A public decision guide to open-weight AI governance
Start with the choice, not the framework. Open-weight release can expand research, competition, scrutiny, and access. It can also make provider controls unable to reach every copy. That leaves a public decision: allow, guide, govern, restrict, or ban—and accept the risks of whichever path we choose.
Lifecycle: Published review edition. The frozen source candidate passed its Phase 12 independent recheck and parent verification. Published July 30, 2026 by explicit instruction for review.
Boundary: Evidence remains frozen at July 28, 2026. This is policy research, not legal or operational advice.
First, name the problem
This site focuses on advanced open-weight AI releases—not open-source software, open science, or every model marketed as “open.” The trained weights may be downloadable even when code, data, licenses, and safeguards differ.
Research, independent scrutiny, competition, local adaptation, education, and access can benefit when people are able to inspect and run a system.
After weights spread, provider monitoring, account controls, forced updates, or withdrawal may no longer reach every copy.
Persistence alone does not establish dangerous capability, misuse intent, human uplift, a severe-harm pathway, or the wisdom of a restriction.
The real public problem
Society has to weigh two kinds of error: releasing too freely when control will be difficult to recover, and restricting too broadly when the evidence is weak. Either mistake can impose costs on people who did not choose them.
The policy menu
Every path protects something valuable and exposes something else. The useful comparison is the likely benefit, burden, failure mode, and reversibility of each choice.
01 · Do nothing new
Potential benefit
Maximizes permissionless research, scrutiny, adaptation, competition, and access while avoiding premature rules built on weak evidence.
Main risk
Relies heavily on action after misuse. Once weights spread, provider controls may not reach every copy, and public costs can fall on people who never chose the risk.
02 · Voluntary safeguards
Potential benefit
Can move faster than law and improve evaluations, documentation, repository practice, and incident sharing without a state mandate.
Main risk
Coverage and incentives are uneven. Actors creating the greatest risk may decline, underinvest, or defect, and harmed people may lack an enforceable remedy.
03 · Govern harmful use
Potential benefit
Targets malicious or negligent behavior rather than a general-purpose artifact, protecting lawful research and lower-risk use.
Main risk
Often depends on attribution, jurisdiction, and proof after distribution—or after harm. It may not address an irreversible release decision in time.
04 · Conditional governance
Potential benefit
Uses staged access, evaluations, documentation, safeguards, review, and appeal to match duties to demonstrated capability and an actual release profile.
Main risk
Evaluations are incomplete and can be gamed. Compliance cost and uncertainty can favor incumbents or chill legitimate work unless thresholds and review are disciplined.
05 · Targeted restriction
Potential benefit
Can prevent a difficult-to-reverse loss of control when evidence supports a severe and credible pathway that weaker measures cannot address.
Main risk
False positives can suppress research, scrutiny, competition, civil liberties, and access; restrictions may shift release into less visible or offshore channels.
06 · Outright ban
Potential benefit
Draws the clearest legal line and seeks to stop public distribution before copying makes recall difficult.
Main risk
Treats radically different releases alike, concentrates power in closed providers and governments, sacrifices lower-risk open use, and creates strong incentives to evade the rule.
A necessary companion to every option
Better evaluations, secure repositories, sector-specific safeguards, provenance, incident response, international coordination, and remedies for affected people can reduce harm under any release regime. They complement—but do not settle—the release decision.
How to choose
A defensible decision should explain both what changes when weights are released and why the proposed response is better than less restrictive alternatives.
This site's answer
The framework that follows favors evidence-triggered, release-specific, actor-specific safeguards. It rejects both automatic laissez-faire and a blanket ban, and it requires stronger measures to defeat less restrictive alternatives rather than merely sound precautionary.
Start with MAPS: Model Access, Permissions, and Safeguards. Then inspect the policy mechanisms, evidence limits, objections, and full paper.
Choose an audit path
Start with the decision in front of you, then move directly to the evidence, objection, or full-paper layer that can challenge it.
MAPS separates access, permissions, and safeguards without turning releases into model classes.
02Thirteen mechanisms assign duties by authority, reach, evidence, review, and defeat condition.
03Thirty-one canonical rows connect exact claim IDs to bounded evidence and source-note custody.
04Seven lenses retain their attack, concession, revision, surviving claim, defeat test, and dissent.
05The complete frozen Phase 6 position paper remains available without editorial shortening.
06Exact source dates, retrieval details, scope, limitations, public citations, and frozen-custody hashes remain visible in the evidence route.
The control boundary
Licenses, notices, patches, lawful action, provenance, preparedness, and response can still matter. Their reach is partial and must be measured; none is represented as a universal delete button.
The provider may monitor use, remove an account, change safeguards, suspend access, or force an update—subject to law, notice, and appeal.
Provider-side withdrawal no longer reaches every copy. Prospective distribution controls, notices, voluntary patches, reachable intermediaries, investigation, preparedness, and response remain partial routes.
MAPS
MAPS is a release-profile framework only—never a model class, capability tier, or safety score. It records what is available and to whom, what the terms permit, and which safeguards depend on provider control or remain usable after copying.
Which exact weights, revision, hosted configuration, or bounded access condition generated the evidence?
What task, population, comparison, uncertainty, and estimand did the study actually measure?
Which release-profile change would remove or preserve which provider-dependent control?
Which policy-level barrier could the measured assistance change, and what barriers remain?
What the record actually says
The evidence can justify better evaluation, preparedness, and a review process without being rhetorically upgraded into an observed catastrophe.
Biological tasks
Zhang et al. estimated a 4.16-fold overall improvement for novice accuracy (95% CI 2.63–6.87). The treatment allowed multiple hosted services; it does not isolate one model, checkpoint, physical outcome, or catastrophe.
Physical-world trial
Hong et al. reported 5.2% versus 6.6%, P=.759, for the tested mid-2025 multi-model access condition. That is not proof of no risk, and it is not attributable to a named release.
Kimi evidence
Kimi K2.5 has independent model-specific biological benchmark evidence and separately available weights. Kimi K3 has hosted cyber evaluation evidence and separately pinned repository metadata. Provenance does not make either pair identical.
Policy consequence
Pin the object, fund independent evaluation, publish methods and nulls, protect reporting and research, document safeguard limits, measure burden and control reach, and fund preparedness and response.
Evidence vocabulary
Observed fact, interpretation, forecast, proposal, and moral commitment are not interchangeable. A source can establish what happened or what an institution said without proving what should happen next.
Human consequence
Catastrophic-risk uncertainty cannot make affected people invisible. Neither can an overbroad response that closes research, accessibility, defensive security, local use, competition, or participation without a supported and contestable reason. Precaution must remain proportionate, reviewable, time-limited, and open to defeat by better evidence.