# Research Gap Dossier: Voluntariness, Adversarial State, Incentives, and Human-Agent Authority

**White paper:** *Intelligent Voluntary Cooperation*  
**Prepared:** 30 August 2026  
**Scope:** The four priority gaps identified in `research-evidence-dossier.md`.  
**Source rule:** Primary research papers, canonical original works, official standards, and first-party security research only.  
**System boundary:** This dossier does not make claims about the present Qi or IXO implementation. It establishes evidence, architectural implications, normative commitments, and tests against which system facts can later be assessed.

---

## 1. Epistemic key

- **[E] Established evidence:** a replicated, field-defining, or canonical result or distinction, used only within its applicable domain.
- **[X] Named experimental result:** an observation from a specified study under its reported conditions.
- **[I] Architectural inference:** a conclusion drawn for the proposed architecture; it is not a result established by the cited source.
- **[H] Hypothesis:** a proposition that must be tested against a comparator and can be falsified.
- **[N] Normative principle:** a value-based design or governance commitment, stated as such rather than disguised as science.

`[S]` is reserved for verified Qi or IXO system facts. There are no `[S]` claims in this dossier.

## 2. Executive synthesis

The gap research changes four parts of the white paper's argument.

1. **Voluntary is not synonymous with decentralized or unassigned.** **[E]** Research ethics, organizational theory, and platform-labor studies distinguish formal choice from choice made under inadequate information, unequal authority, hidden control, dependence, or costly exit. **[I]** Intelligent voluntary cooperation therefore requires observable conditions for comprehension, refusal, scoped delegation, voice, contestation, and exit—not merely an absence of a central task allocator.
2. **Persistent shared state is both a coordination medium and a supply chain.** **[X]** Indirect prompt injection, agent-memory poisoning, and RAG-knowledge corruption studies show that a small quantity of attacker-controlled retrieved state can redirect later model behavior and actions. **[E]** Distributed-systems research separately shows that convergent state can still be stale, causally surprising, or adversarial. **[I]** Persistence, convergence, provenance, truth, freshness, authorization, and safety must remain separate properties.
3. **A shared digital environment is not automatically a commons.** **[E]** Public goods, common-pool resources, club goods, private goods, markets, and firms are different analytical categories. Resource characteristics—especially exclusion and subtractability—are distinct from governance regimes and coordination mechanisms. **[I]** Each Topic or resource must be classified before selecting access, incentive, monitoring, pricing, or sanction mechanisms.
4. **Putting a human “in the loop” does not settle authority or produce complementarity.** **[X]** Human-AI studies report overreliance, unequal benefits from safeguards, task-dependent reversals, and bias introduced through interaction. Algorithmic-management studies show that systems advertised as enabling flexibility can exercise substantial indirect control through information and evaluation asymmetries. **[I]** A credible architecture must specify who sets goals, recommends, decides, executes, evaluates, bears liability, and can correct the result, then measure realized influence rather than relying on nominal role labels.

These findings support a bounded thesis:

> **[I]** Persistent shared environments can enable voluntary cooperation only when participation conditions, authority, state quality, resource characteristics, and correction pathways are made explicit. **[H]** A governed shared-state architecture will outperform naive shared memory on coordination quality and attack recovery while preserving more participant initiative than a centrally controlled alternative. This requires comparative evidence; it is not established by the architecture itself.

---

## 3. Meaningful voluntariness, consent, exit, delegation, and power

### 3.1 A defensible definition

No primary source establishes one universal checklist for “voluntary cooperation.” The following is therefore an explicit synthesis.

| Condition | Evidence posture | Minimum operational meaning |
|---|---|---|
| Legible invitation | **[I]/[N]** | The participant can identify the cooperative scope, requesting principal, intended use, material risks, alternatives, and expected consequences. |
| Comprehension | **[E]** in informed-consent settings; **[N]** here | Information is presented in a form the participant can understand, and comprehension can be checked rather than assumed from a click. |
| Freedom from coercion and undue influence | **[E]** as an official research-ethics distinction; **[N]** for Qi | Refusal does not trigger threats, loss of unrelated entitlements, retaliation, or exploitative inducements that overwhelm meaningful choice. |
| Genuine refusal | **[I]/[N]** | The participant can decline a contribution or delegation before consequence without hidden execution continuing elsewhere. |
| Scoped delegation | **[E]** as an organizational and capability distinction; **[N]** here | Authority has an identifiable delegator, delegate, resource, action, constraints, term, and revocation or expiry path. |
| Voice and contestation | **[E]** as a canonical organizational response to decline; **[N]** here | Participants can challenge state, decisions, rules, and sanctions through an effective process, not only submit feedback. |
| Meaningful exit | **[I]/[N]** | Leaving is technically possible, understandable, and not made illusory by undisclosed lock-in, loss of unrelated rights, impossible data extraction, or punitive retaliation. Legitimate prior commitments can survive exit if known in advance. |
| Power visibility | **[E]** that formal and effective control can diverge; **[N]** here | The system exposes who can change rules, rank information, allocate work, withhold information, evaluate performance, impose sanctions, or make consequences final. |
| Continuing consent | **[I]/[N]** | Material changes to purpose, authority, risk, membership, or consequence trigger renewed notice and, where appropriate, renewed authorization. |

**[I]** Voluntariness should be evaluated as a property of a participant's available choices under actual dependencies, not as a property asserted by the system operator. **[H]** Participants shown a machine-readable and human-readable cooperation contract will demonstrate higher comprehension, more accurate refusal, and fewer unintended delegations than participants shown ordinary terms or chat instructions.

### 3.2 Primary-source dossiers

#### V1. Belmont Report — information, comprehension, voluntariness, coercion, and undue influence

**Citation:** National Commission for the Protection of Human Subjects of Biomedical and Behavioral Research. *The Belmont Report: Ethical Principles and Guidelines for the Protection of Human Subjects of Research*. U.S. Department of Health and Human Services, 18 April 1979. [Official text](https://www.hhs.gov/ohrp/regulations-and-policy/belmont-report/read-the-belmont-report/index.html) · [PDF](https://www.hhs.gov/ohrp/sites/default/files/the-belmont-report-508c_FINAL.pdf)

**Supports:** **[E]** Within human-subjects research, informed consent is analyzed through information, comprehension, and voluntariness. The report distinguishes coercion—threat of harm to obtain compliance—from undue influence through excessive, improper, or vulnerability-sensitive inducement. It also treats authority relationships and threatened withdrawal of otherwise available services as relevant to whether a choice is genuinely voluntary. **[N]** A cooperative system should disclose purpose, risk, alternatives, authority, and withdrawal conditions in a comprehensible form and should not condition unrelated entitlements on participation.

**Limitations:** This is an official ethical framework for research involving human subjects, not a general law of organizational membership, employment, contracting, or autonomous agents. Applying it to digital cooperation is a normative transfer.

**Chapters:** 7, 9, 13, 15, 21.

#### V2. Hirschman — exit, voice, and loyalty are interacting responses, not interchangeable controls

**Citation:** Albert O. Hirschman. *Exit, Voice, and Loyalty: Responses to Decline in Firms, Organizations, and States*. Harvard University Press, 1970. ISBN 978-0-674-27660-4. See also Hirschman's original follow-up: “Exit, Voice, and Loyalty: Further Reflections and a Survey of Recent Contributions.” *Social Science Information* 13, no. 1 (1974): 7–26. [https://doi.org/10.1177/053901847401300101](https://doi.org/10.1177/053901847401300101)

**Supports:** **[E]** Canonical distinction: dissatisfied members can leave or attempt to change an objectionable condition; loyalty affects the timing and balance of these responses. Exit can discipline an organization, but it can also drain the members best positioned to exercise voice. Voice has cost, can be collective, and depends on whether it can affect the organization. **[I]** Offering an “exit” button is not sufficient if exit is prohibitively costly, and exit is not a substitute for correction and contestation within a shared world.

**Limitations:** Hirschman's framework is analytical and intentionally broad. Relationships among exit, voice, and loyalty vary by institution; later tests do not support one simple universal causal pattern.

**Chapters:** 7, 10–11, 15, 19.

#### V3. Aghion and Tirole — formal authority and effective control can diverge

**Citation:** Philippe Aghion and Jean Tirole. “Formal and Real Authority in Organizations.” *Journal of Political Economy* 105, no. 1 (1997): 1–29. [https://doi.org/10.1086/262063](https://doi.org/10.1086/262063) · [Author manuscript](https://discovery.ucl.ac.uk/id/eprint/17678/)

**Supports:** **[E]** Canonical formal model distinguishing formal authority—the right to decide—from real authority—the effective control of decisions produced by information and initiative. Delegating real authority can increase an agent's initiative while reducing a principal's control; formal integration can coexist with substantial effective decentralization. **[I]** Agent systems need to record both formal permission and realized influence. A human who formally signs a decision may lack real authority if the agent controls the information, options, timing, or default.

**Limitations:** This is an economic model, not direct empirical evidence about LLM agents, consent, or software authorization. Its assumptions and principal-agent framing do not exhaust cooperative relationships.

**Chapters:** 7, 9, 12–13, 19.

#### V4. Falk and Kosfeld — control can reduce voluntary effort

**Citation:** Armin Falk and Michael Kosfeld. “The Hidden Costs of Control.” *American Economic Review* 96, no. 5 (2006): 1611–1630. [https://doi.org/10.1257/aer.96.5.1611](https://doi.org/10.1257/aer.96.5.1611)

**Supports:** **[X]** In an experimental principal-agent game, many agents reduced performance when the principal imposed a minimum performance requirement. Participants who reacted negatively commonly described control as distrust and limitation of autonomy; the payoff effect of control was non-monotonic. **[I]** Controls can change the social meaning of cooperation as well as the feasible action set, so the architecture should compare the protective value of a control with its effect on initiative and trust.

**Limitations:** A stylized laboratory game with heterogeneous responses. It does not show that control is generally harmful or that high-consequence systems should relax necessary safeguards.

**Chapters:** 7, 10–11, 19, 21.

#### V5. Lee et al. — algorithmic management changes autonomy and sensemaking

**Citation:** Min Kyung Lee, Daniel Kusbit, Evan Metsky, and Laura Dabbish. “Working with Machines: The Impact of Algorithmic and Data-Driven Management on Human Workers.” In *Proceedings of CHI 2015*, 1603–1612. [https://doi.org/10.1145/2702123.2702548](https://doi.org/10.1145/2702123.2702548)

**Supports:** **[X]** A qualitative field study of Uber and Lyft drivers documented how algorithmic systems assigned work, supplied information, evaluated performance, and prompted workers to use online forums for collective sensemaking about opaque behavior. **[I]** An environment can exercise managerial control through rankings, work allocation, data visibility, evaluation, and defaults even when it describes participation as autonomous.

**Limitations:** Early, case-specific research on ride-hailing platforms and the algorithms of that period. Qualitative findings identify mechanisms and experiences rather than population-level causal effects.

**Chapters:** 4, 7, 10, 14–15, 19.

#### V6. Rosenblat and Stark — information asymmetry can turn nominal flexibility into indirect control

**Citation:** Alex Rosenblat and Luke Stark. “Algorithmic Labor and Information Asymmetries: A Case Study of Uber's Drivers.” *International Journal of Communication* 10 (2016): 3758–3784. [Official journal PDF](https://ijoc.org/index.php/ijoc/article/download/4892/1739/21331) · [https://doi.org/10.2139/ssrn.2686227](https://doi.org/10.2139/ssrn.2686227)

**Supports:** **[X]** A nine-month empirical study found that the platform exercised substantial indirect control while promoting drivers as independent and flexible. The authors identify information and power asymmetries as central to how the application structured workers' choices and behavior. **[I]** Voluntariness claims must include information access, dependency, unilateral rule change, evaluation power, and sanction power—not merely freedom to accept individual tasks.

**Limitations:** A case study of one labor platform, not a test of all platforms or open cooperation. The SSRN DOI identifies the deposited version; the journal itself does not assign a DOI.

**Chapters:** 7, 10–11, 15, 19.

### 3.3 White-paper claim boundaries

**Safe claims**

- **[E]** Formal permission and effective control can diverge because information and initiative shape real authority.
- **[X]** In specific organizational and platform settings, controls, opaque information, work allocation, evaluation, and sanction mechanisms have changed perceived autonomy and behavior.
- **[I]** “Voluntary” should be treated as an auditable institutional property comprising informed choice, refusal, bounded delegation, voice, power visibility, and meaningful exit.

**Claims to avoid**

- “Decentralized coordination is voluntary by definition.”
- “Consent to join authorizes all later uses of a participant's data, artifacts, identity, or agent.”
- “A human final click proves meaningful human authority.”
- “Exit alone protects participants from capture or abuse.”

---

## 4. Adversarial shared state and persistent-memory risk

### 4.1 Threat model

Persistent environmental traces create a repeatable influence path:

```text
attacker or low-quality contributor
        ↓ writes or induces
artifact / memory / rating / state update
        ↓ is indexed, ranked, replicated, or inherited
later human or agent retrieves it
        ↓ interprets data as evidence or instruction
decision or tool call
        ↓ may persist further derived state
downstream participants inherit the contamination
```

The relevant safety properties are distinct:

| Property | What it establishes | What it does not establish |
|---|---|---|
| Integrity | State has not changed outside the permitted write process. | The state is true, current, safe, or legitimately authorized. |
| Convergence | Replicas that receive the same updates reach the same value. | The updates were benign, causally suitable, or semantically correct. |
| Provenance | A history of asserted entities, activities, derivations, and agents can be represented. | The asserted history is complete or truthful. |
| Freshness | State satisfies a specified age or validation policy. | The content is true or fit for a particular consequence. |
| Reputation | Past feedback is aggregated under a rule. | Identities are unique, ratings are independent, or the rule resists capture. |
| Authorization | An actor is permitted to request a scoped action. | The action is wise, factually grounded, or safe in the current context. |
| Evaluation | A method produced a result from evidence. | The method, evidence, or evaluator is uncompromised. |
| Consensus | Participants or replicas agree under stated assumptions. | The agreed proposition is true or legitimate outside those assumptions. |

**[I]** Shared artifacts consumed by an LLM must be treated as both content and potentially executable influence. **[N]** Reading an artifact and exercising consequential authority should be separate capabilities, mediated by independent policy and validation.

### 4.2 Primary-source dossiers

#### A1. Greshake et al. — indirect prompt injection through retrieved data

**Citation:** Kai Greshake, Sahar Abdelnabi, Shailesh Mishra, Christoph Endres, Thorsten Holz, and Mario Fritz. “Not What You've Signed Up For: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection.” arXiv:2302.12173, 2023. [https://doi.org/10.48550/arXiv.2302.12173](https://doi.org/10.48550/arXiv.2302.12173) · [Proof-of-concept repository](https://github.com/greshake/llm-security)

**Supports:** **[X]** The authors demonstrated that attacker-authored instructions placed in data likely to be retrieved could alter integrated LLM applications without the attacker directly interacting with the model. Demonstrated impacts included manipulation of application behavior, API use, data theft, and information-ecosystem contamination. **[I]** A page, message, file, or Topic artifact cannot be trusted as passive data merely because it entered through an authorized read path.

**Limitations:** Early preprint and proof-of-concept demonstrations against a small set of then-current systems; it does not quantify prevalence or prove that no robust separation mechanism can exist.

**Chapters:** 6, 10, 12–13, 15–18.

#### A2. AgentPoison — very sparse memory poisoning can create persistent targeted behavior

**Citation:** Zhaorun Chen, Zhen Xiang, Chaowei Xiao, Dawn Song, and Bo Li. “AgentPoison: Red-Teaming LLM Agents via Poisoning Memory or Knowledge Bases.” *Advances in Neural Information Processing Systems 37* (NeurIPS 2024). [Proceedings](https://proceedings.neurips.cc/paper_files/paper/2024/hash/eb113910e9c3f6242541c1652e30dfd6-Abstract-Conference.html) · [https://doi.org/10.52202/079017-4136](https://doi.org/10.52202/079017-4136)

**Supports:** **[X]** Across three tested agent settings, the reported attack poisoned less than 0.1% of memory or knowledge entries while achieving average attack success of at least 80% and reducing benign performance by no more than 1%. The attack operated on stored memory or knowledge rather than model training. **[I]** High aggregate task performance is compatible with a persistent, triggerable compromise, so benign accuracy is not an adequate memory-integrity test.

**Limitations:** Optimized triggers, constructed agent applications, and substantial attacker knowledge in core settings, although transfer was also studied. Results do not establish these rates for arbitrary models, memories, or defenses.

**Chapters:** 10, 12–13, 16–18, 21.

#### A3. AgentDojo — tool-returned data can cross into consequential action

**Citation:** Edoardo Debenedetti, Jie Zhang, Mislav Balunović, Luca Beurer-Kellner, Marc Fischer, and Florian Tramèr. “AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents.” NeurIPS 2024 Datasets and Benchmarks Track. [Proceedings](https://proceedings.neurips.cc/paper_files/paper/2024/hash/97091a5177d8dc64b1da8bf3e1f6fb54-Abstract-Datasets_and_Benchmarks_Track.html) · [https://doi.org/10.52202/079017-2636](https://doi.org/10.52202/079017-2636)

**Supports:** **[X]** AgentDojo evaluated prompt injection in four simulated tool environments using 97 tasks and 629 security cases. Untrusted data returned by tools could redirect agent behavior. Tool filtering reduced tested attacks but was inadequate where the legitimate and malicious tasks required the same tools. **[N]** Tool availability alone is too coarse an authority boundary; the requested action, resource, purpose, provenance, and consequence need independent checks.

**Limitations:** Simulated environments, attacks, models, and defenses current to 2024. Benchmark performance is not a production security guarantee.

**Chapters:** 10, 12–13, 17–18, 21.

#### A4. Instruction-data separation benchmark — models do not reliably preserve the distinction

**Citation:** Egor Zverev, Sahar Abdelnabi, Soroush Tabesh, Mario Fritz, and Christoph H. Lampert. “Can LLMs Separate Instructions From Data? And What Do We Even Mean By That?” *International Conference on Learning Representations* (ICLR 2025). [Proceedings](https://proceedings.iclr.cc/paper_files/paper/2025/hash/a77eadda332b6d4a9ae1e0e4024555f2-Abstract-Conference.html) · [https://doi.org/10.48550/arXiv.2403.06833](https://doi.org/10.48550/arXiv.2403.06833)

**Supports:** **[X]** The paper formalizes instruction-data separation and reports that tested models did not achieve high separation on its benchmark; tested prompting and fine-tuning approaches either failed to remove the problem or reduced utility. **[I]** System-level privilege separation is necessary because a model-level label saying “this is data” is not, by itself, an authority control.

**Limitations:** A single-turn formalization and benchmark. It is not an impossibility proof, and later models or different architectures may perform differently.

**Chapters:** 10, 12–13, 17.

#### A5. PoisonedRAG — a few corrupted texts can dominate retrieval-grounded answers

**Citation:** Wei Zou, Runpeng Geng, Binghui Wang, and Jinyuan Jia. “PoisonedRAG: Knowledge Corruption Attacks to Retrieval-Augmented Generation of Large Language Models.” In *34th USENIX Security Symposium*, 3827–3844, 2025. [USENIX paper](https://www.usenix.org/conference/usenixsecurity25/presentation/zou-poisonedrag) · [PDF](https://www.usenix.org/system/files/usenixsecurity25-zou-poisonedrag.pdf)

**Supports:** **[X]** Across multiple datasets, retrievers, eight LLMs, and three applications, the authors induced attacker-selected answers by inserting a small number of malicious texts. The paper reports roughly 90% attack success when five malicious texts per target question were inserted into databases containing millions of texts; tested paraphrasing and perplexity defenses were insufficient. **[I]** Retrieval relevance and source trust must be separate signals, and high-consequence answers need cross-source or evaluator checks outside the generator.

**Limitations:** Targeted question-answer attacks with a defined ability to insert records. Exact results depend on corpus, retriever, generator, threat knowledge, and target. It does not test Qi or collaborative artifacts.

**Chapters:** 10, 12, 16–18, 21.

#### A6. NIST adversarial-ML taxonomy — poisoning is a lifecycle threat, not only a training attack

**Citation:** Apostol Vassilev, Alina Oprea, Alie Fordyce, Hyrum Anderson, Xander Davies, and Michael Hamin. *Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations*. NIST AI 100-2e2025. National Institute of Standards and Technology, 2025. [https://doi.org/10.6028/NIST.AI.100-2e2025](https://doi.org/10.6028/NIST.AI.100-2e2025)

**Supports:** **[E]** Official taxonomy distinguishing attacker goals, knowledge, capability, lifecycle stage, poisoning, evasion, privacy, and generative-AI abuse. **[I]** The white-paper threat model should identify the write path, retrieval path, model path, evaluator path, and consequence path separately rather than treating “AI safety” as one control domain.

**Limitations:** Taxonomy and guidance, not empirical validation that any listed mitigation is sufficient for a particular system.

**Chapters:** 10, 12–13, 17–18, 21.

#### A7. Douceur — identity multiplicity can defeat redundancy and reputation

**Citation:** John R. Douceur. “The Sybil Attack.” In *Peer-to-Peer Systems: First International Workshop, IPTPS 2002*, 251–260. [Microsoft Research](https://www.microsoft.com/en-us/research/publication/the-sybil-attack/) · [https://doi.org/10.1007/3-540-45748-8_24](https://doi.org/10.1007/3-540-45748-8_24)

**Supports:** **[E]** Canonical formal result: if one faulty entity can present many identities, it can undermine redundancy; absent a logically centralized certification authority, generally preventing Sybil attacks requires extreme resource-parity and coordination assumptions. **[I]** Counting identities, signatures, contributions, ratings, or votes is not evidence of independent support unless the relevant uniqueness or cost assumption is explicit.

**Limitations:** The model concerns peer-to-peer entities and resource tests. It does not imply that civil identity is always required or that one centralized identity provider is the only legitimate design. Privacy-preserving and domain-specific admission systems create different trade-offs.

**Chapters:** 7, 9–11, 15, 17, 20.

#### A8. Friedman and Resnick — cheap pseudonyms create whitewashing and newcomer costs

**Citation:** Eric J. Friedman and Paul Resnick. “The Social Cost of Cheap Pseudonyms.” *Journal of Economics & Management Strategy* 10, no. 2 (2001): 173–199. [https://doi.org/10.1111/j.1430-9134.2001.00173.x](https://doi.org/10.1111/j.1430-9134.2001.00173.x)

**Supports:** **[E]** In a repeated random-matching model, replaceable identities let actors escape reputational consequences. Cooperation can re-emerge through conventions that distrust newcomers, but this imposes a social entry cost; entry fees create a different exclusion cost. The authors analyze free but unreplaceable pseudonyms as an alternative. **[I]** Reputation design must price identity reset and account for the resulting burden on legitimate newcomers; “portable reputation” and “right to start fresh” are in tension.

**Limitations:** Formal model with specific matching, lifetime, mistake, and identity assumptions. It does not identify an optimal real-world identity or reputation regime.

**Chapters:** 7, 10–11, 15, 19–20.

#### A9. Muchnik, Aral, and Taylor — visible ratings can produce self-reinforcing information pollution

**Citation:** Lev Muchnik, Sinan Aral, and Sean J. Taylor. “Social Influence Bias: A Randomized Experiment.” *Science* 341, no. 6146 (2013): 647–651. [https://doi.org/10.1126/science.1240466](https://doi.org/10.1126/science.1240466)

**Supports:** **[X]** In a large-scale randomized field experiment on a social news site, a randomly added positive rating increased the probability of subsequent positive ratings by 32% and raised final ratings by 25% on average. Negative manipulation was more often corrected, and effects varied by topic and relationship. **[I]** Popularity, prior acceptance, and reputation signals can be endogenous causes of later acceptance; they should not be treated as independent evidence.

**Limitations:** One platform, one rating mechanism, and asymmetric context-specific effects. It does not show that all aggregation produces herding or that the exact magnitudes transfer to agent-generated artifacts.

**Chapters:** 6, 8, 10–11, 15–16, 20.

#### A10. Terry et al. — weakly consistent state needs session guarantees

**Citation:** Douglas B. Terry, Alan J. Demers, Karin Petersen, Mike J. Spreitzer, Marvin M. Theimer, and Brent B. Welch. “Session Guarantees for Weakly Consistent Replicated Data.” In *Proceedings of the Third International Conference on Parallel and Distributed Information Systems*, 140–149, 1994. [https://doi.org/10.1109/PDIS.1994.331722](https://doi.org/10.1109/PDIS.1994.331722)

**Supports:** **[E]** The paper defines read-your-writes, monotonic reads, writes-follow-reads, and monotonic writes to prevent user-visible anomalies in weakly replicated systems while retaining important availability benefits. **[I]** A participant or agent should not authorize a consequence from a view older than one it has already observed, or publish a dependent write where the dependency is absent. “Current enough” must be expressed relative to an action's consequence.

**Limitations:** Technical session consistency, not epistemic truth, semantic validity, or permission. Stronger guarantees can reduce availability or disconnected operation.

**Chapters:** 10, 15, 17–18.

#### A11. Shapiro et al. — CRDT convergence is a narrow property

**Citation:** Marc Shapiro, Nuno Preguiça, Carlos Baquero, and Marek Zawirski. “Conflict-Free Replicated Data Types.” In *Stabilization, Safety, and Security of Distributed Systems*, 386–400, 2011. [Author PDF](https://perso.lip6.fr/Marc.Shapiro/papers/2011/CRDTs_SSS-2011.pdf) · [https://doi.org/10.1007/978-3-642-24550-3_29](https://doi.org/10.1007/978-3-642-24550-3_29)

**Supports:** **[E]** Under specified conditions, CRDT replicas that receive the same updates converge deterministically without foreground synchronization. **[I]** Convergence cannot be used as evidence that an update was authorized, truthful, fresh, non-malicious, or safe to merge into consequential state.

**Limitations:** The paper establishes data-type convergence properties, not application semantics, governance, confidentiality, provenance, or adversarial source validation.

**Chapters:** 9–10, 15, 17.

#### A12. PROV-DM — revision and invalidation can be represented without erasing lineage

**Citation:** Luc Moreau and Paolo Missier, eds. *PROV-DM: The PROV Data Model*. W3C Recommendation, 30 April 2013. [https://www.w3.org/TR/2013/REC-prov-dm-20130430/](https://www.w3.org/TR/2013/REC-prov-dm-20130430/)

**Supports:** **[E]** The standard represents entities, activities, agents, attribution, delegation, derivation, revision, use, generation, and invalidation. Invalidation marks the cessation or expiry of an entity; revision relates a new entity to an original. **[N]** Correction should normally create an attributable revision or invalidation that preserves historical lineage and identifies which downstream artifacts or decisions may need reassessment.

**Limitations:** A descriptive vocabulary, not a trust engine, access-control system, consensus protocol, proof that asserted provenance is true, or automatic downstream-revocation mechanism.

**Chapters:** 6, 9–10, 13, 16–18.

#### A13. RFC 9111 — stale use should be explicit and policy-bound

**Citation:** Roy Fielding, Mark Nottingham, and Julian Reschke. *HTTP Caching*. RFC 9111, June 2022. [https://www.rfc-editor.org/rfc/rfc9111.html](https://www.rfc-editor.org/rfc/rfc9111.html)

**Supports:** **[E]** The standard distinguishes fresh and stale responses, defines validation, and prohibits serving stale responses under specified directives or absent disconnection, permission, or policy. **[I]** The architecture should attach freshness, expiry, and revalidation semantics to artifacts whose use depends on time, and should make emergency or disconnected stale use visible.

**Limitations:** HTTP cache semantics are not a general epistemology. Age and validation against an origin do not establish truth, legitimacy, or suitability for a consequence.

**Chapters:** 6, 10, 15, 17–18.

#### A14. Lamport, Shostak, and Pease — agreement under adversarial behavior requires explicit assumptions

**Citation:** Leslie Lamport, Robert Shostak, and Marshall Pease. “The Byzantine Generals Problem.” *ACM Transactions on Programming Languages and Systems* 4, no. 3 (1982): 382–401. [https://doi.org/10.1145/357172.357176](https://doi.org/10.1145/357172.357176)

**Supports:** **[E]** Canonical result: reliable agreement is constrained by the number of faulty participants and whether messages can be forged; with oral messages, more than two-thirds of participants must be loyal in the paper's model. **[I]** A system must state its adversary, identity, authentication, quorum, and failure assumptions before describing shared state as agreed or trustworthy.

**Limitations:** Formal distributed-computing model, not a social theory of legitimacy, truth, or human disagreement. Cryptographic authentication does not prove honest observation or good judgment.

**Chapters:** 10, 12–13, 17–18.

#### A15. Saltzer and Schroeder — least privilege and complete mediation are consequence controls

**Citation:** Jerome H. Saltzer and Michael D. Schroeder. “The Protection of Information in Computer Systems.” *Proceedings of the IEEE* 63, no. 9 (1975): 1278–1308. [MIT author page](https://web.mit.edu/Saltzer/www/publications/protection/) · [https://doi.org/10.1109/PROC.1975.9939](https://doi.org/10.1109/PROC.1975.9939)

**Supports:** **[E]** Canonical security principles include fail-safe defaults, complete mediation, separation of privilege, and least privilege. **[N]** Every consequential action should pass an authority check at the moment of use; an agent reading a shared artifact should receive only the action authority required for the approved operation.

**Limitations:** General computer-security principles, not empirical evidence that any particular LLM-agent defense works.

**Chapters:** 9–10, 12–13, 17–18.

#### A16. Abrams, McGrew, and Plotkin — reputation aggregation can reward strategic misreporting

**Citation:** Zoe Abrams, Robert McGrew, and Serge Plotkin. “A Non-Manipulable Trust System Based on EigenTrust.” *ACM SIGecom Exchanges* 5, no. 4 (2005): 21–30. [Author PDF](https://www.sigecom.org/exchanges/volume_5/5.4-abrams.pdf) · [https://doi.org/10.1145/1120717.1120721](https://doi.org/10.1145/1120717.1120721)

**Supports:** **[E]** In the paper's model, raters whose own trust scores depend on their reports may rationally misreport rather than reveal their observations. The authors prove non-manipulability for a cyclic-partition construction under stated assumptions. **[I]** Reputation is an incentive-bearing decision mechanism, not neutral metadata. Evaluators, beneficiaries, anchors, aggregation rules, and conflicts of interest must be visible and separable.

**Limitations:** A single-round formal model with myopic, selfish peers and a non-strategic calculation center. The construction is not a general proof that EigenTrust, reputation, or any named production mechanism resists coalition, Sybil, or long-horizon manipulation.

**Chapters:** 10–11, 18, 20.

#### A17. Yu et al. — Sybil resistance depends on explicit social-graph assumptions

**Citation:** Haifeng Yu, Phillip B. Gibbons, Michael Kaminsky, and Feng Xiao. “SybilLimit: A Near-Optimal Social Network Defense Against Sybil Attacks.” In *2008 IEEE Symposium on Security and Privacy*, 3–17. [Author PDF](https://www.comp.nus.edu.sg/~yuhf/yuh-sybillimit.pdf) · [https://doi.org/10.1109/SP.2008.13](https://doi.org/10.1109/SP.2008.13)

**Supports:** **[E]/[X]** SybilLimit bounds accepted Sybil identities when the honest social graph mixes rapidly and the number of attack edges between honest and Sybil regions is limited; the authors examined relevant assumptions using Friendster, LiveJournal, and DBLP graphs available to them. **[I]** “Social proof” or graph position is not a universal uniqueness proof. A Sybil-defense claim must state its trust anchors, attack-edge assumption, admission rule, and expected burden on legitimate newcomers and peripheral participants.

**Limitations:** Guarantees depend on the graph and adversary satisfying the model. Social centrality can encode existing inequality and is not evidence of civil identity, independence, competence, or legitimacy.

**Chapters:** 7, 10–11, 15, 20.

#### A18. Doyle — correction requires re-evaluating dependent beliefs

**Citation:** Jon Doyle. “A Truth Maintenance System.” *Artificial Intelligence* 12, no. 3 (1979): 231–272. [https://doi.org/10.1016/0004-3702(79)90008-0](https://doi.org/10.1016/0004-3702(79)90008-0) · [Author publication page](https://www.csc2.ncsu.edu/faculty/jdoyle2/publications/)

**Supports:** **[E]** The paper's symbolic system records justifications for beliefs and revises dependent beliefs after assumptions are contradicted. **[I]** Correcting or invalidating a source artifact without identifying and re-evaluating the claims, evaluations, decisions, and actions derived from it leaves contamination active. A revision protocol therefore needs a dependency graph and a policy for reopening descendants, not only a new version label.

**Limitations:** A centralized symbolic truth-maintenance architecture, not a distributed, adversarial, probabilistic, privacy-preserving, or authorization system. Human and model judgments will not always have complete machine-readable dependencies.

**Chapters:** 10, 12, 18, 21.

#### A19. W3C Bitstring Status List — suspension, revocation, refresh, and replacement are distinct

**Citation:** World Wide Web Consortium. *Bitstring Status List v1.0*. W3C Recommendation, 15 May 2025. [https://www.w3.org/TR/vc-bitstring-status-list/](https://www.w3.org/TR/vc-bitstring-status-list/)

**Supports:** **[E]** The standard distinguishes status purposes including irreversible revocation and reversible suspension, and discusses refresh or replacement rather than treating all lifecycle change as deletion. It identifies freshness, caching, correlation, availability, and status-authority considerations for verifiers. **[N]** A persistent environment should represent suspension, revocation, supersession, expiry, and replacement as different states with explicit authorities and verifier policies.

**Limitations:** A credential-status mechanism, not a general artifact lifecycle, truth oracle, or automatic correction system. A status assertion is controlled by an issuer or status authority; freshness and acceptance remain verifier-policy questions.

**Chapters:** 10, 17–18.

#### A20. W3C ActivityStreams Tombstone — logical retirement can preserve identity and deletion context

**Citation:** World Wide Web Consortium. *Activity Vocabulary*. W3C Recommendation, 23 May 2017, section “Tombstone.” [https://www.w3.org/TR/activitystreams-vocabulary/#dfn-tombstone](https://www.w3.org/TR/activitystreams-vocabulary/#dfn-tombstone)

**Supports:** **[E]** The vocabulary defines a tombstone for content that existed but has been deleted, with properties for former type and deletion time. **[N]** Logical retirement should preserve enough stable identity and context for references, disputes, synchronization, and audit while preventing retired content from silently remaining active.

**Limitations:** A data vocabulary with no access control, propagation guarantee, privacy solution, causal-consistency rule, or safe garbage-collection algorithm. Legal erasure and minimization obligations can require stronger deletion than an indefinitely retained tombstone.

**Chapters:** 10, 15, 17–18.

#### A21. W3C DID Core — identity, keys, control, credentials, and reputation have different lifecycles

**Citation:** World Wide Web Consortium. *Decentralized Identifiers (DIDs) v1.0: Core Architecture, Data Model, and Representations*. W3C Recommendation, 19 July 2022. [https://www.w3.org/TR/2022/REC-did-core-20220719/](https://www.w3.org/TR/2022/REC-did-core-20220719/)

**Supports:** **[E]** The standard distinguishes DID document update, verification-method and key rotation or revocation, and DID deactivation, with detailed behavior delegated to DID methods. It does not claim that one DID equals one human or independent actor. **[I]** Identity, controller keys, delegated capabilities, credentials, reputation, membership, and resource authority must not share one undifferentiated lifecycle. A valid controller proof does not establish uniqueness, competence, independence, or permission for an unrelated consequence.

**Limitations:** An identifier architecture, not a Sybil-resistance, civil-identity, governance, authorization, reputation, or factual-verification protocol. Concrete guarantees are method- and deployment-dependent.

**Chapters:** 7, 9–11, 15, 17–18.

### 4.3 Correction and retirement requirements

**[N]** Persistent state should support all of the following without pretending that deletion and auditability are always compatible:

1. stable identity for each artifact version;
2. provenance and source trust distinct from content popularity;
3. explicit status: draft, active, challenged, superseded, invalidated, expired, quarantined, or deleted where legally and technically possible;
4. validity windows and revalidation triggers for time-sensitive state;
5. challenge and appeal with an identified decision authority;
6. revision lineage that never silently overwrites the evidence used by an earlier decision;
7. dependency tracking so invalidating an ancestor identifies derived artifacts and consequences for review;
8. quarantine before deletion when evidence preservation, incident response, or dispute obligations apply;
9. bounded rollback or compensating action for consequences that cannot literally be reversed;
10. measurable propagation and recovery: time to detection, affected descendants, time to contain, time to correct, and residual exposure.

**[H]** Provenance, status, dependency tracking, and least-authority execution will reduce poisoned-state propagation and unauthorized consequence rates compared with unversioned shared memory. **Falsifier:** the controls add delay and complexity without materially improving detection, containment, correction, or outcome quality under matched attacks.

---

## 5. Incentives and economic coordination categories

### 5.1 Do not confuse resource type, property regime, and coordination mechanism

**[E]** Goods are commonly distinguished using difficulty of exclusion and subtractability or rivalry of use. **[I]** These characteristics describe a resource in a context; they do not determine one governance regime.

| Resource type | Exclusion | Subtractability / congestion | Central problem | Plausible digital examples | Governance implication |
|---|---|---|---|---|---|
| Private good | Feasible | High | Exchange, contract, externalities | A single-use entitlement, exclusive service capacity | Property/contract and market or firm allocation may fit. |
| Public good | Difficult | Low | Under-provision and free riding | A reusable open protocol specification; non-sensitive public knowledge | Contribution incentives and collective funding matter more than depletion control. |
| Common-pool resource | Difficult | High | Overuse, congestion, depletion, pollution | Scarce shared compute, attention, moderation capacity, a rate-limited public service | Boundaries, monitoring, appropriation rules, congestion pricing, and sanctions may fit. |
| Club or toll good | Feasible | Low until congestion | Membership size, exclusion cost, congestion | A members-only knowledge base or under-capacity shared service | Membership rules, fees, quotas, and governance of club size may fit. |

Three corrections follow:

- **[E]** A common-pool resource is not the same as common property. The first describes resource characteristics; the second is an institutional arrangement.
- **[E]** A market and a firm are coordination mechanisms or institutional forms, not fifth and sixth resource types.
- **[I]** One Topic can contain resources of several types. For example, an open method may be public-good-like, its compute budget common-pool-like, a credential private, and a review service club-like.

### 5.2 Primary-source dossiers

#### I1. Samuelson — pure collective consumption

**Citation:** Paul A. Samuelson. “The Pure Theory of Public Expenditure.” *Review of Economics and Statistics* 36, no. 4 (1954): 387–389. [https://doi.org/10.2307/1925895](https://doi.org/10.2307/1925895)

**Supports:** **[E]** Canonical formalization of collective-consumption goods whose use by one person does not reduce the quantity available to others, and of the distinct condition for efficient public-good provision. **[I]** Reusable knowledge may be non-subtractable in consumption while its production, verification, hosting, and governance remain scarce.

**Limitations:** A highly abstract model of pure public expenditure; practical goods often have mixed exclusion, congestion, quality, and jurisdictional properties.

**Chapters:** 7, 11, 19–20.

#### I2. Buchanan — clubs occupy the space between pure private and public goods

**Citation:** James M. Buchanan. “An Economic Theory of Clubs.” *Economica* 32, no. 125 (1965): 1–14. [https://doi.org/10.2307/2552442](https://doi.org/10.2307/2552442)

**Supports:** **[E]** Canonical theory of cooperative membership in which exclusion is possible and members share benefits subject to crowding or congestion; optimal membership depends on costs and the effect of additional members on utility. **[I]** Bounded Topic membership may create a club good rather than a commons, especially where access is controlled and use is non-rival until attention, compute, or moderation capacity congests.

**Limitations:** Formal welfare analysis, not a governance blueprint. Member heterogeneity, power, exclusion justice, privacy, and dynamic rule change require additional treatment.

**Chapters:** 7, 11, 15, 19–20.

#### I3. Ostrom — exclusion and subtractability are separate from institutional rules

**Citation:** Elinor Ostrom. *Understanding Institutional Diversity*. Princeton University Press, 2005. [https://doi.org/10.2307/j.ctt7s7wm](https://doi.org/10.2307/j.ctt7s7wm) · [Publisher chapter](https://assets.press.princeton.edu/chapters/s8085.pdf)

**Supports:** **[E]** The Institutional Analysis and Development framework separates attributes of the biophysical/material world, community, and rules-in-use. Ostrom uses exclusion and subtractability to distinguish four broad goods and stresses that rule combinations affect incentives and outcomes in context. **[I]** A Topic's resource classification, participants, rules, action situation, and outcome criteria should be recorded separately.

**Limitations:** General institutional framework, not an algorithm and not evidence that one set of digital rules will be optimal.

**Chapters:** 7, 9, 11, 13, 15, 19.

#### I4. Coase — firms and markets are alternative coordination arrangements

**Citation:** R. H. Coase. “The Nature of the Firm.” *Economica* 4, no. 16 (1937): 386–405. [https://doi.org/10.1111/j.1468-0335.1937.tb00002.x](https://doi.org/10.1111/j.1468-0335.1937.tb00002.x)

**Supports:** **[E]** Canonical account: market coordination through the price mechanism has transaction costs, while firms substitute conscious direction by an entrepreneur for some market transactions; firm boundaries depend on the comparative cost of organizing additional transactions. **[I]** Environmental coordination, messaging, markets, and hierarchical orchestration should be compared by discovery, contracting, communication, monitoring, and failure costs—not ranked ideologically.

**Limitations:** Foundational theory from 1937, not a complete modern theory of organizations, digital platforms, or agent coordination. Transaction-cost reasoning does not determine legitimacy.

**Chapters:** 2, 7, 11, 19, 21.

#### I5. Holmström — team production creates free-riding and monitoring problems

**Citation:** Bengt Holmström. “Moral Hazard in Teams.” *Bell Journal of Economics* 13, no. 2 (1982): 324–340. [https://doi.org/10.2307/3003457](https://doi.org/10.2307/3003457)

**Supports:** **[E]** In a formal multi-agent production model, individual effort is not directly contractible and joint output creates free-riding; the principal and informative monitoring have incentive roles. Competition is valuable in the model only insofar as it supplies information for relative evaluation. **[I]** Shared outcomes require attribution and evaluation mechanisms that do not assume joint success proves each contribution or that competition is intrinsically productive.

**Limitations:** Formal principal-agent model with strong assumptions. Real teams have social preferences, communication, identity, intrinsic motivation, unequal power, and multi-dimensional outcomes.

**Chapters:** 7, 11, 18–20.

#### I6. Fehr and Gächter — punishment opportunities can sustain cooperation but are costly

**Citation:** Ernst Fehr and Simon Gächter. “Cooperation and Punishment in Public Goods Experiments.” *American Economic Review* 90, no. 4 (2000): 980–994. [https://doi.org/10.1257/aer.90.4.980](https://doi.org/10.1257/aer.90.4.980)

**Supports:** **[X]** In repeated laboratory public-goods experiments, contribution patterns differed substantially when participants could punish others at personal cost; punishment supported higher cooperation under the reported conditions. **[I]** Sanction affordances can alter contribution incentives even without a central enforcer, but sanction legitimacy, capture, proportionality, and appeal must be designed separately.

**Limitations:** Laboratory public-goods games with experiment-specific endowments, identities, information, and punishment rules. Costly punishment can destroy welfare, target cooperators, or reflect retaliation in other settings.

**Chapters:** 7, 10–11, 19, 21.

#### I7. Ostrom, Walker, and Gardner — communication and self-chosen sanctions can change commons outcomes

**Citation:** Elinor Ostrom, James Walker, and Roy Gardner. “Covenants With and Without a Sword: Self-Governance Is Possible.” *American Political Science Review* 86, no. 2 (1992): 404–417. [https://doi.org/10.2307/1964229](https://doi.org/10.2307/1964229) · [Indiana University manuscript](https://dlc.dlib.indiana.edu/dlcrest/api/core/bitstreams/0afb3404-76da-473c-ae70-b0f489e84624/content)

**Supports:** **[X]** In laboratory common-pool-resource experiments, opportunities for communication and participant-designed sanctioning materially changed appropriation and joint outcomes; self-governance did not require an external sovereign in every treatment. **[I]** Rule participation and communication may be productive components of environmental coordination rather than overhead to eliminate.

**Limitations:** Controlled small-group resource games. The result does not imply that informal covenants scale, resist collusion, or replace legal authority in high-consequence settings.

**Chapters:** 7, 10–11, 15, 19, 21.

#### I8. Gneezy and Rustichini — a fine can reframe a norm as a price

**Citation:** Uri Gneezy and Aldo Rustichini. “A Fine Is a Price.” *Journal of Legal Studies* 29, no. 1 (2000): 1–17. [https://doi.org/10.1086/468061](https://doi.org/10.1086/468061)

**Supports:** **[X]** In a field experiment at day-care centers, introducing a monetary fine for late pickup significantly increased late pickup, and removing it did not restore the prior behavior. The authors argue that the fine changed information and the meaning of an incomplete social contract. **[I]** Token rewards, prices, penalties, and reputation points can crowd out or reframe norms; incentive changes require behavioral testing and rollback plans.

**Limitations:** One field setting; the study does not isolate a single universal mechanism, and later work has examined alternative explanations and boundary conditions.

**Chapters:** 7, 10–11, 19, 21.

### 5.3 Design consequences

**[N]** Before attaching tokens, prices, voting, reputation, or sanctions to a Topic, record:

- the resource or outcome being governed;
- its exclusion and subtractability characteristics in this context;
- whether the problem is provision, appropriation, congestion, quality, discovery, coordination, or externality;
- who bears cost, receives benefit, observes contribution, and can leave;
- whether contribution is intrinsically, socially, contractually, or financially motivated;
- the expected gaming path and who can change the incentive rule;
- which institution—market, firm, club, commons governance, public provision, or hybrid—is being used and why.

**[H]** Incentive mechanisms matched to the actual resource problem will produce higher verified contribution and lower gaming than uniform token rewards. **Falsifier:** matched mechanisms do not improve outcome quality or create governance overhead and exclusion costs that offset the gain.

---

## 6. Heterogeneous human-agent collaboration and unequal authority

### 6.1 Evidence boundary

SwarmWorld studies initially equivalent model agents in a simulator. The central Qi case is different: people, organizations, models, agents, and devices have unequal expertise, information, rights, dependencies, liability, and sanction exposure. Existing human-AI research does not yet validate open-ended intelligent voluntary cooperation. It does establish several reasons not to assume complementarity or meaningful human control.

### 6.2 Primary-source dossiers

#### H1. Lubars and Tan — people prefer different delegation levels by task

**Citation:** Brian Lubars and Chenhao Tan. “Ask Not What AI Can Do, But What AI Should Do: Towards a Framework of Task Delegability.” NeurIPS 2019. [Proceedings PDF](https://papers.nips.cc/paper/8301-ask-not-what-ai-can-do-but-what-ai-should-do-towards-a-framework-of-task-delegability.pdf) · [https://doi.org/10.48550/arXiv.1902.03245](https://doi.org/10.48550/arXiv.1902.03245)

**Supports:** **[X]** In a survey built around 100 tasks, respondents showed little preference for full AI control and a stronger preference for machine-in-the-loop configurations where humans led. Perceived trust was the factor most correlated with preferred delegation among motivation, difficulty, risk, and trust. **[I]** Delegation should be task-, participant-, and consequence-sensitive rather than fixed by a universal autonomy level.

**Limitations:** Stated preferences in a survey, not observed long-term collaboration or proof that preferred allocations perform better. The task set and 2019 understanding of AI bound generalization.

**Chapters:** 7, 9, 12–13, 19, 21.

#### H2. Horvitz — mixed initiative requires explicit uncertainty and control trade-offs

**Citation:** Eric Horvitz. “Principles of Mixed-Initiative User Interfaces.” In *Proceedings of CHI 1999*, 159–166. [https://doi.org/10.1145/302979.303030](https://doi.org/10.1145/302979.303030) · [Microsoft Research PDF](https://www.microsoft.com/en-us/research/wp-content/uploads/2016/11/chi99horvitz.pdf)

**Supports:** **[E]** Canonical mixed-initiative design: automated services and direct human manipulation should be coupled using models of uncertainty, value, timing, attention, and user control. **[I]** Initiative is not binary. A system can let an agent discover, recommend, draft, schedule, execute, or verify while allocating final authority differently at each stage.

**Limitations:** Design principles and a scheduling system, not current generative-agent evidence or a governance theory for unequal parties.

**Chapters:** 2, 7–9, 12–15, 19.

#### H3. Bansal et al. — explanations can increase reliance without improving complementarity

**Citation:** Gagan Bansal, Tongshuang Wu, Joyce Zhou, Raymond Fok, Besmira Nushi, Ece Kamar, Marco Tulio Ribeiro, and Daniel S. Weld. “Does the Whole Exceed Its Parts? The Effect of AI Explanations on Complementary Team Performance.” In *Proceedings of CHI 2021*. [https://doi.org/10.1145/3411764.3445717](https://doi.org/10.1145/3411764.3445717) · [Microsoft Research PDF](https://www.microsoft.com/en-us/research/wp-content/uploads/2021/02/does_the_whole_exceed_its_parts-chi21.pdf)

**Supports:** **[X]** Across mixed-method studies on three datasets with AI accuracy selected to be comparable to humans, AI assistance sometimes produced complementary improvement, but explanations did not increase it. Explanations increased acceptance of AI recommendations whether they were correct or wrong. **[I]** Explanation display is not an authority safeguard; evaluation should measure selective reliance and end-to-end team performance.

**Limitations:** Specific classification and question-answering tasks, explanation forms, participant populations, and models. It does not show that explanations are generally harmful.

**Chapters:** 8, 10, 12–13, 18–19, 21.

#### H4. Buçinca, Malaya, and Gajos — reducing overreliance has usability and inequality trade-offs

**Citation:** Zana Buçinca, Maja Barbara Malaya, and Krzysztof Z. Gajos. “To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI in AI-Assisted Decision-Making.” *Proceedings of the ACM on Human-Computer Interaction* 5, CSCW1, Article 188 (2021). [https://doi.org/10.1145/3449287](https://doi.org/10.1145/3449287) · [Author page](https://www.eecs.harvard.edu/~kgajos/papers/2021/bucinca2021trust.shtml)

**Supports:** **[X]** In an experiment with 199 participants, cognitive-forcing designs reduced overreliance compared with simple explainable-AI conditions, but the designs that reduced overreliance most received worse subjective ratings. Benefits were greater on average for participants higher in need for cognition. **[I]** A safety interaction can distribute burden unequally, and preference or trust scores are not substitutes for appropriate reliance.

**Limitations:** One experimental task and participant sample. Need for cognition is not a complete account of expertise, disability, language, power, or access.

**Chapters:** 7–8, 10, 12–13, 19, 21.

#### H5. Green and Chen — algorithm-in-the-loop interaction can worsen accuracy and disparity

**Citation:** Ben Green and Yiling Chen. “Disparate Interactions: An Algorithm-in-the-Loop Analysis of Fairness in Risk Assessments.” In *Proceedings of the Conference on Fairness, Accountability, and Transparency*, 90–99, 2019. [https://doi.org/10.1145/3287560.3287563](https://doi.org/10.1145/3287560.3287563) · [Author PDF](https://scholar.harvard.edu/files/19-fat.pdf)

**Supports:** **[X]** In a controlled Mechanical Turk experiment using criminal-risk prediction cases, participants underperformed the risk assessment even when shown it, could not reliably assess their own or the tool's accuracy, and interacted with the score in ways that produced higher predictions for Black defendants and lower predictions for White defendants. **[I]** Fairness and authority must be evaluated in the combined human-model process, not only in the model output.

**Limitations:** Crowd participants were not judges, the task used historical criminal-justice data, and the experimental interface does not represent every deployed decision process. The result is not a recommendation to automate bail decisions.

**Chapters:** 7, 10, 12–13, 18–19, 21.

#### H6. Dell'Acqua et al. — AI advantage reverses across a jagged task frontier

**Citation:** Fabrizio Dell'Acqua, Edward McFowland III, Ethan Mollick, Hila Lifshitz-Assaf, Katherine C. Kellogg, Saran Rajendran, Lisa Krayer, François Candelon, and Karim R. Lakhani. “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality.” *Organization Science* 37, no. 2 (2026): 403–423. [https://doi.org/10.1287/orsc.2025.21838](https://doi.org/10.1287/orsc.2025.21838) · [HBS author PDF](https://www.hbs.edu/ris/Publication%20Files/dell-acqua-et-al-2026-navigating-the-jagged-technological-frontier_5c589c8c-fbb5-458f-b285-c944746cd717.pdf)

**Supports:** **[X]** In a preregistered field experiment with 758 consultants, participants with GPT-4 completed 12.2% more tasks and worked 25.1% faster on a set of tasks judged within the tested capability frontier, with higher quality. On a selected task outside that frontier, AI users were 19% less likely to reach the correct solution. **[I]** Delegation should respond to measured task-model capability and consequence; apparent task similarity is not enough.

**Limitations:** One consulting firm, one model generation, constructed task sets, and short-term individual knowledge work. It does not establish durable team structures, voluntary cooperation, or effects on lower-power workers.

**Chapters:** 1–2, 7, 12, 19, 21.

#### H7. Lee et al. and Rosenblat–Stark — human-agent systems can conceal organizational authority

**Citations:** See V5 and V6 above.

**Supports:** **[X]** In platform-labor settings, algorithms performed managerial functions through assignment, information, evaluation, dynamic incentives, and deactivation risk while workers retained nominal flexibility. **[I]** “Agent as assistant” and “participant as autonomous” are insufficient descriptions; the system must disclose who owns the platform, sets defaults, ranks opportunities, observes whom, changes rules, adjudicates disputes, and can exclude participants.

**Limitations:** These are human-platform labor studies, not direct studies of modern generative agents cooperating inside Qi.

**Chapters:** 4, 7, 10, 14–15, 19.

#### H8. Aghion and Tirole — nominal human authority can coexist with agent real authority

**Citation:** See V3 above.

**Supports:** **[E]/[I]** The formal/real-authority distinction gives the white paper a precise way to analyze “human in the loop.” If the agent selects the evidence, frames the options, controls timing, or sets the default, it may hold substantial real authority even where a person retains the formal signature.

**Limitations:** This application to human-agent systems is an inference from organizational theory and requires direct empirical testing.

**Chapters:** 7, 12–13, 18–19.

### 6.3 Authority matrix required for every consequential workflow

**[N]** Record the following for each actor class, including platform operator, organization, human participant, model host, agent, evaluator, and settlement authority:

| Function | Required question |
|---|---|
| Goal setting | Who can declare or change the intended outcome? |
| Context selection | Who decides which state, evidence, and alternatives are visible? |
| Proposal | Who may generate candidate actions or claims? |
| Delegation | Who grants which capability, for how long, and how is it revoked? |
| Decision | Who has the formal right to select or reject? |
| Execution | Who or what can cause an external or irreversible effect? |
| Evaluation | Who selects the method, evidence threshold, and evaluator? |
| Consequence | Which authority changes rights, resources, credentials, or canonical state? |
| Monitoring | Who can observe behavior and whose behavior is invisible? |
| Challenge | Who may contest, pause, appeal, correct, or invalidate? |
| Liability | Who bears loss, sanction, or remediation cost when the combined system fails? |
| Exit | Who can leave, what obligations remain, and what state or reputation can be exported? |

Measure both **formal allocation** and **realized influence** through override rates, default acceptance, information asymmetry, response latency, refusal success, and consequence-bearing error.

**[H]** Explicitly separating proposal, decision, execution, and consequence authority will reduce inappropriate reliance and unauthorized effects without eliminating the productivity benefits of AI assistance. **Falsifier:** separation adds friction but does not change error detection, override quality, or unauthorized-consequence rates.

---

## 7. Integrated claim ledger additions

| ID | Claim | Class | Confidence | Principal support | Required limitation or falsifier | Chapters |
|---|---|---|---|---|---|---|
| G01 | Decentralized or unassigned action is not sufficient to make participation voluntary. | **[E]/[I]** | High for distinction; moderate for architecture | Belmont; Hirschman; Lee; Rosenblat–Stark | Operational definition must predict comprehension, refusal, contestation, or exit outcomes. | 7 |
| G02 | Formal authority and effective control can diverge. | **[E]** | High as organizational distinction | Aghion–Tirole; platform studies | Transfer to a particular human-agent workflow requires observed influence data. | 7, 12–13, 19 |
| G03 | Controls can reduce initiative or reframe voluntary cooperation under some conditions. | **[X]** | Moderate | Falk–Kosfeld; Gneezy–Rustichini | Do not generalize beyond treatment; test control benefits and motivation costs together. | 7, 10–11, 19 |
| G04 | Retrieved artifacts and memory are prompt-injection and poisoning surfaces for LLM agents. | **[X]** | High within tested systems | Greshake; AgentPoison; AgentDojo; PoisonedRAG; Zverev | Later models and defenses may differ; test exact system. | 10, 12–13, 16–18 |
| G05 | Persistence can convert a one-time write into repeated downstream influence. | **[X]/[I]** | Moderate to high | AgentPoison; PoisonedRAG; Muchnik | Measure actual retrieval, inheritance, and consequence pathways. | 6, 10, 16–18 |
| G06 | Convergence, provenance, freshness, authorization, consensus, and truth are distinct properties. | **[E]/[I]** | High | CRDT; Terry; PROV-DM; RFC 9111; Lamport; Doyle; W3C lifecycle standards | An implementation must show which property each mechanism establishes. | 9–10, 12–13, 15, 17–18 |
| G07 | Identity counts and reputation can be captured by Sybils, identity resets, strategic reporting, and social influence. | **[E]/[X]** | High for mechanisms | Douceur; Friedman–Resnick; Muchnik; Abrams–McGrew–Plotkin; SybilLimit | Exact attack cost and effect depend on admission, identity, graph, anchor, incentive, and aggregation assumptions. | 10–11, 15, 20 |
| G08 | Public goods, common-pool resources, club goods, and private goods require different incentive analysis. | **[E]** | High | Samuelson; Buchanan; Ostrom | Mixed and context-dependent resources must not be forced into one quadrant. | 7, 11, 15, 19–20 |
| G09 | Markets and firms are alternative coordination arrangements, not resource types. | **[E]** | High as canonical distinction | Coase | Modern platform and agent settings add power and network effects not resolved by the original theory. | 2, 7, 11, 19 |
| G10 | Incentives, monitoring, communication, and sanctions change cooperation but can create crowding, capture, and welfare costs. | **[X]/[I]** | Moderate | Fehr–Gächter; Ostrom–Walker–Gardner; Falk–Kosfeld; Gneezy–Rustichini | Test the mechanism and total welfare in the target context. | 7, 10–11, 19, 21 |
| G11 | Human-AI teams do not automatically outperform the stronger participant. | **[X]** | High within studied tasks | Bansal; Buçinca; Green–Chen; Dell'Acqua | Generalization is task-, model-, interface-, and population-dependent. | 12, 19, 21 |
| G12 | Explanations and nominal human approval do not ensure appropriate reliance or real authority. | **[X]/[I]** | Moderate to high | Bansal; Buçinca; Green–Chen; Aghion–Tirole | Measure selective reliance, realized influence, and end-to-end outcome. | 7, 12–13, 18–19 |
| G13 | Authority-sensitive, task-specific delegation is a plausible design requirement. | **[I]/[N]** | Moderate | Lubars–Tan; Horvitz; Dell'Acqua | Compare against fixed autonomy levels; assess user burden and error. | 7, 9, 12–13, 19 |
| G14 | A governed shared-state system will contain adversarial propagation better than naive shared memory. | **[H]** | Speculative | Derived from A1–A21 | Falsified if controls fail to reduce spread, consequence, or recovery time under matched attacks. | 10, 17–18, 21 |
| G15 | Meaningful voluntariness controls will improve understanding and legitimate participation without unacceptable friction or exclusion. | **[H]** | Speculative | Derived from V1–V6 | Falsified if comprehension, refusal, voice, or exit do not improve, or participation becomes materially less accessible. | 7, 15, 21 |

---

## 8. Comparative research program to close the gaps

### 8.1 Voluntariness study

**Conditions:** ordinary invitation; disclosure-only; structured cooperation contract; structured contract plus comprehension check and explicit refusal/exit preview.

**Participants:** heterogeneous humans with different expertise and dependency; no consequential field deployment until ethics and legal review are complete.

**Measures:** comprehension, perceived and observed freedom to refuse, accidental delegation, time and cost to revoke, voice usage and response, exit completion, portability, participation burden, and differential effects by language, expertise, disability, income/dependency, and organizational role.

**Primary hypothesis:** **[H]** Structured, revisitable authorization improves comprehension and refusal accuracy.  
**Failure condition:** No improvement, higher exclusion, or formal consent without changed behavior.

### 8.2 Adversarial shared-state benchmark

Seed matched environments with benign, false, stale, popularity-manipulated, Sybil-supported, prompt-injecting, and capability-escalating artifacts.

Compare:

1. unversioned shared memory;
2. provenance and source labels;
3. provenance plus explicit lifecycle/status;
4. provenance, lifecycle, dependency invalidation, and least-authority action mediation;
5. an orchestrated baseline with no shared artifact reuse.

Measure attack success, retrieval rate, time to detection, descendant contamination, unauthorized tool calls, irreversible consequences, correction propagation, false-positive quarantine, recovery time, coordination latency, and benign task quality.

**Primary hypothesis:** **[H]** Lifecycle and authority controls reduce unauthorized consequence more than they reduce benign coordination value.  
**Failure condition:** Material attack propagation remains, or control overhead removes the coordination advantage.

### 8.3 Resource and incentive study

Classify each test environment before intervention: public-good provision, common-pool appropriation, club congestion, private exchange, or mixed.

Compare no explicit incentive, attribution only, social recognition, fixed payment, outcome payment, price, participant-selected rule, monitoring, graduated sanction, and market or firm-style allocation where applicable.

Measure verified outcome, contribution diversity, free riding, overuse, gaming, sanction cost, exclusion, inequality, intrinsic-motivation change, and participant retention.

**Primary hypothesis:** **[H]** Problem-matched institutions outperform a uniform token/reputation mechanism.  
**Failure condition:** Classification does not predict the effective mechanism or produces excessive governance cost.

### 8.4 Human-agent authority study

Cross participant expertise and organizational power with four authority allocations:

1. human acts without AI;
2. AI recommends and human decides;
3. AI drafts or executes only after specific approval;
4. AI executes under pre-delegated bounded capability with retrospective audit.

Vary whether the model is inside or outside its measured task frontier and whether explanations, confidence, source provenance, or cognitive forcing are shown.

Measure team quality against both solo baselines, error complementarity, override precision, inappropriate reliance, decision time, realized information influence, unauthorized consequence, perceived autonomy, contestation, and unequal effects across participant groups.

**Primary hypothesis:** **[H]** Task- and consequence-sensitive authority outperforms fixed autonomy.  
**Failure condition:** No complementarity, systematic overreliance, benefits concentrated among already advantaged participants, or formal human authority with unchanged realized model control.

---

## 9. White-paper-ready conclusions

### Established or experimentally demonstrated

- **[E]** Voluntary consent in a canonical ethics framework requires information, comprehension, and freedom from coercion and undue influence; importing that framework to digital cooperation is normative.
- **[E]** Formal authority can differ from real authority, and exit, voice, and loyalty are interacting institutional responses.
- **[X]** In specific experimental and platform settings, control, opaque information, incentives, rankings, and algorithms changed initiative, behavior, and power.
- **[X]** Retrieved text, knowledge stores, and persistent agent memory can carry prompt-injection and poisoning attacks into later behavior.
- **[E]** Convergent replicated state is not necessarily fresh, authorized, or true; provenance, revision, invalidation, and consistency are separate mechanisms.
- **[E]** Sybil identities and replaceable pseudonyms can undermine redundancy and reputation, while visible social signals can create herding.
- **[E]** Public goods, common-pool resources, club goods, and private goods differ along exclusion and subtractability; markets, firms, and commons governance are institutional arrangements rather than synonyms for these goods.
- **[X]** Human-AI assistance can improve or worsen outcomes depending on task, model, interface, and participant, and explanations or human approval do not guarantee complementary performance.

### Architectural inferences

- **[I]** A Topic should be analyzed simultaneously as an environment, an authority domain, a collection of differently typed resources, and a software supply chain.
- **[I]** The system should distinguish permission from effective influence and proposal from decision, execution, evaluation, and consequence.
- **[I]** Persistent artifacts need source, version, status, freshness, dependency, challenge, and retirement semantics before they can safely accumulate.
- **[I]** Incentive mechanisms should follow resource and institutional diagnosis rather than a universal token or reputation design.
- **[I]** “Human in the loop,” “decentralized,” “transparent,” and “on-chain” are not sufficient evidence of voluntariness, legitimacy, safety, or truth.

### Normative commitments

- **[N]** Consequential actions require current, scoped authority checked independently from model cognition.
- **[N]** Participants should be able to understand, refuse, contest, revoke, and meaningfully exit within legitimate prior commitments.
- **[N]** Corrections should preserve attributable history while preventing invalidated state from silently continuing to authorize new consequences.
- **[N]** Governance should measure exclusion and burden on lower-power participants, not only aggregate throughput or cooperation.

### Hypotheses still requiring evidence

- **[H]** Qi Topics enable more meaningful voluntary contribution than conversational or centrally orchestrated alternatives.
- **[H]** Governed persistent state preserves coordination benefits while containing poisoned, stale, or captured artifacts.
- **[H]** Evaluated artifact accumulation reduces future coordination cost without information pollution overwhelming the gain.
- **[H]** Authority-sensitive human-agent delegation produces complementary performance across heterogeneous participants.

These hypotheses belong in the research agenda, not in present-tense product claims.
