Open by default. Gated by evidence.

The policy

Six claims, tagged like everything else

1. P-01 ¶

   The bundle must be refused.

   Weights cannot be recalled; access can be monitored and revoked; methods diffuse regardless; outputs are speech; judgment is a thing apart. Any policy that draws one line through all five is wrong somewhere by construction. A conceptual claim, open to counterargument.

   OPEN
2. P-02 ¶

   Judgment is democratized unconditionally.

   Every method, protocol, instrument, and audit we produce is public, free, forever — verification capacity is nearly pure defense, and diffusing it makes almost every actor safer. Audit the commitment: the methods carry no wall, the instruments no account.

   VERIFIED
3. P-03 ¶

   A public verifier helps forgers less than it helps everyone else.

   The honest objection to P-02 is the forger’s oracle: an open checker teaches evasion. Our answer: forgers already iterate privately against private checks; a world where everyone can check beats a world where only some can. This is the policy’s weakest load-bearing claim, tagged accordingly.

   OPEN
4. P-04 ¶

   Generation is open by default. The burden of proof sits with the gate, never with access.

   A gate is a claim — that restricting this capability, for these people, prevents this harm — and claims carry burdens. “What would ordinary people do with this power?” has been the gatekeeper’s argument for ten millennia. It is not evidence.

   OPEN
5. P-05 ¶

   Gates are opt-in, not opt-out.

   No fixed territory of “dangerous capabilities” — territorial lines go stale as the open baseline moves. A gate exists only after it passes the test below, and it lapses when its sunset expires. Restriction must re-argue itself; openness never has to. This page enacts the rule; the verdicts below follow it.

   VERIFIED
6. P-06 ¶

   The test counts both tails.

   Misuse risk and concentration risk are priced together. A policy that tallies only what open access might do, and never what closed control might become, is rigged — our own outcome tree puts real probability on lock-in and single-actor dominance. Democratization is partly a hedge against our own branches.

   OPEN

The gate test

Five conditions. A gate opts in by passing all of them.

A gate that fails any condition is UNSUPPORTED — theatre. A gate that passes all five STANDS. Partial passage is a MIXED verdict, published with the failing conditions named. Gate verdicts compose from claim-level checks: every justification a gatekeeper offers is itself a claim, and claims are what we verify.

1. G-1 ¶
   Scoped

   The gate names a specific capability class, not general intelligence. Restricting a whole model for a worst-case slice is the aristocratic move: it selects for institutional incumbency, not for safety.

2. G-2 ¶
   Evidenced

   Published, falsifiable measurement of marginal uplift over the open baseline — open weights, textbooks, search. If older and open models already do it, the frontier gate adds friction, not prevention, and the charge of theatre is earned.

3. G-3 ¶
   Asymmetric

   Demonstrated offense–defense asymmetry and irreversibility of harm, not asserted danger. Biological synthesis pathways clear this bar; most cyber does not, because defenders patch with the same tools.

4. G-4 ¶
   Audited

   Evaluation methodology and results stand open to adversarial external review. Self-graded safety is no mark of safety. There is only evidence, and it points both ways.

5. G-5 ¶
   Sunsetted

   Gates expire by default and must be re-argued against the moving open baseline. The gatekeeper funds the surgical alternative to their own blunt instrument, or the gate lapses.

The policy must be able to approve gates, or it is advocacy wearing verification’s clothes.

Worked verdicts

The test, applied — including to us

Four gates through the five conditions. One stands, one fails, one is mixed, and one is our own.

1. V-01 ¶

   Gates on AI-assisted biological synthesis uplift.

   Scoped to a capability class (G-1 ✓); measured uplift over open baselines published (G-2 ✓); offense–defense asymmetry and irreversibility demonstrated (G-3 ✓); external red-teaming exists (G-4 ✓); sunset and re-argument schedule still missing (G-5 pending). Attach the sunset and this gate is the model of a legitimate one.

   STANDS
2. V-02 ¶

   Whole-model frontier tiering as currently practiced.

   Published capability-specific justifications partially satisfy G-2 and G-3 — the bio slice is real. But the gate is drawn around the entire model rather than the capability (G-1 strained), independent adversarial audit of the evals is thin (G-4 strained), and no tier carries a sunset (G-5 ✗). Verdict: the narrow version would stand; the broad version is incumbency wearing the narrow version’s evidence.

   MIXED
3. V-03 ¶

   Blanket restrictions on AI cyber capability.

   Marginal uplift over open tools unproven (G-2 ✗) — legacy code falls to older and open models already, a documented reality. Defenders patch with the same capability, so the asymmetry claim fails (G-3 ✗). Two failed conditions; the verdict follows.

   UNSUPPORTED
4. V-04 ¶

   Solonic’s own gate on the Occupy AI protocol.

   The policy covers the shipwright. The saturation protocol stays gated by us, under our own test: its mechanism is unmeasured until the pre-registered experiment runs (G-2), and the method generalizes to worse payloads than ours (G-3). The gate carries a sunset — the experiment’s report — after which it must stand re-argument or lapse (G-5 ✓).

   STANDS

The standing offer

Send us your gate