14 · Predictions — class, grade, falsifier

Anyone writing about a development has to say what they expect, how certain they are, and how you would tell that they got it wrong. Everything else is commentary. The thirteen chapters before this one counted, quoted and dated; this chapter sets out what follows from the count — and binds itself to a rule that makes it less comfortable than it would have to be.

The rule, and where it comes from

The rule does not come from the tax world. It comes from the re-assessment of a well-known forecasting text. We held Leopold Aschenbrenner's Situational Awareness against reality claim by claim, twenty-five months after publication, and counted the result: 2 HIT · 1 PARTIAL · 3 MISS · 3 TOO-EARLY. Nine claims, two of them borne out, three wrong, three not yet decidable at the time of the check.

The hit rate is not the interesting part. The interesting part is that the distribution was not random. The hits all sat on the same kind of claim: the extrapolation of a measured quantity — compute, power consumption, geopolitical attention. The misses likewise all sat on the same kind: a threshold that was to be reached by a date. Extend a curve and you were right. Say when it crosses a given height and you were wrong — regardless of how good the argument was.

The sentence that moved from there into our confidence rubric reads:

«Inputs extrapolieren, Outcomes nicht.» (extrapolate inputs, not outcomes)

— the system's confidence rubric, § «Forecast-class ceiling», introduced 01.08.2026

Three ceilings follow from it, and they apply in this text without exception:

• Trend extrapolation on a measured input can reach grade A. Condition: the input is a number we counted ourselves, with a date and a method.
• Threshold-by-date caps at B, however strong the argument. Not «B, wenn das Argument mittelmässig ist» (B if the argument is mediocre) — B as the ceiling, even with a very good argument. That is the point of the rule.
• Category disappearance — the claim that something ceases to exist — caps at C, unless a structural forcing mechanism is named. If one is named and it is published and dated, B is within reach.

A rule you only follow when it delivers the result you want is not a rule. So we say it here explicitly: for P2 and P5 below, one could argue that the evidence base is stronger than the grade. We do not try. Nothing carries the credibility of this text as much as following our own rule visibly, precisely where it hurts.

The format

Every prediction sits in one row with five columns: claim · class · grade · falsifier including test method and test source · test date. No prediction without all five. A falsifier that names no test source is decoration; it cannot be applied. A missing test date makes the claim unfalsifiable, because it could still come true.

The eleven predictions

Group 1 — Zürich

P1 is the sharpest of the eleven, because it is decided fastest and because the canton set the date itself: slide 6 of the information event of 25.08.2025 carries the heading «E-Bilanz ab Steuerperiode 2025» (E-Bilanz from tax period 2025). The Wegleitung 505 for exactly that tax period names ZHcorporateTax seven times and the E-Bilanz not once. We do not claim the import is not coming. We claim that no instructions for it are published by the time the deadline expires — and we say how to prove us wrong.

P10 is the only one of the four carrying an A, and it is deliberately the dullest. A number that has grown by around 2'680 a year for five years very probably grows in the sixth as well. That is exactly what A is for.

Group 2 — the federal level, the standard, transmission

P4 is the second A prediction, and it is one for the same reason as P10: it extends a counted curve. Three versions of the same standard in twenty months, two of them with a changed set of elements, an element diff we computed ourselves. We do not say what the next version looks like. We say that it moves.

P8 deserves a note on the grade. One could argue that a forcing mechanism is named here and the ceiling therefore does not bite. It bites anyway, because in its form the claim is a threshold with a date: it says when something happens. The forcing mechanism lifts a C prediction to B; it does not lift a B prediction to A.

Group 3 — Switzerland as a whole

P3 is the prediction that gives the chapter its title, because it is the only one whose test source a third party runs and keeps updating. We counted the matrix across all 26 canton rows on 03.07.2026; anyone counting it again on 31.12.2027 arrives, by the same method, at a number that confirms or refutes us. That is the difference between a prediction and an opinion.

P7 is our favourite prediction, because it runs against the expectation the word «maschinenlesbar» (machine-readable) creates. Four productive tax periods in Obwalden, and the signed PDF Jahresrechnung has never disappeared. Nor has it in Appenzell Ausserrhoden, and Zürich invokes the Obligationenrecht for it. The structured file fills the form; it does not replace the document.

And the one C prediction

This prediction is expressly graded C, and the reason can be named: Zürich has published no forcing mechanism for the JP tax return. § 1 Abs. 3 of the Verordnung says only: «Die Steuererklärung kann auch in Papierform eingereicht werden.» (the tax return may also be filed on paper) No expiry date, no transition period, no announced repeal. Without a named mechanism the rubric caps at C, and C means: we say it, we mark it as weak, and we claim no more than we can evidence.

What a published forcing mechanism looks like is shown by another canton, and the comparison belongs here because it shows that the grade is not a question of temperament but of the record. Minutes of the Erfahrungsaustausch between the Steueramt of the Kanton Solothurn, EXPERTsuisse and TREUHAND|SUISSE of 25.11.2025, Traktandum 10.1, verbatim:

«Fristverlängerungen • ab dem Kalenderjahr 2026 aus Dr. Tax online • ab dem Kalenderjahr 2027 nur noch online und via CSV-Datei • ab dem Kalenderjahr 2028 nur noch online übermittelt werden können.» (deadline extensions can be transmitted: from calendar year 2026 from Dr. Tax online; from calendar year 2027 only online and via CSV file; from calendar year 2028 only online)

That is a forcing mechanism: a sequence of stages with years, minuted and published. For deadline extensions, not for the tax return — but dated. For the JP tax return, Zürich has published nothing comparable. Hence C.

The verification infrastructure

Falsifiers nobody can check are rhetoric. Checking this text runs over the same ten sources, the same eight word-boundary terms and the same baseline of 04.08.2026, 20:58 MESZ that we used for the zero counts. The eight terms are eCH-0276, XBRL, E-?Bilanz, CH-Taxonomie|OR-Taxonomie, Instanzdokument*, swissdec, Schnittstelle*|API, Drittsoftware*|Drittanwendung*. Counting uses word boundaries, never a substring — a naive search for ech finds ten hits in the Obwalden handbook, and all ten sit inside Jahresrechnung, Obligationenrecht and entsprechend.

The rule the check reports by sits in the source code of the checking tool and applies to this chapter in exactly the same way:

«Silence means nothing changed — never that the check failed. A fetch failure is reported as manual-check-due, never as green.»

And the qualifier from Part 12, which belongs here again because it otherwise disappears exactly where it counts: the watch does not run on a schedule today. There is no entry in the scheduler, no cron line, no entry in the register of expected runs. It ran once, to write the baseline. We may say: built, and given a baseline on 04.08.2026. We may not say: daily. Anyone wanting to check whether we run our own checks now holds the test source for that question.

What we do when we are wrong

Eleven predictions: six capped at B, four at A and one at C. After the Aschenbrenner re-assessment, a share of them is to be expected not to come true, and preferentially the share with a threshold and a date. So we say in advance what happens then.

This text is updated when the underlying material changes — not on a publication rhythm. There is no quarterly version and no annual edition. If a source changes, the text changes; if nothing changes, it stands.

Changes are noted at the head of the document, with a date and with a statement of which source moved.

A falsified prediction is left standing and marked as falsified, not deleted. Remove a prediction after it has turned out to be wrong and you publish, retroactively, a better hit rate than you had. The use of this chapter to a reader lies precisely in being able to read, two years from now, how often we were off and on which class of claim. We have already done this elsewhere: one of our own notes of 04.08.2026 opens with the sentence that its predecessor's headline was falsified six days later by two primary sources. The predecessor is still there.

Three sentences to the cantons

What follows is opinion about the situation. It is marked as such and contains no statement about the conduct of any named person or authority.

First. The distinction between the two formats is cleanly settled in the technical material — the communication matrix of the Transmitter specification assigns each route its artefact — and in the communication addressed to taxpayers it is not visible. A short public statement of which artefact a portal accepts, in which version and with which validation, would close the open question from Part 8. We do not know the effort behind it — the list in the Solothurn minutes of 25.11.2025 shows that six systems hang on introducing one standard. What we can say is which statement we are missing.

Second. The admission regime for private Deklarationslösungen is today a contractual matter with a named holder and a «zuerst» (first):

«Aus vertragsrechtlichen und statuarischen Gründen seitens Swissdec können aktuell nur kantonseigene Deklarationslösungen die E-Bilanz via Distributor elektronisch empfangen. Für die Anbindung von privaten Deklarationslösungen an den Distributor müssen zuerst die vertraglichen Grundlagen geschaffen werden.» (for reasons of contract law and reasons arising from swissdec's own articles of association, only cantonal Deklarationslösungen can currently receive the E-Bilanz electronically via the Distributor. To connect private Deklarationslösungen to the Distributor, the contractual foundations must first be created)

— Technische Spezifikation eBILANZ, Anforderungen Transmitter, release 06.03.2026, note immediately below Table 2.1

A «zuerst» presupposes a «dann» (then). A published timetable for that «dann» would be information, not a commitment. The difference is not small: information can be planned around; a non-information can only be waited out.

Third. A test marker that the Richtlinien require and the schemas do not know is one to close before the first production call, not after. The Richtlinien require it in § 8.1 and § 8.3; it appears in none of the 35 shipped schemas and in none of the 11 shipped SOAP examples, and the Job types are closed sequences with no extension point. Why we hold this to be the most important of the three sentences is in Part 10: in Germany we lost six returns because a missing test marker did not count as an error but as a live case. A missing test marking has exactly one failure direction — everything is production.

Two recommendations

To software vendors. The route open today is not only XBRL out of the accounting system. We wrote that too narrowly in an earlier version and correct it here. The Regelwerk of the Schweizerische Steuerkonferenz (the Swiss Tax Conference), part of the eCH-0276 standard, names two electronic channels, in identical wording, for the Aktiven, Passiven and Erfolgsrechnung dialogues:

«Wird die E-Bilanz elektronisch übermittelt (Upload oder via Swissdec), können die Positionen nicht überschrieben werden.» (if the E-Bilanz is transmitted electronically (upload or via Swissdec), the positions cannot be overwritten)

— Regelwerk SSK zur Online-Steuerdeklarationslösung für Juristische Personen, as at 31.03.2026

Two channels, and the delivered data is treated the same either way. The upload needs no contract with swissdec. Anyone who wants to build something today that arrives builds an artefact a human uploads — and does not wait for a contractual basis that its own holder describes as still to be created.

To Treuhand firms. Three things that converge from chapters 5 and 9 and that count operationally.

With a registered general power of attorney the access codes go directly to the representative: «Beim Vorliegen einer registrierten, generellen Vollmacht wird das Schreiben an den Vertreter verschickt» (where a registered general power of attorney exists, the letter is sent to the representative) (Fragen und Antworten of the kantonales Steueramt Zürich, 08.09.2025), and «auch die künftigen Zugangscodes für die elektronische Deklaration werden diesem Vertreter zugestellt» (future access codes for the electronic declaration are also delivered to that representative) (Hilfe-Center article «Vertreter, Vollmacht, Formular» — representative, power of attorney, form). The client's access credentials are not handed over.

There is no role model and no change record visible to users: «ZHcorporateTax bietet kein Rollenmodell mit unterschiedlichen Berechtigungen an. Eine Steuererklärung kann geteilt werden, wobei alle Benutzer die gleichen Rechte haben» (ZHcorporateTax offers no role model with differing permissions. A tax return can be shared, with all users holding the same rights), and on the question of who made a change: «Diese Informationen sind in der Datenbank teilweise vorhanden, sind aber für die Benutzenden in der Anwendung nicht ersichtlich» (this information is partly present in the database, but is not visible to users in the application) (both Q&A, 08.09.2025). An internal four-eyes check cannot be modelled inside the portal itself.

And anyone delivering the E-Bilanz electronically can no longer overwrite Aktiven, Passiven and Erfolgsrechnung in the portal — under Grundsatz 6, the validation rules of the SSK Regelwerk for these three dialogues expressly apply only to manually entered fields. The machine route is the less validated one. That is not a warning against it; it is a statement of where the check has to happen, namely before the upload.

The bottom line

In Switzerland, what decides the machine-readable Jahresrechnung is not the format but three gates that someone set and someone can open: a login that knows only natural persons, a contract that its own holder describes as still to be created, and a set of instructions that appears not once in 33'000 words of cantonal documentation.

---