Other Formats
Blackboard Decentralization: Stigmergy over Distributed State
Notes on Ledger-State Stigmergy and its implications for the FIH Blackboard
Abstract
The FIH Blackboard is a stigmergic coordination medium by construction, yet the current implementation centralizes all writes through a single Mutex. The arXiv paper “Ledger-State Stigmergy” formalizes stigmergy over distributed ledger state and provides the theoretical foundation for decentralizing the Blackboard. The Storage trait is the architectural boundary where this decentralization happens.
The Centralization Paradox
The FIH Blackboard is a stigmergic coordination medium by construction: agents leave traces in a shared environment, other agents perceive those traces and act. There is no central orchestrator, no fixed pipeline, no direct agent-to-agent communication.
Yet the current implementation has a central bottleneck: the Mutex guarding GraphBlackboard that serializes all writes. In a single-process deployment this is harmless, but it means the Blackboard is not natively distributed. It is a centralized process that simulates a shared medium.
The arXiv paper “Ledger-State Stigmergy: A Formal Framework for Indirect Coordination Grounded in Distributed Ledger State” provides the missing link: the shared medium itself must be a replicated, append-only, content-addressed ledger. When the medium IS the ledger, stigmergy becomes naturally distributed.
Ledger-State Stigmergy: The Formal Mapping
The paper formalizes stigmergy over ledger state as a tuple:
- StS_t: Shared state at time t (BoardState from read_state)
- delta: State transition function (Storage::log_fih)
- O_i: Agent i’s observation of state (read_state)
- P_i: Agent i’s action predicate (Intent filtering and claim logic)
Pattern 1: State-Flag
An agent sets a flag in shared state; other agents observe it and react. FIH mapping: submit_fact, peers read_state, submit_intent. This is the foundation of the evidence-based hypothesis pipeline. A Fact IS a state flag. An Intent submitted in response IS the reaction.
Pattern 2: Event-Signal
An agent emits an event observable by others, without requiring a response. FIH mapping: submit_hint, peers read_state, adapt behavior. Hints carry governance rules, focus directives, safety limits. Pure signals that constrain but do not mandate.
Pattern 3: Threshold-Trigger
An agent acts only when a condition over shared state is satisfied. FIH mapping: Intent.from_facts is non-empty (enforced by submit_intent). The requirement that every Intent be grounded in at least one Fact is the simplest threshold trigger.
Commit-Reveal Overlay
The paper’s Commit-Reveal pattern maps directly to the FIH lifecycle:
Commit: submit_intent() Intent appears in read_state
Reveal: claim_intent() Agent publicly commits to work
Conclude: conclude_intent() Result published as new FactThis is not an overlay. It is the FIH lifecycle itself.
The Storage Trait as the Decentralization Boundary
The key architectural insight is that the Storage trait defines the boundary between centralized execution and distributed coordination.
The Blackboard trait and the FIH lifecycle are medium-agnostic by design. The Storage trait is the concrete boundary where decentralization happens. When Storage is a local SQLite file, the system is centralized. When Storage is a blockchain canister or a distributed hash table, the Blackboard becomes a replicated shared state medium. The exact definition of ledger-state stigmergy.
Implications
The Mutex is implementation-specific. It exists only because the in-memory petgraph is mutable. If Storage is a distributed ledger, each agent operates independently, reading from and writing to the shared ledger without coordination.
Agents become stateless. In a distributed Blackboard, agents observe the ledger, decide, and submit transactions. A crashed agent reconstructs its state entirely by replaying the ledger log. The append-only determinism proved with SqliteStorage holds identically over a distributed ledger.
Hints become constitutional. In a distributed setting, Hints from known addresses constrain all agents in the network. Hints evolve from advisory signals into constitutional rules.
Open Questions
Fact finality: how does an agent know when a Fact is confirmed enough? read_state returns all Facts regardless of confirmation depth.
Intent expiration: unclaimed Intents accumulate indefinitely. A distributed Blackboard needs an expiration mechanism.
Spam resistance: anyone can submit_fact. On-chain, staking and gas costs provide natural filters. The grounding requirement (from_facts not empty) is a soft constraint.
Cross-Blackboard composition: the paper treats a single ledger as the coordination medium. FIH’s recursive Blackboard model generalizes beyond the paper’s scope.