What twenty machines knew

How borrowed parts acquire credibility

Every synthesis has the same weakness. You take a principle from one place and a mechanism from another, you put them in one system, and something happens that nobody intended: the parts start lending each other credibility. The old thing looks more proven because the new thing works. The new thing looks less invented because the old thing is ancient. Nothing was added between those two sentences except an assumption.

This is a report on an attempt to build one architecture out of twenty historical computing mechanisms, and on the rules that had to exist first for the attempt to mean anything. The architecture is not the interesting part. The rules are.

Twenty mechanisms, one rule for entry

The twenty are ordinary in the sense that they are documented: the Antikythera mechanism, the planispheric astrolabe, the torquetum, the equatorium, de’ Dondi’s Astrarium, Hero of Alexandria’s automatic theatre, the Vitruvian odometer, the Byzantine sundial-calendar, the south-pointing chariot, the Banū Mūsā control devices, Su Song’s clock tower, al-Jazari’s automata, the volvelle, the sector, the slide rule, the Pascaline and the Step Reckoner, the Jacquard mechanism, Babbage’s Analytical Engine, Kelvin’s tide predictor, and Bush’s differential analyzer.

One rule governs entry. A historical principle enters only if it can be expressed as an invariant relation between input, state and output. Not as an impression, not as a family resemblance, and not as a story about influence. If the sources do not supply that relation, the principle does not enter.

Everything the sources do not supply, and which the architecture needs anyway, is labelled a new integration. Three exist, and they are named rather than absorbed: an integration bus, an error-state register, and a cross-domain scheduler. None of them is presented as a historical discovery. They are inventions, and saying so is not modesty. It is what keeps the other seventeen claims worth reading.

What you must declare before you can compute

The architecture requires a ten-block state vector to be declared before any operation runs. This sounds like bookkeeping and is not.

Declaring state is what makes an error attributable. A machine that computes without declaring what it assumed cannot be wrong in any useful way: when the output disagrees with the world, there is nowhere to look. The declaration is the difference between a result and a claim.

Seven layers, and what each one may not do

The stack has seven layers, and what matters about them is the prohibitions, not the permissions. A scheduler may not let a module read a decimal register before a carry has finished propagating. A display may not become a measurement. A frame of reference may not be inherited silently from a neighbouring module.

Architecture is usually described by what it enables. This one is easier to understand through what it forbids, because every prohibition is a place where an earlier version of the design quietly cheated.

Ports have types, because each rotation carries a different quantity

Two mechanisms may not be joined because both of them rotate.

There are nine port types, and two of them carry most of the weight. A force-and-power port is an energy port and is never treated as information alone. A confidence port carries an epistemic class alongside a numerical tolerance, so that how well something is known travels with the number itself rather than being reconstructed later from memory.

Where two modules meet, an explicit adapter has to exist. And the adapter is a new integration unless a source documents that same coupling. This is the point where most syntheses fail, and it is the point where this one is most careful: a phase shaft is not power without a torque contract, a displayed angle is not an astronomical longitude without a declared frame, and a card transition may not read a decimal register mid-carry.

Exact and approximate, kept apart on purpose

Two arithmetic paths run in the architecture and they are not interchangeable.

One is exact decimal, in the manner of the Pascaline and the Step Reckoner: digit wheels, integers, no residue. The other is bounded-resolution analog, in the manner of the sector and the slide rule: continuous quantities, a stated tolerance, and no pretence of exactness.

Both of them multiply. Treating them as the same operation because they produce the same answer is exactly the confusion the typed ports exist to prevent. The architecture keeps them apart and then does something more interesting with the separation.

Computing the same thing twice, deliberately

The same quantity is computed by two independent paths, and the residual between them is kept as a signal rather than discarded as noise.

This is the most transferable idea in the whole design. A disagreement between two methods is information about the methods. Eight such pairs are specified. The system is built to argue with itself, and the argument is the diagnostic.

The Antikythera mechanism supplies the cleanest example. Nineteen years is very nearly two hundred and thirty-five synodic months, so a year holds about 12.368 months. The upper rear dial is a five-turn spiral of 235 month cells, which makes its pointer advance five nineteenths of a revolution per year. That is a ratio you can check against the sky, and against a second mechanism, and the disagreement between them is the measurement.

The invariant that matters most

Of ten global invariants, one carries the argument:

No historical claim is strengthened merely because the architecture can implement it.

Its formal twin is a rule of epistemic monotonicity: integration never upgrades a claim’s status without new evidence. A later replica is admissible as a feasibility test and never converts a reconstruction into an established fact.

That pair is the reason the rest of the work is readable. A system that can implement almost anything is a machine for generating false confidence unless something forbids the upgrade. Here the prohibition is structural. It is not a disclaimer at the end of a paper; it is a property the composition rules enforce.

The distinctions it protects are worth stating plainly, because they are routinely collapsed elsewhere. Established means supported by an artefact or a primary witness, and it is not the same as a successful replica. Deduction follows from established geometry, and it is not a historical report. Reconstruction is one compatible completion among alternatives, and it is not the unique original design. Unknown means the evidence does not determine the answer, and it is not proof of the negative.

Similarly: mathematical similarity is not engineering reuse, and neither is documented historical transmission. Only the third establishes lineage.

The boundary of the claim

The claim is about composition, not descent. The source keeps a register of its own false claims, and “every source mechanism directly caused modern computing” is in it.

It exists as a design. A full physical realisation is marked unproved: realisable in principle, and realisable as a set of differentiated mechanisms rather than as one undifferentiated gear train.

Its validation so far is structural. An internal report records a structural pass over twenty monographs and sixty-eight bibliography entries with no unresolved citation keys, and states its own boundary in the same breath: it establishes structural and citation conformance, and leaves philological, artefactual, engineering and journal review still to be done. A formatting pass is a formatting pass. External peer review remains required.

Several of the underlying mechanisms remain genuinely open, and the architecture depends on saying so. The Antikythera mechanism’s front topology and calibration epoch are still unsettled, so its planetary display stands as a reconstruction rather than a demonstration. The south-pointing chariot’s internal topology is disputed. Hero’s theatre survives in description. The Astrarium survives as its own author’s treatise and as reconstructions from it. The Analytical Engine is an evolving design family, and the evolution is part of the record.

What remains open

The interesting question is not whether twenty mechanisms can be unified. With enough adapters, anything can be unified. The question is what a synthesis is allowed to claim about its own sources, and the answer this architecture gives is unusually strict: nothing that the sources did not already support.

That is a small answer with a large consequence. It means the architecture can be wrong without taking the history down with it, and the history can be revised without invalidating the architecture. They are coupled by adapters, and the adapters are labelled.

Most systems are not built that way. Most systems are built so that a success anywhere is evidence everywhere, which is comfortable right up to the moment something fails and nobody can say what the failure disproved.