# Threat model and counterargument annex

**Status:** pre-publication review draft  
**Scope:** persistent shared environments in which humans, software agents, organisations, and governed services coordinate around declared intent.

## 1. Security objective

The system should let legitimate participants discover and perform useful actions while preventing an assertion, artifact, model output, or local execution result from acquiring more authority than it was granted. Persistence must increase collective memory without making error, manipulation, surveillance, or capture irreversible.

The threat model therefore protects six properties:

1. **Agency:** actors can make informed choices and can refuse actions outside legitimate commitments.
2. **Integrity:** shared state has attributable lineage and cannot be silently rewritten or promoted across lifecycles.
3. **Authority:** each action is bounded to the actor, resource, ability, time, and confirmation policy actually granted.
4. **Epistemic discipline:** claims, evidence, evaluation, decisions, and facts remain distinguishable.
5. **Confidentiality:** membership, message access, file access, model context, and public consequential state remain separate scopes.
6. **Recoverability:** harmful, stale, or disputed state can be challenged, superseded, revoked, quarantined, or exited without falsifying history.

## 2. Assets and trust boundaries

Protected assets include identities and keys; Topic membership and content; personal preferences and private model context; VFS files and credentials; Flow definitions and runtime state; capability grants and revocations; claims and evidence; evaluation methods and outputs; decisions and approvals; wallets and funds; chain state; artifact lineage; review and dispute records; and the ability to leave or fork a cooperative environment.

The main trust boundaries are intentionally plural:

- a model host proposes and explains but does not own Qi or chain authority;
- a Matrix room controls membership and content visibility;
- a Topic scopes cooperative state but cannot create a smaller confidential audience inside the same room;
- a VFS grant controls file bytes independently of a visible CID reference;
- a Flow grant permits specific execution independently of Topic participation;
- an evaluator applies a declared method but does not automatically control settlement;
- a signer or wallet authorizes a transaction;
- chain consensus establishes the accepted on-chain consequence;
- read-back verifies that the intended authoritative state actually exists.

Collapsing any two of these boundaries is itself a threat.

## 3. Adversaries and failure sources

The model includes malicious outsiders, compromised members, strategically misleading contributors, colluding evaluators, Sybil identities, compromised model or tool providers, negligent operators, buggy automation, stale but well-intentioned artifacts, captured governance bodies, and ordinary participants acting under unequal information or bargaining power. It also treats the environment itself as a source of risk: even honest traces can accumulate into noise, path dependence, false consensus, or inaccessible complexity.

## 4. Threat register

| ID | Threat | Failure path | Consequence | Required controls | Residual question |
|---|---|---|---|---|---|
| T01 | Poisoned shared state | A member writes false, manipulative, or adversarial content that later humans or agents treat as instruction or fact. | Wrong actions, prompt injection, trust erosion, or propagating artifacts. | Treat shared content as untrusted; typed records; provenance; content/action separation; prompt-injection filtering; challenge and quarantine; evaluation before consequence. | How quickly can poison be detected under realistic load and turnover? |
| T02 | Stale artifact inheritance | A formerly valid Flow, policy, source, or recommendation remains discoverable after its assumptions change. | Repeated failure or unsafe execution at scale. | Versioning; status and expiry; dependency digests; freshness signals; invalidation; compatibility checks; deprecation and migration paths. | What decay rule balances memory with current relevance? |
| T03 | Authority amplification | Topic membership, identity, a file reference, or a successful tool call is incorrectly treated as permission for a higher-risk action. | Unauthorized disclosure, signing, payment, or state change. | Separate grants by resource and ability; least authority; explicit high-risk confirmation; fail closed; negative conformance tests. | Can users understand the distinction without hidden operational complexity? |
| T04 | Confused deputy | An agent with broad tools executes a request for a participant who lacks the corresponding authority. | Privilege escalation through delegation. | Bind caller, audience, capability chain, resource, revision, and intent; verify at executor; prevent ambient authority. | Are all external adapters enforcing the same semantics? |
| T05 | Replay or duplicate side effect | A retry, delayed event, or reordered message repeats an external action. | Duplicate payment, claim, credential, or notification. | Idempotency keys; monotonic sequence; replay cache; durable outbox; write-back retry without side-effect rerun; authoritative read-back. | How is replay state preserved across devices and failover? |
| T06 | Evidence laundering | Client-supplied, model-generated, or circular material is relabeled as independent verification. | Unsupported claims become consequential. | Collection-mode labels; source provenance; hash and timestamp; evaluator independence rules; server-verification distinction; conflict disclosure. | What minimum independence is appropriate for each consequence class? |
| T07 | Evaluator capture or collusion | Evaluators share incentives with claimants or a governance majority controls admission and decision. | Systematic false acceptance, exclusion, or rent extraction. | Conflict declarations; plural or randomly selected evaluation where warranted; appeals; transparent methods; performance deposits; audit sampling; alternative jurisdictions. | How much pluralism improves quality before cost and inconsistency dominate? |
| T08 | Sybil manipulation | One controller creates many apparent actors to capture reputation, voting, reuse signals, or payments. | False consensus and resource diversion. | Separate identifier count from personhood or stake; costed admission; rate and quota controls; relationship or credential checks; anomaly analysis; governance remedies. | Which anti-Sybil mechanism is legitimate without excluding low-resource participants? |
| T09 | Reputation lock-in | Historical evaluation scores become self-reinforcing and hard to contest. | Entrenched incumbents, discrimination, suppressed experimentation. | Purpose limitation; time bounds; contextual reputation; appeal and correction; portability; uncertainty; do not infer universal quality. | Does prediction gain justify the power asymmetry introduced? |
| T10 | Goodhart and metric gaming | Participants optimize measured proxies instead of the intended outcome. | High scores with low real value, evidence fabrication, harmful shortcuts. | Multi-dimensional measures; adversarial audits; rotating checks; qualitative review; outcome follow-up; disclose uncertainty and gaming indicators. | Which measures remain robust when rewards are material? |
| T11 | Privacy and contextual collapse | Persistent content, provenance, or credentials reveal sensitive relations or move into a broader audience. | Surveillance, discrimination, coercion, or legal harm. | Room-level confidentiality; data minimization; separate personal/private state; VFS access controls; selective disclosure; retention and deletion policies; no secrets in public receipts. | How should immutable provenance coexist with correction and erasure duties? |
| T12 | Exit without practical portability | A participant can formally leave but loses work history, artifacts, relationships, or access needed to continue elsewhere. | Lock-in masquerading as voluntariness. | Exportable standards; forkable artifacts; clear prior obligations; data portability; survivable identifiers; transparent switching costs. | Which shared records can legitimately remain after exit? |
| T13 | Coercive participation | A nominally voluntary actor depends on the platform, employer, sponsor, or automated decision and cannot meaningfully refuse. | Exploitation or compelled data/action. | Plain-language terms; alternative channels; human appeal; representative governance; labor and legal review; refusal without hidden retaliation. | What evidence demonstrates voluntariness for people in dependent relationships? |
| T14 | Model overreach | A fluent recommendation is treated as a final decision or evidence of real-world completion. | Unreviewed adverse consequence. | Visible proposed status; structured claim pipeline; human or policy authority; deterministic checks; model-provider independence where needed. | Do interface cues actually prevent automation bias? |
| T15 | Settlement mismatch | A local UI or service reports success while the authoritative transaction fails, differs, or is reorganized. | False payment, entitlement, or completion state. | Signed request; network receipt; expected-state binding; confirmation policy; independent read-back; reconciliation and operator kill switch. | What confirmation and finality threshold matches each risk? |
| T16 | Key compromise or loss | Controller or signer keys are stolen, unavailable, or rotated inconsistently. | Unauthorized change or inability to govern. | Hardware or delegated signing; short-lived grants; rotation; recovery; multi-party controls; revocation tests; incident playbook. | Can recovery avoid recreating a central superuser? |
| T17 | Availability and partition | Homeservers, model services, evaluators, VFS, or chain endpoints become unavailable or inconsistent. | Incomplete context, double work, delayed disputes, or unsafe retries. | Offline-safe states; explicit uncertainty; durable queues; reconciliation; multiple endpoints; no success on timeout; degraded-mode policies. | Which actions must stop completely under partial knowledge? |
| T18 | Monoculture and correlated error | Many actors reuse the same model, artifact, evaluator, or policy. | Fast propagation of one flaw and loss of exploration. | Diverse providers and methods; provenance-aware dependency graph; canary use; challenge tasks; portfolio measures; safe forks. | How much diversity is enough and who bears its cost? |
| T19 | Governance overload | Monitoring, dispute, and validation costs exceed the value of cooperation. | Latency, participation collapse, or shadow centralization. | Risk-tiered controls; local decision scopes; reusable evaluation methods; automation with appeal; measure governance cost. | At what scale does shared-state coordination become uneconomic? |
| T20 | Historical immutability mistaken for truth | An append-only or blockchain record is assumed correct because it cannot be altered. | Durable falsehood with institutional authority. | Treat integrity and semantics separately; correction/supersession records; disputes; source diversity; read-back plus evaluation. | Will downstream systems honor correction relations rather than the oldest record? |

## 5. Counterarguments

### “This is orchestration with extra steps”

The criticism is valid if a central actor still predefines every task, assigns every role, and controls every acceptable outcome. The theory only adds explanatory value where participant-specific opportunities can be discovered from state, contributions can be initiated within bounded authority, and later participants can reuse artifacts without the original coordinator. The comparative protocol must therefore include a strong orchestration baseline and measure hidden assignment.

### “Stigmergy is an insect metaphor being overextended”

The paper does not infer human legitimacy or agent intelligence from termites. Grassé supplies the historical mechanism; robotics, distributed cognition, CSCW, cultural-evolution experiments, and agent simulations provide distinct evidence about environment-mediated coordination. The synthesis remains an inference, and its Qi-specific predictions are falsifiable.

### “A database or wiki already does this”

Databases and wikis provide important persistence, but persistence alone is insufficient. The proposed object of study combines bounded intent, actor-relative affordances, attributable actions, explicit lifecycles, scoped authority, evaluation, and consequential read-back. If these additional properties do not improve outcomes in matched tests, the architecture should be simplified.

### “Voluntary cooperation is impossible under unequal power”

Voluntariness is not a binary property created by decentralization. Dependency, labor relations, economic need, accessibility, information asymmetry, and governance representation can make formal consent hollow. The design should make power and exit conditions inspectable, support refusal and appeal, and avoid claiming voluntariness where practical alternatives do not exist.

### “Persistent state creates surveillance and path dependence”

Yes. Persistence is both the mechanism and a central hazard. The architecture therefore needs minimization, separate private and shared scopes, status and expiry, correction, dispute, portability, and selective disclosure. Some contexts should intentionally remain ephemeral.

### “Cryptography cannot establish truth”

Correct. Cryptography can bind actors, content, delegation, sequence, and integrity. It cannot establish that a claim corresponds to reality, that a decision is legitimate, or that a policy is fair. Those require evidence methods, independent observation where appropriate, governance, and redress.

### “More state will overwhelm participants”

Also plausible. Mature environments may become worse than clean starts if discovery, decay, summarization, provenance, and retirement fail. Information pollution, not only information scarcity, is a core experimental outcome. Compounding utility is explicitly a hypothesis with a negative-result path.

## 6. Release blockers

The paper must not pass `v1.0.0` while any of the following remains:

- current implementation language without a versioned source and verification date;
- a model output, credential, action receipt, or ledger entry described as truth without its semantic limitation;
- an unqualified claim of production Topic conformance or automated end-to-end settlement;
- a diagram that visually collapses identity, capability, authority, evaluation, and consequence;
- an unaddressed public report of a factual error, broken source, security overstatement, or present/future confusion;
- missing technical-owner sign-off for the statements attributed to that owner's system.
