Summary

  • RFC 981 described an experimental AX.25 packet-radio algorithm that passively read other stations' source-route headers, built a graph of observed and inferred links, and ranked a primary route with alternatives.
  • The prototype distinguished a link actually heard from a downstream path merely inferred, penalized unverified links in its heuristic weights, and worked within small node and link tables on an LSI-11 system.

Somewhere in the packet-radio traffic around Washington, DC, a station sends a frame through a chain of repeaters. Another station hears it and writes down the route. Is that enough to send its own packet the same way? D.L. Mills's experimental Wiretap Algorithm, published in March 1986, behaved like a detective with a radio: listen, draw a map, mark the uncertain streets, and keep a second way out in case the first street goes dark.

The problem came from a broadcast channel with hidden stations. A source-route header listed callsigns from origin to destination, perhaps through several digipeaters. Hearing a frame told the observer something about the part of the journey that reached it. It did not grant magical hearing beyond its antenna. In the RFC's example, the listener could mark links on the observed way toward itself as heard; the continuation toward the destination was a presumption unless further evidence arrived. Certain frame types suggested a virtual circuit had been established and strengthened that presumption.

Radio gossip could be useful, but only if the listener remembered which detail was witnessed and which was deduced.

Direction made the detective work harder. A link heard one way was not automatically heard the other way; congestion and hidden stations could give the reverse route different properties. The prototype marked heard, reciprocal and synchronized evidence separately. Then it turned its accumulating node and link tables into a graph and calculated routes ordered by a weighted distance. It wanted the best current route and additional candidates to use if that route failed, or to discover a better one not yet tried. The word “best” belonged to the chosen estimates and weights, not to an all-seeing view of the radio channel.

Mills did not have infinite room for the case file. The LSI-11 prototype's node table was capped at 75 entries, its link table at 150. A node entry took 26 bytes, a link entry four. A weighted least-recently-used cache made room as observations arrived. In a 24-hour off-the-air example, the stabilized tables normally held about 60 nodes and 100 links. These were figures from this prototype and example, not a census of Internet routing. They make the constraint tangible: this listener had to decide what to remember before it could decide where to send anything.

The price list shows its bias. Each hop cost 30 in the prototype's heuristic. An unverified link added 50, a nonreciprocal link five, and a link without synchronized evidence another five. There were node penalties too, including complexity and whether a station had been seen acting as a repeater. The proposed dynamic congestion measure was still a plan in the memo, not an implemented feature of this prototype. The weights were educated first guesses, open to change.

Still, the central moral of the arithmetic was sharp: a route drawn from overheard fragments could compete, but it had to pay for what the listener had never directly observed.

The RFC Editor record dates this experiment to March 1986. The document reports a working prototype and initial testing on a demanding local packet-radio channel. Its reach was experimental, its antenna limited and its chosen scores heuristic. That limitation is exactly why it belongs in Internet history. Routing a packet sometimes meant managing incomplete evidence in real time, on a small machine, with the clock and the radio refusing to wait. The router eavesdropped; the honest part was that it knew when it was guessing.