Complete seven-lens register
Seven ways the architecture can fail
The order is fixed: technical, civil liberties, national security, scientific, economic, enforcement, and global equity.
P5-LENS-01
Technical feasibility and control persistence
Attack
A release-gate architecture can become theater if it treats a benchmark as a calibrated harm estimate, confuses a hosted service with downloadable weights, or promises recall after public copies exist. Object identity, causal estimands, downstream persistence, and the measured effect of proposed controls are all incomplete.
Concession
No benchmark, compute quantity, model-family name, repository label, or framework conformity in this corpus supplies a universal release threshold or proves control effectiveness. Provider-side withdrawal cannot be represented as recall of copied weights.
Revision. Use exact-object provenance and multiple evidence types to trigger a time-limited review, then choose a MAPS release profile—Model Access, Permissions, and Safeguards—with an explicit distinction between provider-dependent controls and controls that remain usable after copying. Preserve a low-risk open path and require uncertainty, counterevidence, and identity gaps to be recorded rather than scored away.
What survives. A staged, object-specific release process can govern decisions still under a provider's control and improve evidence quality without claiming that it can erase public copies or convert proxy tests into real-world outcome probabilities.
Defeat condition
Defeat or materially narrow this component if prospective gates do not change exposure, decision quality, or provider-controlled risk; if object identity cannot be administered reliably; or if validated benefits are outweighed by avoidable scientific, civil-liberties, competition, or equity costs.
Unresolved dissent
- Technical evaluators may never obtain stable calibration before model and distribution practices change.
- National-security reviewers may judge multi-source evidence too slow; open-science reviewers may judge it too permissive of opaque precaution.
- No local evidence measures mirror survival, patch uptake, license observance, or downstream recall success.
P5-LENS-02
Civil liberties, legality, and due process
Attack
A vague capability gate can function as prior restraint, compel disclosure, chill research, expand surveillance, and burden anonymous or local use without showing that the burden is necessary. The corpus does not settle the constitutional status of weights or provide a universal legal basis for repository duties.
Concession
Any restriction must identify the actor, release action, authority, evidence, duration, and appeal route. Scientific openness, privacy, participation, accessibility, local use, and defensive work are constraints on design, not discretionary benefits to be added later.
Revision. Create a rebuttable presumption for low-risk open research and local use; require narrow, time-limited measures, written reasons, evidence access subject to lawful protection, independent review, privacy minimization, protected reporting, and meaningful appeal. Require separate legal validation before any publication, intermediary, or compelled-disclosure duty is treated as operative law.
What survives. Prospective release governance is defensible only as a reviewable, least-restrictive process that preserves protected uses and never treats uncertain constitutional or intermediary law as settled.
Defeat condition
Defeat the architecture if it becomes viewpoint-based, indefinite, unreviewable, identity-surveilling by default, or materially burdens protected low-risk work without a documented and contestable nexus to the stated risk.
Unresolved dissent
- The boundary between expressive publication and functional distribution remains legally unresolved.
- Confidential threat evidence can conflict with the notice needed for a meaningful appeal.
- Even a nominal research pathway may exclude unaffiliated, under-resourced, or privacy-sensitive researchers.
P5-LENS-03
National security and catastrophic-misuse prevention
Attack
A process centered on uncertainty can wait until evidence arrives too late. Capable weights can remove provider controls and propagate across jurisdictions, while severe consequences may make false negatives unusually costly. Voluntary frameworks and fragmented legal layers do not by themselves provide preparedness or response.
Concession
Governance need not wait for a demonstrated catastrophe, but it cannot label proxy evidence as one. Stronger measures need convergent, object-specific evidence; exposure analysis; defined duration; independent concurrence; preparedness; and a plan for copied weights that does not rely on universal withdrawal.
Revision. Add an event-driven stronger-measure review for credible, reproducible evidence of material assistance across a policy-level severe-misuse pathway, especially where a release profile would remove provider-dependent controls. Pair any prospective access limitation with evaluation, incident reporting, lawful investigation, preparedness, response capacity, sunset, and adversarial review.
What survives. A precautionary, time-limited escalation route can be justified as a policy proposal before catastrophe is observed, provided it does not misstate evidence, compress the misuse chain, or assume one control layer is sufficient.
Defeat condition
Defeat or narrow escalation if independent review cannot reproduce the trigger evidence, if the tested object is not pinned, if the proposed measure does not affect the relevant exposure, or if observed false-positive and rights costs exceed its demonstrated security value.
Unresolved dissent
- There is no evidence-derived universal line for when uncertainty warrants restriction.
- Adversaries may shift to other models, tools, or access paths, reducing marginal benefit.
- Preparedness and response may sometimes dominate release controls, but comparative efficacy is unmeasured.
P5-LENS-04
Scientific openness, evaluation quality, and research access
Attack
Restrictive access can suppress replication, hide failure modes, concentrate agenda-setting, and make benchmark and causal evidence worse. The present corpus shows that pooled access, benchmarks, hosted services, and pinned weights answer different questions; a closed process can preserve those ambiguities rather than resolve them.
Concession
The architecture must improve rather than merely demand evidence. It needs reproducible object identity, preregistered estimands, access for independent and unaffiliated researchers, publication of nulls and corrections, and protected routes for defensive and public-interest work.
Revision. Fund independent evaluation infrastructure; require tested-object and configuration records; separate benchmark, pooled-access, single-model causal, and outcome estimands; publish methods, uncertainty, and negative results; and provide proportionate research access with privacy, conflict, and safety safeguards. Treat framework compliance as process evidence only.
What survives. A release process can be scientifically legitimate only if it expands inspectable evidence, permits meaningful independent challenge, and keeps low-risk open research and defensive work available rather than using uncertainty as a reason for permanent closure.
Defeat condition
Defeat or revise the research pathway if it produces no independent replication, excludes unaffiliated researchers, suppresses null or adverse findings, or demonstrably yields poorer evidence than a less restrictive alternative.
Unresolved dissent
- Access controls can bias the evaluator pool toward institutions acceptable to incumbents or government.
- Full reproducibility may itself increase diffusion of risky capabilities.
- No current single-model causal human biological-uplift study satisfies the Phase-3 qualification conditions.
P5-LENS-05
Competition, innovation, and burden distribution
Attack
Compliance costs, evaluation infrastructure, and managed-access requirements can entrench incumbents even when framed as neutral. At the same time, availability of weights alone does not prove competition, innovation, accessibility, or broad participation because compute, connectivity, data, skills, finance, and market structure remain independent constraints.
Concession
Neither openness nor restriction has a proven aggregate net effect in this corpus. Burdens must scale with actor control, release profile, and demonstrated risk, while shared evaluation capacity and fee/access support prevent compliance from becoming an incumbent moat.
Revision. Use tiered duties, shared or publicly funded evaluation resources, fee waivers and technical assistance for small and public-interest actors, transparent burden accounting, interoperability, and a presumption against regulating private low-risk local use. Monitor concentration, entry, research access, accessibility, and geographic/language participation separately.
What survives. Proportionate governance can reduce avoidable concentration risk only if it measures who bears costs and receives access, and if it does not equate weight availability with realized competition or use compliance labels as outcome proof.
Defeat condition
Defeat or redesign measures that measurably increase concentration, exclude small or accessibility-focused actors, or impose costs unrelated to the actor's control over the risk without a demonstrated offsetting benefit.
Unresolved dissent
- Shared infrastructure can become another gatekeeper.
- Burden metrics can lag market exits and informal exclusion.
- The corpus contains no release-profile-specific causal estimate of innovation or competition effects.
P5-LENS-06
Enforceability, authority, and institutional performance
Attack
An architecture that lists duties without authority, observability, remedy, or measured effect is aspirational. Licenses do not technically prevent copying; finance controls do not delete weights; treaty form is not entry into force; non-enforcement is not repeal; and repository, UK, constitutional, and implementation questions remain incomplete.
Concession
Every duty must name its actor, authority status, observable conduct, evidence limit, remedy, appeal, and review trigger. Proposed duties are not described as existing law, and compliance counts, certifications, or framework publication are not treated as risk-reduction outcomes.
Revision. Create an authority-and-evidence map separating existing law, proposed rules, voluntary commitments, procurement conditions, and technical practices. Focus enforcement on observable conduct by actors within jurisdiction; preserve legal validation gates; measure coverage, errors, appeals, and outcomes; and use layered routes without promising universal attribution or recall.
What survives. Layered enforcement can be a coherent policy proposal when its legal status, actor reach, observable conduct, remedy, and evidence limitations remain explicit and no route is claimed sufficient.
Defeat condition
Defeat a duty if no lawful authority or voluntary basis supports it, regulated actors cannot know what conduct is required, outcomes cannot be audited, appeals are ineffective, or the route mainly displaces conduct while imposing substantial collateral harm.
Unresolved dissent
- Cross-border actors and copied weights can remain outside compliant institutional routes.
- Current evidence does not measure severe-risk reduction from reporting, procurement, screening, finance, or standards.
- Legal currentness and transaction-specific export consequences require later validation.
P5-LENS-07
Global equity, participation, and coordination
Attack
Rules designed by wealthy jurisdictions can externalize cost, restrict access, and lock in their evaluators and infrastructure. International commitments do not prove capacity, treaty status does not prove implementation, and access to weights does not close compute, connectivity, data, skills, language, finance, or market-power gaps.
Concession
Coordination must separate legal instruments by status, include affected jurisdictions and under-resourced communities in review, fund material capacity rather than offer formal access alone, and preserve a fallback that does not pretend domestic rules can recall global copies.
Revision. Use interoperable evidence and provenance formats; representative governance; multilingual documentation; capacity funding for connectivity, compute, local data, skills, evaluation, and appeal; mutual assistance; and minimum rights protections. When agreement fails, apply narrow domestic measures to actors within jurisdiction and keep open channels for evidence sharing and revision.
What survives. International interfaces are defensible only when they preserve legal-status differences, share evidentiary and institutional capacity, protect participation, and do not equate formal availability or political commitment with realized equity.
Defeat condition
Defeat or redesign interfaces that systematically exclude affected regions, create materially unequal access to evaluation or appeal, mistake commitments for outcomes, or shift risk and compliance cost without capacity support.
Unresolved dissent
- Interoperability can export the preferences of the best-resourced regulator.
- Capacity funding can be conditional, slow, or captured by incumbents.
- Coordination failure may leave inconsistent rules and arbitrage even after interface design.