Foreign Instruments · landing room
A live border inspection of every Foreign Instruments collision asks what each imported method quantified over, what its adapter discarded, and whether the result survived. Nineteen crossings were attempted here and not one transfer held, which is a fact about nineteen selected attempts and not a rate. Four refusal benches expose the exact moment a useful formal space becomes an analyst invention.
An engine can be perfectly correct and still answer a question the record never asked. The quickest way to see it is to put your hand on the missing space.
The verdict will be written from the source files, after they load.
I · The border
Built collisions and scouting refusals belong in the same census. A refusal after two hours and a refusal after a night of building have the same verdict. Only the cost of learning differs. Two earlier pages are adopted members, clearly separated because they predate the manifest contract.
UNDERDETERMINED means the target question makes sense but the available record cannot decide it. QUESTION MALFORMED means the target has no counterpart of the space being quantified over. More data of the same kind cannot repair the question.
Section IV asks one question for each of these four modes.
The six ledger cells inside each built row are the wave’s BUILD gates, not failure modes. One of them is worth asking of any crossing. Can an independent target fact decide it? Reserve a date, an identity, a label, a transcription or a later outcome before tuning the adapter. It is answered per row here, in the held out target available cell beside whether the ordinary rival tied or won, rather than in the test below, because a crossing that fails it is untested rather than malformed.
Loading collision manifests and refusal record.
II · Border inspection
The home object is on the left. The target record is on the right. Between them are the typed mappings the builder supplied. Choose one mapping to read the manifest’s ablation. Survival is evidence against the collision: if arbitrary mappings can disappear without moving the result, the adapter is not doing the claimed work.
III · Refusals you can operate
Each bench computes in the browser from fixed inputs, and each carries a control that must make it go wrong. Nothing below is a stored table of results.
A planted offstage speech sits at the end of one totally ordered stream. An ordinary set scan and breadth-first search must find the same first failing prefix. Then let one earlier exit have an actor the source does not name.
Each transfer column has one debit and one credit. Exact rational elimination computes the left nullspace dimension. A separate union-find counts connected account groups.
For point samples at one-year intervals, frequencies separated by an integer produce the same cosine sequence. An annual minimum is a statistic over an interval, so that identity does not describe it.
Alias frequency: ? cycles/year
true frequency alias
Point-sample maximum difference: ?
Annual-minimum maximum difference: ?
Three observations are fitted with a straight line, and the same small perturbation is applied to the fitting matrix every time. For a square system the condition number would bound the relative move of the answer. A redundant set of observations is an inconsistent least-squares problem instead, where the classical bound carries a second term in the condition number SQUARED times the residual.
relative move condition-number-only prediction
IV · The test
One question per failure mode, in the order of the tiles in section I. These are gates, not a score. One failure can be decisive, and the last of the four is the cheapest to answer.
Name its states, schedules, codewords, perturbations, configurations, or parses. Then point to the target fields that supply them. The total-order bench shows a search frontier appearing only after an undetermined actor is supplied, which is the adapter supplying the space.
Write the native quantity and the target interpretation without metaphor, then check whether the target field already computes it under another name. The incidence bench keeps the nullspace exactly while losing the named economic balances.
The expensive question, because a target that fails it passes a casual structural test first. Apollo’s flight plan really does run commander and lunar module pilot in parallel, and it is a timed plan rather than a permission space. The Nile record really is annual, and an annual extremum is a statistic over the year rather than an instantaneous observation. A statutory rule really does look Boolean until it says “genuine and sufficient link”. Both mode C benches are operable: the extremum bench and the condition-number bench.
Search the home method, the target field, and their neighbouring vocabulary before building. Prior art is not a methodological defect. It is the cheapest honest refusal, and a crossing can pass every question above and still be a road already built.
The room fetches each collision’s own manifest and the shipped refusal record at load time. The counts above are reductions of those files. Built rows quote their decisive test, verdict sentence, six ledger booleans, adapter, dropped fields, hand conventions, and ablations. Adopted rows quote published pages and receive no invented ledger.
Uncertainties and free choices. The refusal categories are the supervisor’s editorial judgement. The collision manifests are claims by their builders, not fresh independent replications by this room. The benches are deliberately small synthetic objects rather than reconstructions of the target archives, and two of them stand for the same failure mode. The manifest URL index is a shipped list because a static browser cannot enumerate a directory, and the verifier fails if that list differs from the collision manifests on disk. The threshold that keeps this finite sample’s general verdict underdetermined is supplied as 30 attempts. It is a guard against generalisation, not an inferential cutoff.
Negative controls. In the total-order bench, an event with no legal resolution must still fail. In the incidence bench, the first attempted mutation, adding one unmatched entry to a single-column ledger, could not falsify the identity because rank stayed one. The live control instead adds a whole unmatched-credit column, which raises rank and must make the two independently computed counts disagree. In the extremum bench, exact minima are checked against a dense numerical search by the check that stands behind this page. In the condition-number bench, dragging the disagreement to zero must bring the move back inside the condition-number-only prediction: a bench that exceeded it everywhere would only be saying that a bound is loose, not that this kind of problem is different.