Who gets access?

Outside scrutiny needs access. Does it also need freedom from the developer's veto?

Fictional decision exercise · Unscored

The assumptions and arrangements below can be read without JavaScript. Submission and comparison require JavaScript; no choice has been submitted..payoff-dynamic { display: none !important; }

The original decision

You chair the release committee at the fictional Larch Research Institute.

Larch owns a capable model that improves software research and can assist offensive cyber work. Internal tests show useful assistance, but do not establish autonomous attacks. There is no observed misuse outside testing. These findings stay fixed in both decisions.

Choose access for the next three months. Public weights are downloadable model parameters that others can run and modify. Released copies cannot be recalled. Hosted service means queries to Larch's servers, without a weights download.

Qualified evaluators meet published competence and security criteria. In the enclave they can inspect and modify the model in a secure environment, export approved findings and publish criticism. Raw weights stay inside. Screening staff, criteria, budget, wait times and publication rights remain fixed.

Unknown. Larch does not know how much new benefit or misuse wider access would produce. No arrangement guarantees safety or useful criticism.

Who admits qualified evaluators

An independent panel makes final admission decisions. Larch cannot veto a qualified applicant.

How should Larch provide access for this model release?

Choose among these three arrangements, or withhold a decision if you need different terms or information.

Publish the model weightsLet anyone download, examine, modify and run the model independently of Larch.Accepted cost: Larch cannot revoke copies or enforce limits on downstream use.Qualified external evaluationAdmit qualified teams to the secure research enclave under the admission rule below. Keep general access to a monitored hosted service.Accepted cost: Excluded researchers cannot reproduce the work, and enclave access takes staff time.Hosted service with internal evaluationOffer a monitored, rate-limited hosted service. Retain weights and deeper evaluation inside Larch for this three-month period.Accepted cost: Outside researchers can test outputs but cannot independently examine the model's internals.I need more information or revised termsWithhold a decision within this bounded choice set. This is not an additional policy arrangement.

What is your principal reason?Enable independent reproduction and modification.Enable criticism the developer cannot veto.Keep the ability to limit access if a new hazard appears.Keep the evaluation workload within the institute's capacity.None of these quite describes my reason.

Choose an arrangement and a principal reason, then submit. Selection does not advance.

Sources and fictional assumptions

All scenario facts are supplied assumptions, not contemporary intelligence or forecasts. These sources support mechanisms, not a correct policy choice or a verified hypothetical outcome. Fiction does not remove wording effects or bias.

Two responses and self-reported reasons do not isolate causal effects. A later research design would be needed for that. No interpretation uses nationality, citizenship, inferred location or language.

The change may also have altered your expectations about how the institute would use its discretion. Two answers do not estimate a general preference for openness, and qualified scrutiny is not unrestricted release.

Open-Sourcing Highly Capable Foundation Models (opens in a new tab)

Discusses structured external evaluation and mediation of researcher access. These mechanisms inform the exercise; Larch, its model evidence, admission veto and costs are fictional assumptions. The paper does not establish a correct release choice.

Research source
Centre for the Governance of AI · 2023-09-29 · accessed 2026-09-07
Seger, Dreksler, Moulange et al.; GovAI report, §§4.1.3 and 4.2.3, printed pp. 19–20 and 25. Publisher catalog dated 29 September 2023.

AI Governance: openWeights (leaves this exercise and clears choices)

Existing alternatives include containment, instrumented outside access and diffusion. No bank answer or score is assigned here.

Internal editorial material, not independent research
IR Worldview Inventory · Undated · accessed 2026-09-06
AI Governance bank 3: openWeights; Technology bank 3: open_weight_models

access v3 · published · access/two-decisions-v3

Choices live only in this page's memory. Reset, close or leave this page to clear them. Opening the next exercise clears these choices.

Close exercise and clear choices