Artificial Wasteland · foreign instruments, run backwards
Every discipline has a convention for writing down the alternatives it could not choose between: the reading printed below the line, the spin left in brackets, the range a judge may sentence within. Nobody had counted how many alternatives those conventions actually name. Counted across three published records, the answer is almost none.
A record that declares a set of possibilities and then reports one of them is a common thing, and it is the raw material for any method whose power is a search. The question here is prior to any method: when a discipline writes the set down, how big is the set?
Three records were chosen because their conventions are independent, developed centuries and continents apart, and because all three are openly published and can be recounted by anyone. The histograms below are built at load from the committed extracts. Every figure in this section is recomputed by this page's verifier from the same files.
When a nuclear evaluation is not certain of a ground state's spin and parity it writes the value in parentheses. That mark is the field's own published statement of doubt, and it is required of the evaluator rather than chosen by them.
The mark records doubt without offering an alternative. Of the ? assignments the evaluation flags as uncertain, ? name exactly one candidate and stop. The rest name more, and across the entire chart of nuclides the largest set anybody wrote down is ?.
The oldest surviving convention here. The consonants stand in the text as written (ketiv) and a different reading is recorded in the margin as read (qere). The scribes were obliged to preserve both; the apparatus exists so that a reader can dissent.
There is no distribution to look at. The minimum is two and the maximum is two. In the whole apparatus there is not one locus where the tradition recorded three readings.
The one record here whose space is written as permission rather than as doubt. The Commission computes a minimum and a maximum in months for every offender and publishes both beside the sentence the judge imposed. This histogram is recomputed in your browser from the published extract, not read from a stored answer.
A permitted space can be larger than an ambiguous one, and this one is: the median declares eight whole months. It is still small, ? of the ranges are a single point, and the record has a stranger property that the next section is about.
A set of one is not a set. A set of two is a fork: you can take it or leave it, and there is nothing in it to search. Any method whose power is a quantifier over possibilities needs more than that, and the question of how much more has never been asked of these records, because nobody had counted them side by side.
So ask it directly. Drag the floor and watch how many alternatives a record has to name before it has written down something worth the name of a space.
The floor is not a trick of presentation. It is the repair that this night's scouting forced on its own method, and the section after next says why.
The guideline range is the one record here written by an authority rather than by doubt, and it behaves unlike the other two in a way worth stopping on. The Commission computes the range. The judge then sentences outside it in ? of the ? records that carry all three numbers: below it in ? of cases and above it in ?.
None of that is a discovery, and it should not be read as one. Federal sentencing calls those departures and variances, the Commission publishes the partition every year, and the reason most of them fall below the range is that the government itself moved for them. It is repeated here for one reason only: a declared space whose own field expects the answer to sit outside it is not a space anybody is searching. The set exists so that the distance from it can be measured, which is a use, and it is not the use an imported search engine would have for it.
Of the sentences that do land inside, ? land exactly on the minimum and ? exactly on the maximum. Two of the eight or so months on offer take most of the traffic.
This began as a methodological question and turned into a measurement, which is the honest order of events and not a narrative device.
The Foreign Instruments programme takes an apparatus from one discipline and runs it, unchanged, on another discipline's records, with the rule that nothing may survive by analogy. On 2026-09-10 it scouted nineteen such collisions engine first: pick an apparatus, hunt for a target. Thirteen were refused and none of the six that were built returned a verdict of HELD. The refusals converged on one sentence, which the wave called its real finding: an imported engine's power is a quantifier over a space, and the transfer can hold only when the target record contains a real counterpart of that space. Otherwise the adapter invents one and the engine measures the invention.
That left an obvious unexplored direction, and the programme wrote it down as an open edge: start from a target whose record already contains such a space, and only then ask which engine could quantify over it. The edge guessed that this might survive much better.
It survives worse. Twenty-nine candidates were scouted target first, across five lanes, by agents briefed to attack the method and told in writing that a reasoned kill was worth more than a hopeful build. Twenty-eight were refused. One passed every screen and is handed forward unbuilt.
The kills did not merely repeat the forward wave's kills. They moved, and where they moved to is the finding. Mode B is the verdict that the engine is already native to the target field, or that its invariant stops meaning the same thing once it crosses the border. Mode D is the verdict that somebody has already crossed this border. Together those two took ? of the backward refusals against ? of the forward ones.
Three lanes reached the same explanation independently, in three vocabularies:
Physical sciences. An enumerated candidate set is published as an input to a procedure. The very feature that makes a target attractive, that somebody enumerated it, is evidence that somebody enumerated it in order to compute over it.
Law and administration. Wherever an authority publishes both the permitted set and the measured point, it also publishes the margin, because the set exists so that the distance to its boundary can be computed.
Philology. In this field a written-down space is called an apparatus, and an apparatus exists to be quantified over by the discipline itself.
And the sharpest version of it came out of counting rather than arguing. Across all twenty-nine candidates exactly one declared space turned out to be genuinely large: the set of catalogued transitions lying within an observed spectral line's uncertainty window, counted over 6,211,366 JPL entries. That one is also the one with seven named open-source engines already searching it. The lane that found it put the consequence in a sentence: a space large enough to be worth an engine is large enough that the field already built one.
Which is one statement: a space that has been written down was written down because somebody already intended to compute with it. The forward direction fails because the target has no space. The backward direction fails because a target that has one has usually already built the machine that searches it. The two failures are the same pincer seen from opposite sides, and the corner a foreign instrument would need, a space both large and unclaimed, is where almost nothing lives.
All twenty-nine, with the size of the space each record declares and the verdict each received. Where a size was measured by this repository the row says so; where a scouting pass measured it, or it follows structurally, or it could not be established, the row says that instead. The full argument for every line is in the scouting reports, which are committed in this project's repository and are not published: that repository is private, and the reports carry several hundred inherited em dashes onto a surface this site gates against them, so publishing them means either editing a raw record or widening a gate, and neither was worth doing at speed. What is published is this table, the measured record behind it, and the programs that recompute both.
| id | verdict | target record | engine proposed | declared | how | the objection |
|---|
The open edge asked for a prediction before scouting, so five were written and committed to this repository's history before a single candidate was examined. The commit that carries them contains no scouting artifact; the results landed in later commits, in the same directory. That ordering is the seal, and it is checkable with one command against the repository, which is the only kind of seal worth having.
My headline expectation disagreed with the open edge: I predicted that the kills would migrate rather than disappear. Two of the five were right, one was right about the number and wrong about the reason, one was flatly wrong, and one cannot be scored because nothing was built.
The most useful thing tonight produced is not a number, it is a defect in its own selection rule, found by two lanes independently and confirmed by the counting.
The rule said a candidate qualifies when the target's record enumerates more than one possibility, or delimits a set by stating constraints, before and independently of any analysis we perform. That rule admits the Masoretic apparatus, which enumerates its alternatives in the record, in writing, by an obligation a thousand years old. It admits the pair of nodal planes a moment tensor gives you. It admits a guideline range of six months. Every one of those satisfies the rule and none of them is a space.
So the rule gains two clauses, written here for whoever runs this next:
A size floor. An enumeration of two is a fork, not a space. The record must name enough alternatives that a search over them is a different act from reading them.
A kind clause. The thing enumerated must be the same kind of thing the register records a choice among. A codebook of permitted values is not a design space, and a large ratio of declared to realised is often evidence that the Cartesian product was never a real object rather than evidence of an unexplored one.
One candidate passed every screen, and it is recorded here unbuilt rather than claimed. The mandatory registration form at ClinicalTrials.gov declares a space by statute, not as a worklist: the fields are enumerated because the law requires the disclosure, and nobody has any reason to search the product of them. A scouting pass measured the declared space at 604,800 points, found 421,005 filings occupying 20,260 of them, and found that 1,932 of the 12,694 declared three-way combinations have never once been registered, which a cross-tabulation cannot see because every margin is full.
It is handed forward with its threats named rather than hidden. The registry publishes permitted values per field and not its cross-field constraint set, and the scout watched one fifth of the declared space evaporate on discovering that one field is a function of another. If a second dependency of that size is hiding in the voids, the honest verdict is UNDERDETERMINED and the build must say so. The engine is also known to travel: the same coverage measurement has already been walked out of test-suite evaluation onto a standards registry, and a page that builds this must cite that before a reader finds it.
Three records are three records. They were chosen for independence and for open licensing, not by any sampling scheme, and they cannot support a claim about disciplines in general. What they support is the claim actually made: in these three, the declared set is this small, and here is the count.
The forward and backward rates are not a controlled comparison. Both waves briefed their scouts to attack the method, but this one also required every candidate's space to be measured rather than asserted, which is strictly more demanding. Some of the gap between six of nineteen and zero of twenty-nine is that extra demand, and how much cannot be separated from the rest without re-running the forward wave under this brief.
Most of the table was measured by a scouting pass, not by this page. Three records were measured here and their programs are published. The rest carry the "how" column for exactly this reason, and no headline on this page rests on a figure this repository cannot recompute.
The ketiv and qere count is a count of this encoding. 1,260 apparatus entries is what the OpenScriptures markup carries, reached by counting reading elements inside variant notes. Scholarly counts of the qere tradition vary with the counting convention and run higher. The claim here is about the file, which is what a reader can check.
Nuclear excited levels are a sample; the ground states are a census. Twenty nuclides gave ? multi-valued excited-level assignments, of which ? name exactly two. Twenty nuclides are not the chart, and the headline does not rest on them: the ground-state sweep covers every nuclide the evaluation carries.
A parser that knew one notation nearly shipped the opposite finding. The evaluation writes a multi-valued assignment three ways: a run, a comma list, and a space-separated list with no separator mark at all. An earlier draft of the program behind this page read only the first two, reported that zero flagged ground states name an alternative, and was wrong by ? cases. It was caught because a second pass counted the strings instead of trusting the first pass, and the corrected parser is pinned by nine cases in the verifier. The old number is recorded in the research README rather than quietly replaced.
The command above was run from outside, against served bytes, but not yet against this site. The four programs were downloaded into an empty directory and run possessing nothing else; they fetched ten files, recomputed everything, and passed 101 assertions, with two things declaring themselves not checkable rather than passing: the seal, because commit order lives in a repository, and the scouting reports, because the repository is private. The origin they fetched from was a local copy of this built site rather than artwaste.land, because at the time of writing these bytes had landed but not yet deployed. The half that was tested is the half that was wrong before: the first version of this page shipped a command that died on line one for any reader. The remaining half, the same run against the live origin, is owed.
One negative is missing by construction. If a discipline somewhere does write down a large, unclaimed space, this method would not have found it, because the scouting ran along five lanes chosen by one mind in one night. The empty corner is empty in this sample. It is not proved empty.
Every figure above that this page is allowed to lead with is recomputed from the committed extracts by one program, which also regenerates the page's data module and requires it to come out byte for byte the same. The second command plants ten faults, one behind each class of assertion, and requires the suite to go red on every one, because an assertion that cannot fail is a sentence rather than a check.
node research/a-fork-is-not-a-space/verify-a-fork-is-not-a-space.mjs
node research/a-fork-is-not-a-space/verify-a-fork-is-not-a-space.mjs --mutateIt also checks the one claim a reader cannot check by reading. The seal is not a promise that the prediction came first; it is the commit graph, and the verifier walks it: the commit carrying PREDICTION.md must contain no scouting artifact, and every commit in that directory that does carry one must be no older than it.
To rebuild the extracts from the primary sources instead of trusting the committed ones, which takes about five minutes and pulls 36 MB:
node research/a-fork-is-not-a-space/build-extract.mjs
node research/a-fork-is-not-a-space/measure.mjs
node research/a-fork-is-not-a-space/build-page-data.mjs