Prior Art

NTL’s core mechanism — routing weights learned from delivery outcomes — is not new. It has a thirty-year literature, and pretending otherwise would be both dishonest and strategically foolish: the first question any informed reader asks is “isn’t this just Q-routing?” It partly is. This document says what NTL takes from its lineage, what it adds, and where existing systems are simply the better choice.

​

Summary

​

Reinforcement-Learned Routing

​

Q-routing (Boyan & Littman, 1994)

The direct ancestor. Each node holds a Q-table estimating delivery time to each destination via each neighbour, updated from the delivery times actually observed. Q-routing beats shortest-path routing under congestion, because it learns that the short path is slow when it is busy. NTL takes: the central claim, which is that routing learned from observed outcomes outperforms routing computed from static topology. NTL differs:

• Q-routing estimates delivery time to a destination. NTL estimates suitability of a synapse for a signal type — a different and smaller learning problem, because most NTL traffic is not addressed to a specific destination at all.
• Q-routing lives in the router. NTL lives at the application transport layer, so it can see semantics (signal type, tags) that a router cannot.
• Q-routing assumes a cooperative network. NTL assumes an adversarial one, which is why it carries influence caps and a threat model — the literature largely does not, because it predates open networks as a deployment target.

Honest assessment: if you control your network end to end and only need to minimise latency to known destinations, Q-routing’s formulation is simpler and better studied. NTL’s added machinery earns its keep when the network is open and the traffic is heterogeneous.

​

AntNet (Di Caro & Dorigo, 1998)

Ant-colony routing: probe agents deposit “pheromone” on paths that work, and pheromone evaporates over time so stale paths fade. NTL takes: two mechanisms almost directly. Weight decay is pheromone evaporation, and stochastic path selection is the probabilistic transition rule. The learning-model says so explicitly rather than reinventing the vocabulary. NTL differs: AntNet’s exploration comes from dedicated probe agents; NTL explores with real traffic, which is cheaper but means exploration costs real delivery attempts. That is why NTL’s exploration temperature is tunable and bounded, rather than a fixed probe rate.

​

Why This Lineage Was Never Widely Deployed

Learned routing has been known to work since 1994 and is essentially absent from production networks. Any honest positioning has to engage with why:

1. Convergence is hard to reason about. Operators prefer routing they can predict and debug over routing that is on average better.
2. BGP and OSPF were good enough, and the switching cost of replacing working infrastructure is enormous.
3. No adversarial model. The academic work assumes cooperative nodes. On the open internet, learned weights are an attack surface — which is precisely the gap threat-model exists to fill.
4. Nothing to learn from. Classical routing lacks a per-flow outcome signal at the transport layer.

NTL’s bet is that (1) and (2) do not apply at the application transport layer — there is no incumbent to displace and no operator expecting BGP-like determinism — and that (3) and (4) are addressable with signed receipts and per-identity influence caps. That bet may be wrong. It is at least a specific bet, rather than an assumption that the idea is new.

​

Content-Based Networking

​

Named Data Networking / ICN

NDN routes on content names rather than host addresses: a consumer expresses interest in a name, and the network finds it, caching along the way. Parallel: NTL’s Gradient scope is close in spirit — signals route toward nodes that have historically handled that type, which is type-directed rather than address-directed routing. Differences: NDN’s names are hierarchical and its caching layer is central to the design. NTL has neither: signals carry types and tags, not names, and NTL caches nothing (its dedup cache prevents loops, it does not serve content). NDN is the better choice for content distribution; NTL is aimed at signal transport with side effects.

​

Delay-Tolerant Networking

Directly relevant to NTL’s target deployments. DTN assumes no contemporaneous end-to-end path exists — intermittent links, long delays, frequent partition — and uses store-and-forward with custody transfer. This is the reality of the intermittent-connectivity deployments NTL is built for, and DTN’s literature (Bundle Protocol, RFC 9171; PRoPHET; Spray-and-Wait) is the most operationally honest body of work in this space. NTL takes: the premise. Partition is normal, not exceptional, which is why delivery-semantics does not pretend to offer guaranteed delivery and why timeouts are penalised weakly — on an intermittent link, a timeout is often not a routing failure. NTL differs: DTN forwarding heuristics are mostly hand-designed (encounter history, delivery predictability). NTL learns them. Whether learning beats a well-tuned DTN heuristic on a genuinely sparse network is an open question, and not one NTL has yet answered with data.

​

Peer-to-Peer and Messaging Systems

​

libp2p

The closest thing to a shared substrate. libp2p supplies transport negotiation, peer discovery (Kademlia DHT), NAT traversal, and pluggable crypto — much of NTL’s synapse and topology layers, already battle-tested. Shares: multi-transport, pluggable crypto, peer identity from public keys, no central registry. Adds: libp2p’s DHT routing is structural — it routes by XOR distance in key space, which is deterministic and ignores whether a peer is actually good at handling your traffic. NTL’s routing is learned from outcomes. NTL also adds activation backpressure and transform-at-synapse, neither of which libp2p addresses. Honest assessment: NTL could plausibly have been built on libp2p rather than beside it, and a reasonable reviewer will ask why it was not. The answer is control over the synapse abstraction — libp2p connections are not weighted, stateful, learning objects — but this is a real trade-off against a great deal of proven engineering, not a clear win.

​

MQTT

The incumbent for constrained devices. Lightweight, broker-based, three QoS levels, and enormous deployed reach. Shares: the constrained-device target; QoS 1 is NTL’s acknowledged class. Adds: MQTT requires a broker, which is a single point of failure and a deployment prerequisite. NTL is brokerless and peer-to-peer. MQTT’s topic hierarchy must be designed up front; NTL’s type affinity is learned. Honest assessment: if a broker is acceptable, MQTT is the pragmatic choice — mature, tiny, and universally supported. NTL is for cases where a broker is not acceptable: no reliable central infrastructure, or mesh topologies where “where do I put the broker” has no good answer.

​

NATS

Subject-based messaging with clustering, at-least-once delivery via JetStream, and excellent operational characteristics. Shares: subject/type-based routing, at-least-once semantics, lightweight wire protocol. Adds: NATS routing within a cluster is configured; NTL’s adapts per node from observed outcomes. NATS assumes a reasonably stable cluster; NTL assumes churn. Honest assessment: for a datacentre or a stable cloud deployment, NATS is more mature and almost certainly the better engineering choice today.

​

What Is Actually Novel

Stripped of marketing, NTL’s genuinely new contribution is the combination, not any single mechanism:

1. Learned routing at the application transport layer, with per-signal-type affinity. Q-routing learns per destination; NTL learns per type, which suits traffic that is mostly not addressed to a specific node.
2. An adversarial model for learned routing. The academic literature assumes cooperative nodes. Per-identity influence caps, signed reward receipts, and an explicit out-of-scope list appear to be new to this space, and are the parts most likely to matter in practice.
3. Activation backpressure as a first-class protocol mechanism. Threshold-and-refractory admission control instead of rate limiting.
4. Transform-at-synapse. Data reshaped in transit as a property of the connection rather than of an endpoint.
5. Storage as a specified interface rather than an assumed database.

Items 2 and 3 are the strongest claims. Item 1 is a modest extension of existing work. Items 4 and 5 are engineering choices rather than research contributions.

​

What Would Falsify NTL’s Thesis

Stated so the project can be held to it:

• Learned weights fail to beat static configuration on a realistic intermittent network. Then NTL is a worse libp2p with extra machinery.
• Influence caps prove insufficient against cheap Sybil identities in practice. Then learned routing is unsafe on open networks, and NTL is a consortium-only protocol.
• The exploration cost is too high. If probing with real traffic degrades delivery more than better routing improves it, the model is not worth its price.
• Nobody emits receipts. Without outcome feedback the model has no gradient, and NTL is an ordinary pub/sub system carrying unused machinery.

None of these is currently settled by evidence. The benchmarks exist to start settling the first and third.

​

References

• Boyan, J. & Littman, M. (1994). Packet Routing in Dynamically Changing Networks: A Reinforcement Learning Approach. NIPS.
• Di Caro, G. & Dorigo, M. (1998). AntNet: Distributed Stigmergetic Control for Communications Networks. JAIR.
• Zhang, L. et al. (2014). Named Data Networking. ACM SIGCOMM CCR.
• Burleigh, S. et al. (2021). Bundle Protocol Version 7. RFC 9171.
• Lindgren, A. et al. (2003). Probabilistic Routing in Intermittently Connected Networks. SIGMOBILE.
• Sutton, R. & Barto, A. (2018). Reinforcement Learning: An Introduction, 2nd ed. — contextual bandits and exploration policies.

---

Prior Art — The Bundu Foundation “Standing on the shoulders of giants is only useful if you admit whose shoulders they are.”