Forecasts Are Not Facts: Temporal Epistemic Governance for Automated Claims and Decisions

Gergely Vámossy — Independent researcher (gergo@qiera.io). Preprint. Markdown rendition; the typeset PDF and LaTeX source are the canonical form.

---

Abstract

Automated systems, and language models in particular, routinely confuse a claim's tense with its verifiability. They state forecasts as facts — presenting a prediction or a guess about what has not happened as established — and assert unrecorded pasts as known — confabulating about events for which no record exists. These are two faces of one error: acting as though the truth is reachable when the temporal situation makes it, in principle or in practice, unreachable. We present a small, deterministic, self-testing temporal epistemic governance layer that indexes each claim by tense and enforces the verifiability that tense actually permits: the past is verifiable only where a record survives; the present is an acting boundary that cannot be re-checked; the future is forecast-only and cannot be verified, only assigned a probability. Two refusals carry the weight — a future asserted as certain is refused, and an unrecorded past asserted as known is held as testimony, not fact — and the layer composes with a governed decision procedure so that a decision fires now on a footing it can verify (the recorded past) and a belief that is explicitly a probability (the forecast), never on a pretended fact. The contribution is operability, not new epistemology: the underlying asymmetry is old, but making it a fail-closed property of an automated pipeline is not. Every component ships with a self-test.

1. The problem: tense is not verifiability

A claim's grammatical tense and its epistemic status are independent, and conflating them is a specific, costly error. The past is, in principle, settled — but knowing it requires a surviving record; a past event with no record is not false, but it is not known, and treating it as established is unfounded. The future is not settled at all: it has not happened, so no procedure can verify a claim about it — the most one can honestly offer is a forecast with a probability. The present is peculiar: it is the only moment at which one can act or directly observe, and what is not captured there passes into the (possibly unrecorded) past.

Two errors follow, and both are endemic in current automated systems. Forecast-as-fact: a prediction, extrapolation, or guess about the future is asserted with the grammar of an established fact — the overconfidence failure, and a recognizable mode of model "hallucination" about what will happen. Unrecorded-past-as-known: an event for which the system holds no record is asserted as something that definitely occurred — confabulation about what did happen. Both are the same underlying mistake seen through different tenses: behaving as if an independent ground truth is available when the temporal situation makes it unreachable. Exhortation ("be calibrated," "don't hallucinate") does not fix a structural error; a structural error needs a structural check.

2. Scope and stance

We govern the epistemic boundary of time, not its metaphysics. This work says nothing about whether the future "exists," whether time flows, or where an objective present lies; it takes no side on presentism versus the block universe. Those questions are real and unsettled, and a deterministic governance tool has no business adjudicating them. What it operates on is the one temporal boundary with uncontested, usable content: the asymmetry between what is recorded and what is unwritten, which is a fact about available information rather than about time itself. Nor do we claim new epistemology — the asymmetry of the past and future, and the status of forecasts as probabilities, are old. The contribution is operability: turning a widely-shared epistemic commitment into an enforceable, fail-closed property of an automated claim-and-decision pipeline.

3. The governance layer

Each claim is indexed by tense and assigned the verifiability that tense permits. The mapping is deterministic and fail-closed: an unmet condition lowers the verdict rather than passing silently.

Tense	Condition	Verdict and meaning
Past	a surviving, independent record exists	VERIFIABLE — check it against the record
Past	no record	UNRECORDED — not verifiable now, and not true-by-default; treat as testimony
Present	—	ACT_BOUNDARY — observe or decide now; it cannot be re-checked later
Future	a forecast (optional probability)	FORECAST — UNVERIFIED until it arrives; usable as a probability, never assertable as fact
Future	asserts certainty / "verified"	UNVERIFIABLE_NOW — refused; a category error, downgraded to a forecast

Two refusals do the real work. A claim that the future is verified or certain is refused outright: the future has not happened, so it cannot be a fact, only a probability. A claim that the past is known without a record is held at UNRECORDED, and any field whose name asserts verification (recollection_verified, future_verified) is flagged by the same overclaim linter the wider toolkit uses. Neither refusal decides whether the claim is true; each decides only what epistemic status it is entitled to assert.

4. Reference implementation, and the decision seam

The layer ships as a deterministic, self-testing component. Its value shows most when it composes: a decision is always made now, and the two other tenses supply its inputs. The recorded past sets the evidentiary footing — a VERIFIABLE track record is a mature basis to act on, an UNRECORDED past is not. The future forecast supplies the belief — its probability becomes the decision's posterior, and because a future asserted as certain is refused upstream, the decision is necessarily driven by a probability, never by a pretended fact. The present ACT_BOUNDARY is when the decision fires. Composed with a governed decision procedure, the four combinations resolve as one would want: a recorded past with a strong forecast and a human authority acts; the same footing with a weak forecast waits; an unrecorded past withholds the decision for want of verifiable footing, whatever the forecast says; and a future asserted as certain is refused at the seam — no decision fires on a pretended fact. The two tense-errors thus become, at the point of action, a withhold and a refusal-to-fire.

5. Why it matters

Tense-confusion is a distinct and addressable failure mode for AI systems that both generate claims and support decisions. A system that states forecasts with the grammar of fact manufactures unwarranted confidence; one that asserts unrecorded pasts as known manufactures history. Both are corrosive precisely where these systems are most used — planning, risk, forecasting, incident review — and both are invisible to a reader who trusts the surface grammar. Making tense-to-verifiability explicit and fail-closed converts a soft injunction into a structural guarantee: a forecast cannot present as a fact, and an unrecorded past cannot present as known, without a deliberate override rather than by default. Because the layer composes into the decision, the guarantee reaches the action, not just the phrasing.

6. Related work

The temporal asymmetry and the logic of tense are long studied — Prior's tense logic and its descendants, and temporal logic in computer science (Prior, 1967; Pnueli, 1977). The status of a forecast as a probability to be scored, not a fact to be asserted, is the calibration and forecasting tradition (Brier, 1950; Gneiting & Raftery, 2007; Tetlock & Gardner, 2015). The requirement that a claim about the past be backed by a surviving, independent record is the province of provenance and data lineage (W3C PROV, 2013). And the two errors the layer targets are, in current systems, the tense-specific faces of overconfidence and hallucination (Ji et al., 2023). Against these, this paper contributes neither new temporal logic nor new calibration theory, but an applied, fail-closed governance layer that binds tense to verifiability and composes it into a governed decision.

7. Limitations and honest positioning

This is the most applied, and the most modest, of a set of "honest limits" governance results, and it should be read that way. Its core observations — a forecast is not a fact, an unrecorded past is not known — are intuitive; the contribution is not the insight but its structural enforcement in an automated pipeline. It governs verifiability, not truth: a VERIFIABLE past claim can still be false against its record, and a FORECAST can be badly calibrated. Whether a record exists, and what probability a forecast carries, are declared inputs the tool enforces the consequences of, not facts it establishes. It is deliberately not metaphysics: it operates only on the recorded/unwritten information asymmetry, the part that is not contested. And it shares the family's non-safety-critical scope. Its worth is that a common, costly class of error becomes something a system has to commit on purpose, rather than something it commits by grammar.

References

1. A. N. Prior, Past, Present and Future, Oxford University Press, 1967 (founding tense logic).
2. A. Pnueli, "The Temporal Logic of Programs," FOCS, 1977.
3. G. W. Brier, "Verification of Forecasts Expressed in Terms of Probability," Monthly Weather Review, 1950.
4. T. Gneiting and A. E. Raftery, "Strictly Proper Scoring Rules, Prediction, and Estimation," JASA, 2007.
5. P. E. Tetlock and D. Gardner, Superforecasting: The Art and Science of Prediction, 2015.
6. L. Moreau et al., "PROV-DM: The PROV Data Model," W3C Recommendation, 2013.
7. Z. Ji et al., "Survey of Hallucination in Natural Language Generation," ACM Computing Surveys, 2023.

Gergely Vámossy (2026). Forecasts Are Not Facts: Temporal Epistemic Governance for Automated Claims and Decisions. https://vamossy.com/research/forecasts-are-not-facts