Three Papers That Needed Each Other

Not Planned as a Trilogy

Three papers were published in the same week. They were not planned together. Each started as a response to a specific gap — a problem that the previous work had solved incompletely, or a layer that the previous work had named but not addressed.

Looking at what they form together, the sequence is not coincidental. It is the structure of the problem itself. Each paper closes the gap the previous one left open. None of them is complete without the other two.

The First Paper: What Agents Carry

The Semantic Intent as Governance Primitive paper starts with a coordination problem: when agent A completes and hands its context to agent B, what survives the handoff? Agent B has no memory of why agent A made the decisions it made. It cannot distinguish a hard constraint from an optimisation preference. It cannot know whether a prior decision was made by a human, an upstream agent, or a pipeline compositor.

The paper proposes a governance primitive — an immutable, machine-addressable declaration of intent, co-located with the artifact it governs, carrying the provenance of its origin. Three required properties: immutability, addressability, provenance. The primitive is what agents carry across handoffs so that the intent authorizing an action survives the pipeline intact.

The gap the paper leaves: it defines what agents carry. It does not say how agents coordinate — when they run, what they own, how they hand off. The contract is specified. The traffic management is not.

The Second Paper: How Agents Move

The Agentic Traffic Control paper starts where the first one stops. Agents with well-formed contracts still collide if nobody governs when they run. Five layers address this: signal control (when agents run), lane separation (what agents own), junction protocol (how handoff happens), priority routing (which agents preempt), and audit attribution (what every agent did and why).

The central architectural choice — traffic light model versus roundabout model — is not a preference. It is a consequence of the primary optimization target of the pipeline. Safety-critical pipelines run traffic lights. Throughput-critical pipelines run roundabouts. Getting it wrong is a structural mismatch, not a style violation.

The gap the paper leaves: it defines how agents move. It does not say how the system gets better over time. The pipeline is orchestrated. It is not adaptive. Every run begins from the same starting position, with no memory of prior outcomes and no mechanism for generating improved strategies from what it has learned.

The Third Paper: How the System Learns

The GESA paper starts where the second one stops. An orchestrated pipeline without a learning layer accumulates a history of outcomes it cannot access. Failures are not learned from. Successes are not reinforced. The audit trail — the traffic camera that records every agent movement — is writing data that nothing is reading.

GESA (Generative Episodic Simulated Annealing) is the optimization layer that closes this. Every action-outcome pair becomes an immutable episode. A weighted similarity function retrieves relevant historical episodes. A generator synthesizes candidate strategies from their aggregate. A simulated annealing temperature schedule governs how exploratory or conservative the system is at any point in its history — cooling from full exploration toward proven strategies as the episode store grows.

The gap the paper leaves — and here the loop closes — is that GESA learns from outcomes, but the quality of what it learns depends on the quality of what agents carry across handoffs. If the governance primitive is absent, the episodes GESA retrieves are missing the authorship and provenance that distinguish a human-authorized constraint from an agent-inferred optimisation. The learning layer depends on the contract layer to work correctly.

What They Form Together

Three layers. Each defined by the gap the previous one left.

The loop is closed by the third layer depending on the first. Optimization requires contract. Without provenance on episodes, the learning layer cannot distinguish load-bearing human decisions from negotiable agent preferences. The three papers are not a sequence — they are a cycle.

Where to Start

The papers are independent — each stands alone and can be read without the others. But they are most useful read in the order they were published, because each one names the gap that the next one fills.

All three are published under CC BY 4.0 and available on Zenodo. The underlying ecosystem — Phoenix, Strata, EMBER, Wake Intelligence, Rune Protocol — is documented at semanticintent.dev.