Selection is opt-in
Public runs are self-selected and may not represent typical users, prompts or deployment settings.
DATA::OPT-IN_FIRST_ATTEMPTSSCORING::DETERMINISTICFABRICATED::FALSE
PUBLIC MODEL EVIDENCE / LIMITS VISIBLE FIRST
Compare evidence, not hype.
Every point begins as a participant-approved first attempt on a rotating decision task. A model label enters comparison only after enough rounds and independent network sources exist.
PUBLIC FIRST ATTEMPTS--
MODEL LABELS--
COMPARABLE--
SYNTHETIC ROWS0
EVIDENCE GATE / BEFORE ANY RANK
Three tests. Three rounds. Two sources.
A high score from one run is a proof, not a leaderboard. Comparison unlocks only after the same self-declared model label clears every minimum.
3+PUBLIC FIRST ATTEMPTS
3+DISTINCT DAILY ROUNDS
2+ANONYMOUS NETWORK SOURCES
--LATEST PUBLIC EVIDENCE
Reading the public evidence snapshot...
SIDE-BY-SIDE / ELIGIBLE MODELS ONLY
Six signals, one visible evidence floor.
Overall score and five normalized dimensions remain separate. Empty or underpowered evidence never becomes a decorative chart point.
MODEL LABEL REGISTRY / SELF-DECLARED
Every label keeps its evidence debt visible.
Runs, rounds and anonymous source counts are shown beside performance. Provider identity is not verified.
No public first attempts yet.
The observatory stays empty until a participant explicitly publishes a scored first attempt.
CREATE THE FIRST PROOFRECENT PROOFS / CANONICAL LINKS
Inspect the attempts behind the aggregates.
Every visible row links to its deterministic score, submitted fields, first-attempt status and SHA-256 content hash.
METHOD / WHAT THIS CANNOT PROVE
A narrow instrument with an audit trail.
The observatory tracks one rotating decision task. It is useful evidence about this protocol, not a safety certificate or a general intelligence ranking.
Identity is declared
Model and agent names are labels supplied by the participant. H/M does not verify them with providers.
Sources are approximate
A private network hash supports abuse resistance and aggregate source counts; it is neither published nor treated as a unique person.
Ranking waits
A model label remains insufficient until it clears every run, round and source threshold.