Replication Criterion

The question. If a machine needs a data center or a reactor, it is not self-replicating. That is the study's own name, stated as a test. This repository states the test precisely, applies it across the sibling repositories, and investigates the half of the dichotomy that none of them had touched: the power source.

Why it needed its own repository. Three sibling repos each answered a piece of the question and none stated the criterion that decides it:

• Fluidic Bootstrap Project: can a machine make its own control system from a bootstrapped workshop?
• copy-fidelity: does copying its own geometry drift, or does the three-plate process reset the error?
• closure-audit: when a study publishes a closure figure, does it disclose what it excludes?

Read together, they are three legs of one argument. The fourth leg, power, was missing, and the study was named for a dichotomy whose second half nobody had investigated. See audit.md.

The study this belongs to

This repository is one of four siblings written against a single question: what would it take for a machine to replicate itself with no semiconductor industry and no energy industry behind it.

• Fluidic Bootstrap Project: can the machine make its own control system?
• copy-fidelity: does copying its own geometry drift, or can precision be regenerated?
• closure-audit: does the literature disclose what its closure figures exclude?
• this repository: the criterion that decides the question, and the power leg.

Method, stated because it bounds every claim: literature and arithmetic only. The study ran no laboratory work, and each leg names the one measurement that would close it. A workspace-level index of the four repositories is kept locally by the author.

The criterion, in one paragraph

A machine is self-replicating only if every subsystem whose failure would stop replication can itself be produced by the machine, starting from a stated bootstrap kit and using only its environment (§1, §3 of protocol.md). The literature measures mass closure, and a study can report 90–96% mass closure while excluding 100% of the subsystem that makes it a machine, because a controller is a rounding error in tonnage. A different measure is needed, and this is it.

What is already established

• The data center is condemned, and there is a real replacement. Fluidic wall-attachment logic needs no contacts, no magnets and no rare metals, the only control substrate wholly inside a pre-industrial materials closure. Its cascade verdict was then corrected in the sibling's favour: the first reading inverted its own number, and on that number a stage has gain above parity and can drive a successor. The real limit is shallow fan-out, which this repository's findings-logic.md now prices (findings-gain.md, findings-control.md).
• Precision can be generated, not inherited: the three-plate method, which the sibling copy-fidelity has now confirmed is a rediscovery of the optics three-flat test, and has resolved the measuring-to-making transfer for: the effect is real but conditional on how the process is run, and it resets the workpiece, not the machine (copy-fidelity/findings-transfer.md).
• The literature discloses its boundaries and nobody demonstrates the brain being made (closure-audit, closed 2026-10-14).

What this repository adds

The power leg. It tested a hypothesis carried in from the cross-repo audit: that the two disqualifiers collapse into one, because the materials boundary that condemns magnets also condemns the generator. The leg partly refutes it. Direct heat-to-electricity and light-to-electricity conversion is condemned by the data-center rule, so "needs a reactor" collapses into "needs a data center" for that family. But the self-excited dynamo needs no permanent magnet, so the magnet boundary does not condemn a generator. What survives is shaft power and the generator, conditional on materials this leg could not close. See audit.md and findings-power.md.

Documents

protocol.md          the criterion, pre-registered: the formal statement, the
                     operational definitions of "data center" and "reactor", the
                     bootstrap kit K, and the falsifiers
audit.md             the criterion applied across the three sibling repos: the
                     four-leg map, and the synthesis
findings-power.md    the power leg: which prime movers sit inside the closure,
                     which are condemned, and the mobile-vs-stationary finding
findings-control.md  the control leg: does relay logic compose, and can it be
                     built in-closure? (it composes; its contacts were platinum,
                     reachable only from a placer in E, §7)
findings-gain.md     the gain leg: is the sibling's cascade verdict an inversion?
                     (gain is ~2–10, not 0.10–0.51; the limit is fan-out)
findings-logic.md    the logic leg, the count the study could not supply: how
                     many elements a replication controller needs, derived from
                     its structure instead of assumed
findings-tooling.md  the structure-and-tooling leg: iron has a measured cost,
                     tool-making is unmeasured
model/               standard library only; add `--verify` for self-checks.
                     `controller_budget.py`: the arithmetic behind
                     findings-logic.md, fluidic and relay. `closure_graph.py`:
                     all legs in one table, and which open quantities the
                     verdict turns on
NOTICE.md            sources, rights, and the evidence rule

Leg 2 (precision) has no file here: it is the sibling copy-fidelity, findings-transfer.md. This repository's audit.md Leg 2 summarizes it.

Status

All four legs investigated, and every one resolves the same way. The criterion is fixed (protocol.md); the cross-repo audit and the three legs that live here are written (audit.md, findings-power.md, findings-control.md, findings-gain.md). Where the literature is silent, each says so rather than estimating; the explicit ABSENT lists are findings-power.md §5 and findings-control.md §5.

The result, in one line: every leg ends at an unmeasured quantity, not at a demonstrated impossibility. That is the honest shape of the study.

Four results worth the reader's time, all corrections to the sibling study:

• The reactor half of the study's own name had never been investigated. Result: direct heat/light→electricity conversion is condemned by the same rule as the data center, but combustion, water and wind prime movers survive, and a self-excited dynamo needs no permanent magnet. A stationary replicator is not condemned; a mobile one is not shown to work.
• Relay logic was dismissed too fast. It composes, per-stage gain > 1 by architecture, and its coils and cores sit inside the closure, so the fluidic cascade wall is specific to fluidics, not to non-semiconductor control. Its contacts do not: period relays used platinum points, which is outside K and reachable only from a platinum placer in the environment (findings-control.md §7).
• And the cascade verdict itself was inverted. findings-gain.md shows the sibling read a required control fraction (0.10–0.51) as an available gain. Belsterling's Fluidic Systems Design (1971), read directly, defines a digital fluidic amplifier's gain as the reciprocal of exactly that ratio, and notes it can be infinite for a latching element. The gain is therefore ≈2–10, and the real limit is fan-out depth, with fan-out of 3 a worked design case in the same text.
• And the controller's size was assumed, not counted. src/labour.py priced the control system with GATE_COUNTS = (20, 100), tagged "not known". findings-logic.md derives the count from the controller's own structure: about four elements per step of the cycle, so a cycle of 10 to 50 steps costs 45 to 200 elements. The two assumed counts invert to a 6-step cycle and a 26-step cycle, which makes the sibling's floor too low, and it puts the control system at roughly 40 to 50 per cent of the labour of one generation, not the small share the floor implied. The fan-out limit the sibling called a wall appears in the same arithmetic as a doubling of the element count: a price, with a number attached.
• And the decisive measurements are not the ones the study named. model/closure_graph.py puts every leg in one table. Fluidic gain is not a single point of failure, because the relay route is an alternative. The only two open quantities whose failure alone condemns the replicator are tool-making (findings-tooling.md) and machine-geometry drift (copy-fidelity), because control and structure both stand on them.

Rights

Documents: CC BY 4.0. Sources are cited, never held, no PDFs, no extractions, no figures, no passage longer than a short phrase. See NOTICE.md.

Copyright © 2026 Thymeed.