Can a fleet of AI agents and devices keep itself healthy as verifiable, append-only state—without depending on a central server, elections, or synchronized clocks?

What you are looking at

Two simulated machines, their local clocks, and a frame list. Each frame contains the fingerprint of the frame before it.

Do this

1. Run the guided sample.
2. Switch to wall-clock sorting, then back to dependency order.
3. Tamper with frame 1's predecessor link.

Watch for

Time sorting can place a later action before its predecessor. Dependency order remains intact until the deliberate link mutation turns chain verification red.

What this demonstrates

Dependency links—not claimed timestamps—establish the sequence, and recomputation detects a changed link. The page does not model signatures or transport.

What you are looking at

Five devices, a hash-linked evidence log, and separate meters for scan responses and repair devices touched.

Do this

1. Run the guided loop.
2. Introduce one drift, scan, and heal it.
3. Run the wasteful full-sweep mutation.

Watch for

All five devices report, but repair touches only the exhaust targets. Two identical reports declare a stable floor; the full-sweep mutation turns repair scope red.

What this demonstrates

A fleet-wide baseline can drive follow-up proportional to implicated or locally changed devices, then stop at a stable floor. Silent changes still require a trustworthy local signal or another baseline scan.

What you are looking at

Three ranked machines, a logical beat counter, one active lease, and a SHA-256-linked control log. Rank 1 starts as leader.

Do this

1. Leave rank 3's observation lag enabled.
2. Run the guided scenario and inspect the claim beats.
3. Inject the unsafe rank-3 claim.

Watch for

Rank 2 waits through expiry and grace, rank 3 becomes eligible even later, and rank 1 returns only at the interim boundary. The unsafe claim turns the semantic verifier red.

What this demonstrates

Precommitted rank plus staggered waiting can produce deterministic failover without election messages. The teaching model does not authenticate the roster or prove partition liveness.

What you are looking at

Separate public and private frame stores, exact counts and hashes, and a derived private view that may incorporate accepted public facts.

Do this

1. Verify and read public data inward.
2. Attempt a private-to-public export.
3. Inject a conflicting public frame and try the inward read again.

Watch for

Valid public knowledge enriches the private view. Outward attempts leave public bytes unchanged, and conflicting input is refused.

What this demonstrates

The modeled transfer gate can admit verified public data while rejecting private-frame export. It does not rule out separate side channels.

What you are looking at

One stranded frame, its declared read names and content addresses, and an ordered ladder of possible bases across local and public dimensions.

Do this

1. Compare the stranded reads with each candidate.
2. Run the fixed-order search and note the first accepted base.
3. Tamper with one declared address and run it again.

Watch for

Contradicting bases are rejected exactly. The first contradiction-free base wins, provenance is recorded, and changing an address changes the result predictably.

What this demonstrates

Recovery can be a deterministic set comparison when declared reads and candidate bases are complete. No language model has to guess between the supplied histories.

What you are looking at

Two independent histories containing an aligned shared dimension, a contradictory shared dimension, and data present on only one side.

Do this

1. Run the deterministic comparison.
2. Inspect the fidelity numerator and denominator.
3. Enable the overwrite mutation and compare the assertions.

Watch for

Aligned and one-sided dimensions merge. Contradictions remain visible as parallel alternatives; the mutation turns preservation red.

What this demonstrates

Independent histories can combine useful agreement without hiding conflict or appointing an arbitrary winner. The equal-weight fidelity score is not a measure of business importance.