, ephemeral calls, environmental brittleness, and weak task closure.The paper frames these as architectural limits: agents query external environments rather than inhabit persistent workspaces, and lack durable state, recovery, and verification.Source: §3.1.2 · p. 23

3.2 The OpenClaw Era: Persistent Workspaces for Task Closure

• OpenClaw-style systems make the environment a persistent host for files, sessions, logs, tools, permissions, and skills, shifting the goal from attempting actions to delivering an inspectable, governed final state.Actions become workspace operations whose effects can be inspected, validated, or rolled back, supporting durable task closure and final-state verification.Source: §3.2 · pp. 25–26

3.2.1 Workspace Intelligence and Skill-Based Task Closure

• Workstation agents package skills as reusable file-system units, building on Voyager [51] and formalized in Agent Skills as dynamically loaded folders with instructions, scripts, and resources.Modular skills can encode expertise but also become stale, over-specific, incompatible, or malicious, making provenance, versioning, dependencies, and sandboxing important requirements.Source: §3.2.1 · pp. 26–27

3.2.2 Evaluation, Reliability, and Governance Challenges

• Workstation intelligence changes what counts as evaluation, and reliable task closure requires agents to inspect actions, repair failures, validate outputs, and deliver artifacts through an auditable execution harness.SWE-bench [58] and Terminal-Bench [792] assess real repository and command-line work; OpenClaw’s significance is productizing a workstation stack rather than introducing a new reasoning algorithm.Source: §3.2.1 · p. 27, §3.2.2 · p. 28
• OSWorld [49], WebArena and WorkArena [48] [50], and TheAgentCompany evaluate agents on verifiable final states across realistic, stateful environments and consequential professional tasks.These benchmarks target robust long-horizon execution under asynchronous, partially observable conditions, rather than language understanding or isolated tool selection.Source: §3.2.2 · p. 28
• Reliability requires evaluating repeated execution, perturbation robustness, predictability, recoverability, bounded error severity, and trajectory-wide evidence that the final state meets user intent.Computer-use agents can succeed once and fail on a repeated run because execution is stochastic, task specifications are underspecified, and environmental changes compound over long trajectories [794]; reliability dimensions [795] and verifier design [796] address these challenges.Source: §3.2.2 · p. 28
• Persistent credentials, files, permissions, communication channels, and memory expand the attack surface; security evaluations identify risks from poisoned capabilities and system-level failures [797] [798].Taming OpenClaw [799] and Don’t Let the Claw Grip Your Hand [800] address OpenClaw threats, while OS-Harm evaluates misuse, indirect prompt injection, data exfiltration, and system-level harm.Source: §3.2.2 · p. 28
• Runtime governance enforces policy at action boundaries, while forensics must reconstruct traces across the agent loop rather than inspect a single prompt-response pair.OpenClaw PRISM [65] covers lifecycle stages including tool execution and persistence; ClawGuard [66] checks file, command, network, and skill operations against user-confirmed constraints and audit policies; forensic reconstruction is also needed [802].Source: §3.2.2 · p. 28
• OpenClaw-style systems organize work around persistent workspaces and reusable skills, making task closure, verification, skill provenance, permission control, and runtime governance central requirements.The integrated harness combines persistent workspaces, modular skills, closed-loop execution, verification, and runtime governance.Source: §3.2.2 · p. 29

4.1 Workspace as the Execution Substrate for Agentic Work: Stateful Context for Tasks

• Persistent workspaces extend isolated tool calls into long-horizon task delivery by preserving artifacts, logs, and recoverable execution state.Complex work requires failures to be diagnosable, commands rerunnable, and final workspace states inspectable, capabilities Figure 7 contrasts with simple tool invocation.Source: §4.1 · p. 29, Figure 7 · p. 30

4.1.1 From Ephemeral Tool Calls to Persistent State

• Durable files, terminals, histories, repositories, and memory let work accumulate coherently, while workspace design determines what persists and can be audited.For software and knowledge work, editable and checkable artifacts make task completion repairable and handoff-ready rather than merely plausible.Source: §4.1.1 · p. 30

4.1.2 From Answer Generation to Authorized Work Delegation

• Workspace agents shift interaction from prompting every step to delegating bounded objectives with permissions, constraints, success criteria, and risk limits.Users supervise progress and intermediate traces, intervene or roll back when needed, and accept final artifacts; autonomy can vary with task risk and reversibility.Source: §4.1.2 · p. 31

4.2 Skills for Workspace Agents: Reusable Procedures for Repeatable Work

• Skills package procedures, scripts, examples, dependencies, checks, rollback strategies, and safety constraints, making reusable capability the procedural half of the paradigm.Their value becomes fully visible when executed inside persistent workspaces.Source: §4.2 · p. 32

4.2.1 From Ad-hoc Prompts to Composable Capability Packages

• Skills externalize task-family procedures into maintainable, composable packages that can be parameterized, combined, refined, inspected, and reused across tasks.Executable scripts and templates reduce dependence on fragile natural-language reasoning; skills translate task intent into repeatable workspace procedures.Source: §4.2.1 · pp. 32–33

4.2.2 From Skill Libraries to Integrated Digital Workers

• Workspace + Skill systems combine persistent artifacts and context with reusable procedures, then execute and verify work to produce an inspectable final state.OpenClaw illustrates this pattern through skill activation, workspace actions, and verification; reliable use also requires permission boundaries, provenance, sandboxing, audit logs, and rollback. Figure 8 depicts the combined paradigm.Source: Figure 8 · p. 32, §4.2.2 · pp. 33–34
• Reusable skills can become brittle under environmental drift or mislead through overfitting, while persistent workspaces can accumulate stale, inconsistent, or misleading state.Skills need versioning and regression tests; workspace hygiene, provenance labeling, snapshots, and rollback help preserve trustworthy state.Source: §4.2.2 · pp. 34–35
• Skills and workspaces also introduce supply-chain risks and operational overhead, making permission control, sandboxing, recovery, and trajectory-level evaluation necessary for reliable deployment.Evaluation must judge trajectories and final workspace states, not only single responses; the paradigm shifts the bottleneck from prompt design to scalable system operations.Source: §4.2.2 · pp. 34–35

5 Part IV: Data & Evaluation — Paradigm Shifts Behind the Scenes

• Training data and evaluation must adapt to systems that reason over long traces, orchestrate tools, modify environments, and produce verifiable final states.Static text corpora and answer-level benchmarks remain useful for linguistic fluency but are inadequate for assessing these capabilities [874].Source: §5 · p. 35
• Training signals should shift from flat instruction-response pairs to state–action–observation trajectories that capture tool outputs, UI states, workspace changes, and final-state evidence.Figure 9 contrasts static prompt–response pairs with reasoning traces and trajectories for agentic and OpenClaw-style systems.Source: §5 · p. 35, Figure 9 · p. 36
• Evaluation should move beyond semantic similarity to prioritize end-to-end task closure, system reliability, and operational safety under diverse real-world conditions [882] [883].Source: §5 · p. 35

5.1.3 Agent & OpenClaw Era: State–Action–Observation Trajectories

• Agent and OpenClaw data record interactive state–action–observation trajectories, including tool outputs, interface states, environment feedback, and final-state evidence.Workspace traces can also include terminal errors, file changes, browser states, screenshots, permissions, and task history.Source: §5.1 · p. 36, §5.1.3 · p. 39

5.1.1 Chatbot Era: Human-Annotated Dialogue Data (SFT)

• Chatbot-era systems learned from static text and human-annotated instruction–response pairs, and were evaluated on static inputs for accuracy, preference, and instruction following.These measures tested whether models could speak fluently and answer as requested, not whether they could complete a production task end to end.Source: §5.1.1 · pp. 37–38

5.1.2 Thinking LLM Era: Chain-of-Thought and Process Reward Data (CoT / PRM)

• Thinking LLM data extend beyond final answers to reasoning paths, including intermediate steps, revisions, checks, and process-level feedback.Process reward models supervise individual reasoning steps, unlike outcome reward models that score the complete solution; evaluation also tests reasoning-path reliability.Source: §5.1.2 · pp. 38–39
• Thinking LLM evaluation expands from answer correctness to difficult knowledge and reasoning tasks, code-pass rates, changing benchmarks, and assessment of faulty-step detection.ProcessBench and PRMBench evaluate whether models identify faulty reasoning steps and whether process reward models are reliable.Source: §5.1.2 · p. 39

5.1.3 Agent & OpenClaw Era: State–Action–Observation Trajectories

• OpenClaw adds reusable skill assets—procedures, checklists, and recovery steps—to the workspace, making learned procedures and experience part of task execution.Skills should specify versions, dependencies, preconditions, postconditions, failure cases, rollback behavior, and permissions, and remain stable as environments change.Source: §5.1.3 · p. 40
• Agent evaluation requires realistic executable workspaces, trajectory logs, final-state verification, and measures of task success, reproducibility, execution reliability, efficiency, and trajectory-level safety.Simulated services and snapshot/restore support reproducible testing while limiting irreversible effects on live services.Source: §5.1.3 · pp. 40–41

Challenges

• Verifiable action trajectories are scarce and costly to annotate, can become outdated as tools and interfaces change, and require a trade-off between simulation fidelity and safety.Synthetic tasks and self-play can scale data generation, but without state verification and safety filters they may reinforce incorrect habits, unsafe tool use, or false completion.Source: Challenges · p. 42

5.2.1 Stage I: Final-Output Accuracy and Answer Correctness

• Final-output evaluation scores answers with accuracy or reference-based metrics, scaling well for clear-label tasks but failing to reveal whether the reasoning process was valid.MMLU uses multiple-choice accuracy, while GSM8K and MATH assess mathematical reasoning by final-answer correctness [402] [414] [933]; BLEU and ROUGE compare generated text with references.Source: §5.2.1 · p. 44

5.2.2 Stage II: LLM-as-Judge for Long-Chain Reasoning and Process Verification

• Reasoning evaluation combines harder math and science problems, executable-code tests, and LLM judges or process verifiers that inspect long traces and intermediate errors.ProcessBench and PRMBench test detection of incorrect intermediate steps [990], while LiveBench and ARC-AGI-2 emphasize dynamic, abstraction-oriented evaluations to reduce contamination [988] [989].Source: §5.2.2 · p. 45

5.2.3 Stage III: Agent & OpenClaw Era: Task Closure Rate

• Agentic evaluation measures task closure: whether tool use transforms an initial environment into the intended final state, not merely whether an answer or reasoning trace appears correct.SWE-bench tests whether patches resolve real GitHub issues and pass tests; WebShop and WebArena test completion of interactive web tasks [50] [58] [996].Source: §5.2.3 · p. 46

5.2.4 Stage IV: Workspace/OpenClaw Capability and Safety

• Workspace/OpenClaw evaluation measures task success, reliability, efficiency, reproducibility, and safety, including risks from privacy leaks, harmful side effects, and malicious skills.ClawBench evaluates online task completion, while ClawsBench measures capability and safety in simulated workspaces [999]. Auditing requires fixed initial states, trajectory logs, replayable actions, and final-state diffs [1108] [1110] [1111].Source: §5.2.4 · pp. 47–48

• The trajectory from language generation through reasoning and tool use leads to workspace-level execution in digital environments.Source: §6 · p. 48
• Greater autonomy broadens failure risks to include unsafe tool calls, corrupted workspace states, incomplete task closure, and untraceable long-horizon behavior.Source: §6 · p. 48
• The section identifies reliable autonomy as the central challenge and proposes engineering models, contexts, tools, skills, workspaces, and governance as an integrated whole.It frames this direction as the development of self-evolving AI ecosystems.Source: §6 · p. 48

6.1.1 Long-Horizon Reliability and Task Closure

• Long-horizon agents can propagate tool-call errors, leave inconsistent workspaces, and discover planning mistakes only after irreversible actions, so task closure needs monitoring, verification, repair, and rollback.Composed skills can produce syntactically valid but semantically unsuitable outputs, or fail silently and leave misleading traces.Source: §6.1.1 · pp. 48–49

6.1.2 Safety, Governance, and Permission Boundaries

• Workspace safety is operational: agents can leak data, overwrite files, trigger side effects, execute untrusted skills, or make unauthorized decisions, requiring permissions, validation, audit trails, rollback, and oversight.Governance must define resource access, approval requirements, and logging or sandboxing while keeping high-impact actions inspectable and controllable.Source: §6.1.2 · p. 49

6.1.3 Human–AI Collaboration Ethics and Data Boundaries

• Workspace agents may widen access to professional work while compressing apprenticeship, accelerating work rhythms, shifting labor toward oversight, and homogenizing creativity through templates.They can also make exploration and rapid prototyping cheaper; systems should preserve human agency, attribution, contestability, and escalation. Data governance requires tenant isolation, minimization, purpose limitation, provenance, retention controls, and rules for trajectory reuse; organizational know-how should remain governed as an asset [65] [66] [802].Source: §6.1.3 · p. 50

6.1.4 Memory, Context, and Persistent State

• Persistent memory must maintain goals and context across sessions, while avoiding the cost, retrieval difficulty, and instability of relying solely on ultra-long context windows.Memory can take token-level, parametric, latent, or external forms; it serves working, factual, experiential, and procedural functions, and needs lifecycle management to form, update, retrieve, and discard memories without accumulating stale assumptions or privacy risks [1112].Source: §6.1.4 · pp. 50–51

6.2 Future Directions: Toward Self-Evolving AI Ecosystems

• Self-evolution is an engineered feedback loop that converts operational experience into validated system assets, with the model as cognitive core and surrounding layers enabling experience to accumulate.Trajectories can become memories, repeated workflows skills, failures regression tests, tool errors wrapper updates, and safety incidents permission rules.Source: §6.2 · p. 51, Figure 12 · p. 52

6.2.1 From Prompt Engineering to Harness Engineering

• Prompts specify intent, context supplies task-relevant state, and harnesses shape consequential actions; future progress therefore depends on execution substrates that capture, test, and govern experience.Longer context windows do not remove the need for curation and can let irrelevant history bury decision-critical state; harnesses should make actions inspectable, constrained, replayable, and learnable.Source: §6.2.1 · pp. 52–53

6.2.2 AI-Native Workspaces as Digital Embodiment

• Persistent workspaces make cognition actionable and shift evaluation from plausible answers to verified state transitions, while snapshots, replay, rollback, permissions, and provenance also enable learning.Final-state diffs and evaluator hooks help determine whether an update is beneficial; shared workspaces externalize experience for accumulation, auditing, and reuse.Source: §6.2.2 · p. 53

6.2.3 Beyond-Gradient Learning and Continual System Maintenance

• Beyond-gradient learning lets systems improve executable assets—wrappers, memories, skills, tests, and permissions—without replacing weight learning, but requires provenance, compatibility, lifecycle management, and curation.Experience must become reusable assets small enough to retrieve, structured enough to test, general enough to reuse, and safe enough to deploy; systems must also decide what to retain, merge, quarantine, or verify.Source: §6.2.3 · p. 54

6.2.4 Composable Skill and Multi-Agent Ecosystems

• Reusable skills need contracts and software-style lifecycle and security disciplines, while multi-agent ecosystems need orchestration, explicit authority, shared-workspace coordination, and accountability to avoid unsafe delegation and activity without task closure.Skill contracts specify schemas, preconditions, side effects, permissions, failure modes, and validation; orchestration assigns roles and routes work while governance sets authority scopes and audit obligations.Source: §6.2.4 · pp. 54–55

6.2.5 Self-Evolving Systems: From Chatbot to Digital Colleague

• Governed self-evolution turns operational traces into validated, versioned, reversible updates, enabling persistent digital colleagues to accumulate project memory, improve shared tools, and participate in workflows.The loop observes traces, diagnoses outcomes, abstracts patterns, validates updates in sandboxes or tests, deploys under permission constraints, monitors effects, and rolls back degraded changes.Source: §6.2.5 · pp. 55–56

Reasoning work

• Prior surveys cover prompting, inference-time scaling, process supervision, reinforcement-learning-driven reasoning, self-improvement, and reflection, while this paper focuses on operational reasoning with tools and persistent states.Chain-of-Thought and Tree/Graph-of-Thought variants showed the value of intermediate reasoning for multi-step problem solving [15] [31] [412] [421] [464]; recent work studies inference-time scaling and related reasoning methods [30] [166] [493], alongside reflection feedback loops [32] [33].Source: §7 · p. 56

Agent architectures and tools

• Wang et al. [4] and Xi et al. [5] survey broad agent architectures, while tool-use studies establish an observe-plan-act loop that this paper contrasts with durable workspace operation.Related work examines long-horizon planning, memory, and safeguards [768], cognitive frameworks [8] [638], reasoning with external steps [7], API invocation and tool use [17] [18] [649] [732], and open-ended skill accumulation [51].Source: §7 · p. 57

Interactive benchmarks

• Interactive benchmarks span general agent abilities, browser and desktop tasks, software engineering, and stateful tool use, motivating evaluation by task closure rather than plausible text alone.Examples include WebGPT, WebShop, Mind2Web, and WebArena [50] [996] [997]; OSWorld and WorkArena [48] [49]; SWE-bench, SWE-agent, and OpenHands [46] [47] [58]; and APIBank, ToolAlpaca, ToolLLM, ToolSandbox, and -bench [18] [895] [992] [1000].Source: §7 · p. 57

Novelty

• Existing work treats reasoning, action loops, memory, feedback, and verification largely as separate ingredients; this paper re-examines them through a workspace-centered perspective.Files, terminals, browser sessions, logs, permissions, snapshots, and skill assets jointly shape what agents can perceive, modify, verify, and recover; durable workspaces make reliability, provenance, rollback, permissions, and governance core architectural concerns.Source: §7 · p. 57

• Thinking LLMs advance beyond next-token “fast thinking” by leveraging inference-time computation.Source: §8 · p. 57
• OpenClaw-style workstation systems replace ad hoc tool-calling with persistent workspaces, skills, and governance.The “Workspace + Skill” paradigm supports state persistence, reusable procedures, and task closure.Source: §8 · p. 57
• Data and evaluation shift from instruction-response pairs and static benchmarks toward State-Action-Observation trajectories and auditable, self-evolving ecosystems.The conclusion identifies persistent environments, reusable skills, and safety governance as requirements for reliable digital colleagues.Source: §8 · p. 57