Dispatcher decoding

Dispatcher decoding is the interpretation of the signals of the Dispatcher network by machine-learning models, developed by H&A from 2022. The models were trained on intercepted signals paired with human observers' records of what the parasites did after each signal, and they give each reading as an estimate with a confidence figure. Dispatcher signals have no key and no fixed code, and are not deciphered: the meanings the decoders return are the categories that observers recorded, matched to the signals by correlation. The largest decoder is a block of the Beast, H&A's internal model, and runs only in a few guarded data centres. Smaller readers are licensed to other companies, among them Lumen.

The first working decoders, in 2023, were the first means of anticipating the movements of parasite organisms from their own network, and they renewed investment in large AI models in both the European Federation and the United States (see AI race).

The signals

Recording

Dispatcher signals are received by antennas at H&A listening posts along the edges of infested ground, and recorded as radio-frequency traces. Signal-processing software converts each trace into a string of discrete symbols, by frequency, timing and intensity, and the decoders read only these strings. A short dispatch transcribed this way and its reading look like this:

|J8C@vz&iZ=Xy>Dk/Vv
→ 3 Rupters moving northeast, elevated (confidence 71%)

Properties

A Dispatcher signal carries meanings, such as a direction, a count, an urgency or a kind of organism, and not letters or words, so there is no alphabet to attack. The same meaning is expressed by many different symbol strings, and different meanings sometimes share the same symbols. Meanings that share symbols are almost always of the same kind: compass directions are confused with neighbouring directions and "a few" with "several", but a direction is never confused with an organism. Parts of a message are sometimes missing altogether, and some symbols carry no meaning at all.

The relation between symbols and meanings changes slowly over time. Dispatchers do not use a fixed code that changes on a schedule; the mapping drifts continuously, and a decoder trained on old traffic reads new traffic less and less well.

H&A's early documents called the signal "the cipher", on the assumption that a message lay underneath that could be recovered with the right key. After 2023 the company's documents call it "the channel".

Development

Observation and annotation

The work began in 2022, after H&A Hyperbiology showed that Dispatcher emissions changed in step with the movements of nearby organisms. Observers at the edges of infested ground recorded Dispatcher signals and, alongside each one, what the parasites did in the minutes and hours that followed: how many organisms moved, of what kind, in which direction and how fast. Their notes were reduced to a fixed vocabulary of categories, and each recorded signal was labelled with the categories of what followed it.

The annotation became a worldwide effort. Military units, clearance teams, H&A field staff and contracted observers in held and contested territory have contributed to it, and as of 2028 H&A employs or contracts about 60,000 annotators, many of them in low-paid shift work checking and correcting labels. The corpus holds several billion annotated intercepts.

Training

The decoders are trained on the annotated intercepts, with the symbol strings as input and the observers' categories as output. Given a new signal, a decoder returns the most probable reading, weighted by what Dispatchers in that region usually signal, and a confidence figure for it. The models have no language on the input side and cannot be addressed in words; they read symbols and return categories.

The first decoder able to forecast organism movements better than chance was tested on the northern Adriatic coast in 2023. Decoders were in regular military use by the end of the year. In 2025 H&A moved its main decoder into the Beast, where it occupies a block of the model whose size has not been published.

Calibration and testing

Real traffic cannot be graded directly. H&A grades its decoders on a held-back set of annotated intercepts, and on synthetic test signals produced by its own software, whose meanings are known in advance. The confidence figures are calibrated against the annotated set, so that a reading given at 70% is correct about 70% of the time on traffic of the kind the decoder was trained on.

The same dispatch is often intercepted many times as it passes between Dispatchers, and a reading drawn from several intercepts is more reliable than one drawn from a single intercept.

Retraining

The drift of the channel means that a decoder's accuracy falls from the day its training ends. A decoder left without retraining loses about half its accuracy within a year, and old intercepts are read better than new ones. H&A retrains its decoders continuously on fresh annotations.

Purification

The observers' records describe everything that followed a signal, while the signal may have carried only part of it: Dispatchers drop meanings from their messages, and the dropped meanings remain in the annotations. A decoder trained on these pairs learns to connect symbols with meanings they did not carry, and returns meanings that were never in the signal. Since 2026 H&A has worked to remove these false connections, with the aim of a decoder that returns only what a signal contains.

The work is done by Parsec, an internal model of Lumen's Lumi family that has not been released to the public. H&A built its own language models for this work at first, and now trains the Beast and its decoders with Parsec. For each annotated intercept, Parsec estimates which of the recorded meanings the signal actually carried, and reduces the weight of the rest in the training data. The decoder is retrained on the corrected corpus, Parsec repeats its estimates against the new decoder, and the cycle continues. The decoder takes no part in correcting its own training data.

Each round is checked against signals whose content is known. Some are synthetic signals produced by H&A's own software. Others are encoded by the Dispatchers of the 2nd and 3rd Vagrants, who have worked with H&A since 2024: messages of agreed content are sent through their Dispatchers and intercepted, and H&A records what was encoded and which meanings were lost in transmission. A decoder version passes when it returns the meanings that were encoded, and fails if it omits them or adds a meaning that was not there.

The cycle has run at the scale of whole data centres since 2027, and is the largest single use of the Federation's computing capacity.

Limits

A decoder can only return meanings that observers thought to record. Whatever Dispatchers signal beyond the vocabulary of the annotations is not in any reading, and analysts have found structures that recur in signals across the world and correlate with no recorded observation.

There is also a limit set by the channel itself. Meanings that share symbols, parts of messages that are missing and the continuous drift destroy information that no amount of computation recovers, and accuracy approaches a ceiling below certainty. Each further point of accuracy costs more computing capacity than the one before. H&A has defended the cost of its data centres on the ground that each point of accuracy is measured in lives saved.

Confidence figures are reliable on recent traffic of familiar kinds, and less so on stale or unusual signals, on which the largest model has been found to be overconfident.

Readers

The Beast

The decoding block of the Beast is the largest decoder in operation. Its signal weights are dense, and every input passes through all of them; certain inputs are also routed to further blocks of the model, at much greater computational cost. It runs only in a few heavily guarded data centres, each of which it occupies entirely, and serves H&A's divisions and the joint H&A-EF command.

The Beast's output is not read directly by people. Other H&A models take its readings, in an internal representation or in plain language, and work out what each means for the front concerned and how it should be acted on. The results reach military staffs as intercept reports, with the confidence figure for each reading.

Licensed readers

Since 2026 H&A has licensed smaller signal readers, distilled from its decoders, to other companies. Lumen's Lumi models can call one to read a few intercepted signals at a time. The licensed readers are far less accurate than H&A's own decoder and cannot process traffic at the scale of a region, but the language model that calls them can explain a reading and discuss what lies outside its categories.

Use

The joint command's forecasts of parasite movement draw on decoded intercepts together with reconnaissance and Deterrent reports. Decoded traffic gives warning of organisms moving toward a position, of the arrival of reinforcements from a Dispatcher's stores, and of attacks on rail lines and convoys. In 2027 an H&A signals team tapped the Dispatcher relay network, and analysis of the traffic it intercepted showed that the behaviour of the Primitive Beckons of 2020 and 2021 had been a controlled test of how people and governments respond.

Reception

As of 2028 decoder readings are attached to every intercept report that reaches the joint H&A-EF command. Critics have pointed to the confidence figure printed beside casualty estimates in intercept reports, which reads the same whether the reading is well calibrated or drawn from stale traffic, and to the conditions of the annotation workforce. The name of the Beast's decoding block, which is sometimes called its "Dispatcher block", is often confused with the organism.

See also

• Dispatcher
• The Beast (language model)
• H&A
• AI race
• Lumi (language model)