Claim-Level Review and Comparative Fairness Audit

Status

Current implementation and evidence report for IARPG-OPS-2 2.0.13-wip. Review state: current with explicit limitations. The review covers the 30 canonical source-derived reports introduced in 2.0.12 and records 167 selected claim dispositions. It does not certify every sentence in the preserved source bodies, establish legal or clinical compliance, or replace later freshness review.

Purpose

Establish a claim-level publication boundary for the 30 reports introduced from the latest supplied source collection; correct stale or unsupported assertions; apply equal evidentiary standards across countries, institutions, products, and populations; protect source provenance; and route compact current implications into active .uai memory without copying report bodies into hot memory.

Scope

The audit covers current laws and government instruments, public-agency identity, clinical and rights guidance, age and consent requirements, named vendors and products, market and adoption numbers, software capabilities, accessibility claims, security claims, virtual-economy parameters, mental-health representation, institutional framing, and comparative international fairness. The audit reviews only the 30 reports introduced in 2.0.12. The other canonical reports retain their existing review states and must be assessed separately when reused.

Executive Summary

All 30 targeted reports now contain a Reviewed Synthesis section that controls current reuse while leaving the preserved source-derived body intact for provenance. The machine-readable register contains 167 dispositions: 76 corrected, 25 rejected, 25 not established, 25 design hypotheses, 12 jurisdiction dependent, 2 verified-current, and 2 self-reported. No report is promoted to unrestricted factual authority. Every targeted report is instead marked reviewed with limitations.

The most consequential corrections are that current U.S. federal AI policy cannot be described through the rescinded Executive Order 14110; NIST AI RMF is voluntary guidance rather than a general law; IARPA is a real U.S. research organization and must not be imitated as a fictional authority; automated accessibility checks are evidence rather than WCAG certification; COPPA is not a universal adult-content age gate; European and United Kingdom age-assurance approaches remain jurisdiction-specific; clinical diagnosis and hospitalization must not be shorthand for danger, guilt, deception, or punishment; restrictive game mechanics must not be represented as treatment; local inference does not automatically establish privacy; Colibri performance numbers remain upstream project claims; and design/economy parameters remain hypotheses until tested.

Evidence Reviewed

Repository evidence

• Claim-review register
• Claim-review register schema
• Source manifest
• Report manifest
• Complete source-to-report map
• Provided-report intake audit
• All 30 canonical wrappers and their immutable source-file checksums
• Active .uai records, durable pointer ledger, public UAI mirrors, site templates, generated discovery files, and local validators

Current primary or authoritative external sources

• European Commission — AI Act application timeline
• White House — rescission order list
• White House — Executive Order 14179
• NIST — AI Risk Management Framework
• UNESCO — Recommendation on the Ethics of Artificial Intelligence
• Council of Europe — Framework Convention on Artificial Intelligence
• W3C — WCAG 2.2
• IARPA — About IARPA
• WHO/OHCHR — Mental health, human rights and legislation
• WHO — Guidance on community mental health services
• NIMH — Understanding Psychosis
• FTC — Children’s Online Privacy Protection Rule
• Ofcom — age assurance duties
• European Commission — EU approach to age verification
• JustVugg/colibri — official repository
• TinyRustLM — browser-local runtime publication
• MiRust — TinyRustLM technical guides
• OWASP — Application Security Verification Standard

Findings

Disposition totals

Review coverage by area

Report-by-report register

Current law, policy, and institutional identity

1. The EU AI Act is applied in phases. Reports must name the instrument, duty, actor, application date, and jurisdiction rather than describing “the EU” as a unitary moral position.
2. Executive Order 14110 was rescinded in January 2025. Current United States federal framing must distinguish Executive Order 14179, the current administration’s plan, statutes, agency action, and state law. No report may treat a rescinded order as current policy.
3. NIST AI RMF is voluntary, non-sector-specific risk-management guidance. It does not replace binding sector or jurisdictional duties.
4. UNESCO’s Recommendation is a normative instrument; the Council of Europe convention’s effect depends on signature, ratification, and domestic implementation. Neither should be flattened into a universal global law.
5. IARPA is a real U.S. Intelligence Community research organization and states that it has no operational mission. IARPG must avoid agency imitation, seals, names, claims of affiliation, or interfaces likely to be mistaken for an official service.

Clinical, dignity, and representation boundary

1. The website and reports must not equate psychosis, diagnosis, hospitalization, disability, or help-seeking with criminality, violence, enemy status, untrustworthiness, or incompetence.
2. Confinement, forced labor, surveillance, loss of rights, release through compliance, or punitive “mental hospital” loops must not be described as treatment or clinical realism. Where retained as fictional game-system research, they require non-clinical terminology, meaningful player agency, appeal, proportionality, and rights review.
3. The site is not a clinical service. It must not diagnose players, give individualized treatment instructions, or present source prose as medical advice. Real-world concerns should be routed to qualified local care and emergency resources.
4. International representation must include local language and legal context, disability and service-user perspectives, affected-community review, and uncertainty rather than universalizing one country’s health system or coercive history.

Age, consent, AI identity, and privacy boundary

1. For product planning, 18+ is a conservative adult-facing boundary, not a universal legal definition. Market launch requires jurisdiction-specific law, platform-policy, privacy, consumer-protection, and content review.
2. COPPA addresses certain U.S. services directed to children under 13 or with actual knowledge of collecting from children under 13. It is not a universal age-verification law for adult content.
3. European and United Kingdom age-assurance mechanisms are examples under evolving local regimes. They are not interchangeable universal deployment instructions.
4. Synthetic characters must be identified at material interaction points. Consent must be specific, revocable, and separate from general terms. Realism does not justify hiding AI identity or inducing intimate disclosure.
5. Voice, gaze, emotion, biometric, or intimate data must be minimized, purpose-bound, disclosed, and separated from engagement optimization. Local inference is only one component of privacy; optional memory, peer-to-peer, telemetry, crash-reporting, model-download, and companion paths require their own review.

Accessibility, usability, and security boundary

1. WCAG 2.2 is the current W3C Recommendation. Automated structural checks, browser fixtures, and deterministic tests are useful evidence but do not establish full conformance, native screen-reader behavior, or lived usability.
2. Time-to-understand, bounce-rate, conversion, retention, capacity, latency, and engagement thresholds found in source reports are hypotheses or test targets unless measured under a defined protocol.
3. A prose “black-box audit” does not prove a live host secure, private, fast, or reliable. Claims require a defined scope, current host evidence, reproducible results, safe testing, and explicit limitations.
4. Fictional government-like aesthetics must be paired with conspicuous fiction, non-affiliation, and current-product boundaries on every material entry surface.

Vendor, product, market, and software boundary

1. All unsupported market sizes, growth rates, adoption numbers, investment figures, hardware counts, and vendor comparisons are excluded from the reviewed synthesis unless a dated primary dataset and consistent definitions are supplied.
2. Colibri’s official repository supports the narrow claim that the project implements a C runtime for streaming mixture-of-experts weights from disk and publishes upstream hardware-specific demonstrations. Those demonstrations are self-reported and not general performance guarantees.
3. TinyRustLM’s public site supports the narrow statement that local mode runs .slm models in the browser and says prompts are not sent to a server. Optional memory, companion, peer-to-peer, retrieval, or hosted features need separate data-flow and security review.
4. No tested TinyRustLM-to-Colibri converter, common model format, direct transplantation path, or production integration is established. Similar routing ideas are architectural analogies, not equivalence.
5. Instancing, local inference, economy, progression, moderation, and governance numbers are implementation hypotheses until benchmarked and tested with representative environments and people.

Comparative fairness audit

The audit uses a common evidence matrix rather than identical prose: instrument or institution, legal effect, implementation, enforcement, oversight, rights impact, affected communities, local terminology, source diversity, uncertainty, and gameplay consequence. Equal evidence burden does not mean ignoring local context. It means that no country, institution, culture, religion, diagnostic group, or player cohort receives stronger threat language, weaker sourcing, fewer rights considerations, or harsher gameplay consequences without explicit evidence and review.

The reviewed synthesis therefore prohibits moral typecasting such as “the EU regulates,” “the United States innovates,” “China controls,” “hospitals confine,” “people with psychosis are dangerous,” or “fugitives are inherently deceptive.” It also prohibits treating the supplied archive name as a country classification. Each report is routed by its actual subject and provenance.

Decisions or Recommendations

• Treat each report’s Reviewed Synthesis as the only current reusable portion of that report.
• Preserve the original source file and SHA-256 lineage; correct canonical wrappers and the claim register rather than silently rewriting source evidence.
• Require a freshness check for every high-sensitivity legal, agency, clinical, age/consent, vendor, product, market, and software claim before later publication.
• Keep unsupported numbers and named-product comparisons out of public prose unless they are newly sourced and bounded.
• Replace coercive or stigmatizing hospital and diagnosis framing before it reaches missions, UI labels, economy systems, or NPC behavior.
• Keep AI identity conspicuous and consent revocable; do not use controller opacity as a product goal.
• Use equal evidence standards plus local-language, legal, cultural, disability, and affected-community review for country or institutional content.
• Maintain machine-readable claim dispositions and deep links from active .uai summaries to report #reviewed-synthesis sections.

Risks and Limitations

• The audit is not legal advice, medical advice, a regulatory conformity assessment, a WCAG certification, a clinical peer review, or an independent security audit.
• The selected 167 claims cover high-impact reuse risks; they are not a sentence-by-sentence fact check of the full archival bodies.
• External law, policy, product features, vendor claims, and project repositories can change after the review date.
• Some authoritative sources are global or regional; country-level implementation still requires local primary sources and qualified review.
• Repository validation can prove paths, checksums, syntax, structured records, and deterministic behavior; it cannot prove dignity, comprehension, fairness, safety, or enjoyment for real people.
• The release remains WIP because native assistive-technology, unrestricted external-host browser, field-performance, production-edge, live crawler, and moderated human evidence remain incomplete.

Validation Performed

• Verified all 30 targeted reports contain exactly one Reviewed Synthesis section and a stable #reviewed-synthesis anchor.
• Generated and schema-checked a 30-report, 167-claim machine-readable register with unique claim IDs and disposition counts.
• Verified all 349 individually preserved source files against their recorded SHA-256 checksums after canonical-wrapper changes.
• Verified report-template sections and bidirectional hot/durable links locally.
• Regenerated report and source manifests, search, sitemap, public UAI mirrors, machine-readable discovery, pointer checksums, and documentation checksums during finalization.
• Re-ran syntax, schema, source-integrity, memory-link, anchor, discovery, browser-fixture, accessibility-structure, staging-host, production-edge, delivery-budget, archive, extraction, and external-checksum validation for the final package.

Memory References

• Current progress
• AI source validation
• International fairness
• Rights and bias audit
• Representation safety
• UI shell
• Memory operating model
• Long-term-memory pointer ledger

Related Durable Documents

• [Hero clarity, report integration, and UAI routing report](hero-clarity-report-integration-and-uai-routing-report.md#executive-summary)
• [UAI memory organization report](uai-memory-organization-report.md#executive-summary)
• [International comparative fairness method](global-methodology-for-comparative-intelligence-and-security-institutions.md#5-methodology-source-hierarchy-and-confidence-framework)
• Provided-report intake audit
• Source-to-report map

Supersession Status

Current for IARPG-OPS-2 2.0.13-wip. It supersedes the 2.0.12 blanket unverified reuse boundary for these 30 reports with claim-level reviewed with limitations dispositions. It does not supersede the immutable source files, other canonical reports, current repository implementation, tests, or qualified external legal and clinical review.