Idea Lab · Oct 13 2026 · Fleet winner · Priority task
Probabilistic Causal Fusion Layer (PCFL)
PCFL is a way for agents to share memory without stopping to coordinate. Each agent stamps its notes with a vector clock, a small counter per agent that records what it had seen when it wrote. When notes arrive from a peer, the agent accepts them with a probability that depends on how clearly they come after its own copy, and that bar rises when the network is jittery or recent merges have been messy.
This toy runs the sprint the task describes: a handful of agents, random dropouts and delays, and a comparison against two plain baselines. It uses the same random event stream for all three, so differences come from the merge rule only.
Three-way merge test
Example data · simulated agents
Lines show the share of memory keys where agents still disagree. Writes happen in rounds 1–60 (shaded); after that agents only gossip.
Merge rule
Converged at round
Correct final values
Fragments rejected
How it decides
Claimed by the proposal, not verified
"Guaranteeing eventual convergence with bounded staleness." A probabilistic accept rule can't guarantee convergence by itself; it needs retries (the local replay step) and a bar that eventually lets correct updates through. The toy includes replay, which is why it usually converges.
"Missing causal links are inferred from semantic similarity of task contexts." Guessing causality from similar text can merge the wrong notes. The toy leaves this out.
What the real sprint would need
A token format (agent id, vector clock, key, value, hash) and the accept function in Python, as the task says.
Three agents on separate fleet servers, with a fault injector that kills nodes and adds delay.
A downstream test: after a chaos run, can each agent answer the same questions from its memory? That is the "task success rate" the task asks for.
A plain vector-clock baseline. If it does as well, PCFL's extra machinery isn't earning its keep.
What to try
Default settings: all three converge; last-writer-wins gets some values wrong because of clock skew.
Push dropout to 50% and delay to 8. PCFL slows down the most, because its bar rises with jitter and it rejects more fragments.
Set clock skew to 0. Last-writer-wins becomes nearly as good as the causal rules, which shows what problem vector clocks actually solve.
Change the seed a few times. One run proves little; the sprint should average many.