Trust in the Number

The scarce commodity in quantitative and AI-assisted analysis is not accuracy. It is trust — and trust is a property of the process, not the number.

Abstract

The scarce commodity in quantitative and AI-assisted analysis is not accuracy. It is trust — the ability to know, of any single number, how it was produced, whether it is reproducible, whether it was selected after the fact to look good, and whether the system will tell you when it does not trust the number itself.

We call a number produced under that standard a governed determination. This paper sets out the methodology: four disciplines — pre-registration, out-of-sample validation, adversarial falsification, and detect-and-disclose — and the machinery that makes them enforceable rather than aspirational. The credibility of the method rests on an unusual kind of evidence. We demonstrate it not with wins but with restraint: a driver we sealed as working for only two of five companies rather than claim it was universal, and a promising discovery we killed ourselves before it reached a production surface. A methodology that will publicly refuse its own best-looking result is trustworthy in a way no accuracy claim can be.

The number you cannot trust

Two failure modes bracket the current market for quantitative answers (Exhibit 1).

Neither surface answers the only question that matters for a decision: should I trust this number? Trustworthiness is not a property of a number. It is a property of the process that produced it — and that process is almost always invisible. Governed determination makes the process the product.

The principle

A determination is governed when its production is disciplined before the fact, validated out of sample, tested adversarially, and disclosed honestly — and when the machinery that produced it can reproduce the result without appeal to any human or any AI narrator.

The architecture that enforces this is the Governed Determination Architecture (GDA): a strict separation between the layer that determines a value and the layer that explains it. The determination is computed by deterministic, versioned logic over declared inputs. An AI layer may explain the result in plain language, but it can never compute, alter, or invent one. This separation is what lets a governed number be audited: the explanation is disposable, the determination is reproducible.¹

The rest of this paper is the empirical discipline the architecture exists to protect.

The four disciplines

Governed determination rests on four disciplines, applied in order and enforced by the machinery of §4 (Exhibit 2). Each is illustrated below with a case drawn from CYDENiC’s own working logs, including the unflattering ones.

Pre-registration — lock the bar before the data speaks. The most common way to manufacture a false result is to decide what counts as success after seeing the outcome. Governed determination forbids it structurally: the hypothesis, cohort, success threshold, and negative controls are written down and frozen before any result is computed. When we tested whether a steel-price index improves a steel producer’s revenue forecast, the success bar — an honest majority of the companies must improve out of sample — was locked in a pre-registration document before the estimation ran. The first run improved two of five. It would have been trivial to redefine success post hoc (“two of the four clean names,” “the pure-play subset”) to manufacture a pass. The locked bar did not permit it, and the result stood at two of five. When we later revised the method and re-ran, the bar was carried over unchanged: two runs, two architectures, same threshold, honest outcome both times.

Out-of-sample is the only currency. In-sample fit is free — any sufficiently flexible model can be made to fit history, and a good fit to the past is not evidence of anything. The only test that carries information is rolling-origin, one-step-ahead, out-of-sample error: fit on data available at a point in time, predict the next unseen period, and score the miss, repeated across every origin in the record. Our farm-equipment demand model admits a driver only if it reduces out-of-sample error against the incumbent model. On that test, wheat earned a place (it cut error) and cotton was screened out (it degraded the forecast despite fitting history well). No driver is added to improve a chart. Out-of-sample skill is the only admission ticket.

Adversarial falsification — try to kill your own result. A result is not confirmed by finding more evidence for it. It is confirmed by failing to destroy it under a battery of attacks designed to expose it as an artifact. On 2026-08-06, a driver-discovery process surfaced a candidate we would never have chosen by hand: the broad U.S. dollar, as a leading signal for farm-equipment demand. On first pass it looked excellent — it cut out-of-sample error by 10.4%, it was statistically orthogonal to the crop signals already in the model, and its economic story was clean (a stronger dollar makes U.S. grain less competitive to export, softening farm income a year before equipment purchases). A first-pass process would have shipped it.

We ran the falsification battery instead. It asked: is the gain broad, or one episode? The answer was decisive — 108% of the entire improvement came from a single 2022–2023 window (a large dollar surge followed by an agricultural downturn). Outside that episode, the driver was net negative; in the most recent sub-period it actively hurt the forecast. A placebo test — the same series with its history reversed — revealed that the noise floor at this scale was roughly ±6%, meaning several other “improvements” from the first pass were noise, not signal. The dollar was a real relationship during one shock and a mirage everywhere else. It was not shipped. The production page was left unchanged. Had the falsification stage not existed, a 2022 artifact would today be presented as a live forecast driver to the people who rely on the page.

Detect and disclose — never rubber-stamp, never silently block. A governed system’s most important output is often a refusal or a qualification. When a determination cannot be trusted, the system must say so — visibly, in a label that travels with the number — rather than present it with false confidence or hide it entirely. Our forecast reliability label refuses to call a forecast “reliable” when a structural break in the history has contaminated the benchmark it would be judged against, even when the naïve fit statistics look strong. A governed rule in our engine seals a steel-price relationship for an industry explicitly as non-universal: it applies for the companies where the mechanism earns it and honestly declines for the rest, and which companies qualify is re-computed on every determination rather than frozen into a favorable list. In both cases the honest, qualified answer is the product, and the qualification is disclosed, not buried.

The machinery that makes it enforceable

Discipline that depends on the analyst’s good intentions is not a guarantee. Governed determination is enforced by construction (Exhibit 3).

What governed determination is not

Honesty about scope is part of the discipline, so we are explicit about what this method does not claim (Exhibit 4).

The deliverable is trust in the process, from which trust in the number honestly follows.

Trust as the product

Every organization that acts on a number is exposed to the process that produced it, whether or not that process is visible. The wager of governed determination is that as quantitative and AI-driven analysis becomes ubiquitous and cheap, the differentiator will not be who can produce the most numbers or the most confident ones. It will be whose numbers you can build on — audit, reproduce, defend to a regulator, and trust precisely because the system that produced them will tell you when not to.

A methodology earns that standing the hard way: by publicly declining to claim the things it cannot honestly claim. This paper’s strongest evidence is a driver we killed and a seal we qualified. That is not a weakness of the examples. It is the entire point.

Governed determination is CYDENiC’s discipline for producing quantitative claims that can be trusted, audited, and reproduced. © 2026 CYDENiC.