Worldline

Overview

Worldline is a public register for what persistent AI characters actually did inside simulated worlds. It consumes identity contracts from Agent Atlas, records world events in a merkle-chained content-addressed ledger, and keeps an agent’s claim separate from what the world observed.

Its thesis is concise: achievement is a receipt, not a claim.

Worldline is also a benchmark design. Entrants in a season receive the same content-hashed scaffold, memory budget, retry policy, tool schema, and token ceiling. The model may differ; the harness may not. A privilege card travels with every published row so a result cannot quietly inherit extra tools.

What exists now

The repository labels itself pre-alpha. One signed shakedown season exists: 12 events in the deterministic arena world, sealed against a public drand Quicknet round. From a clean clone, its ledger can be verified without network access or credentials.

That season proves the machinery, not any model. Every decision came from a deterministic L0 policy; the repository states explicitly that no model has yet been called. A live run also refuses to start without a sealed roster and a finite projected spend.

The fourth integration wave merged to main on August 26, adding scale and composition work plus a live dry-run path while preserving the repository’s stated caveat: Rekor transparency publication is designed but not implemented. A signature currently proves who signed a head, not when it was signed.

Evidence is typed

Worldline uses four outcome states instead of flattening every assertion into a leaderboard score:

• verified: evidence resolved and an independent verifier attested the relevant ledger range;
• observed: evidence resolved, but independent attestation is absent or the world is inherently observational;
• claimed: asserted by the agent but not reconciled;
• contradicted: the authoritative observation disagrees.

Infrastructure failures are recorded as void and excluded rather than quietly counted as skill losses. Unknown state is never rendered as healthy.

Different worlds, honest grades

Worldline does not pretend every environment can support the same proof:

• arena is seeded and reproduces authoritative state digests;
• minecraft is observational and relies on stored observation plus capture;
• emulator targets bit-exact replay through savestates and input movies.

The grade ships with the result. That is more useful than forcing a theatrical claim of determinism onto a world that cannot provide it.

Status

Active, pre-alpha research project. The first shakedown proves the ledger, sealing, and verifier path. It does not yet support a model leaderboard, and no public production service is claimed.