Clock Time as a Fixed Point for Reasoning About Time in Historical Text

When a nineteenth-century novel says the family "sat down to dinner," a modern reader, and a modern language model, assumes evening. In 1820 the phrase usually meant midday. Many of the words that let us infer time without a clock have changed meaning over the centuries: meals, lamps and candles, "the post," "church," "the evening." Each change silently shifts the timeline a reader builds. This proposal measures how much these shifts break machine temporal reasoning on historical text, and fixes it in a specific way. Clock times written in the text ("at three o'clock") mean the same thing in 1780 and 1980. We therefore use them as a fixed reference to learn, separately for each era, what the drifting time words actually implied. We then feed those era-specific estimates into a timeline-building reasoner as soft constraints.

Put in two sentences: modern temporal reasoners assume that time-signalling words have their present-day meaning, which is wrong for old text. Wherever an old text states an explicit clock time next to such a word, the clock time records what that word meant in its own era, and this gives a self-supervised, alignment-free way to correct the reasoner.

Where the gap is

The temporal reasoning literature has made fast progress on contemporary text and knowledge bases. Several systems already beat frontier models by large margins on modern benchmarks. TempT5 trains with date masking and reinforcement learning that penalizes near-miss errors (Towards Benchmarking and Improving the Temporal Reasoning Capability of Large Language Models). Timo combines math pretraining with preference optimization (Timo). TISER has the model write an explicit timeline and check its reasoning against it (Learning to Reason Over Time). ChronoFact aligns claim and evidence timelines with soft logic (ChronoFact). All four are evaluated on Wikidata-derived or modern-English data and treat the meaning of time words as fixed. FAITH (Faithful Temporal Question Answering over Heterogeneous Sources) goes further: it discards any evidence without an explicit timestamp. On archival text, that rule throws away most of the usable signal.

The work that motivates this proposal comes from the drift side. What time is it? Temporal Analysis of Novels predicts wall-clock hour across more than 50,000 HathiTrust novels, using a local BERT classifier followed by a dynamic-programming pass that keeps time moving forward. Along the way it documents the "dinner" shift and the post-1880 rise of after-10pm scenes after electric light. It treats drift as a finding and does not model it: its classifier is one static model across eras, and it answers no ordering or duration questions.

CHRONOBERG builds a dated 1750–2000 Gutenberg corpus with a mean dating error of about 3 years. It finds roughly 16,000 words whose emotional tone shifted, and shows that sequentially pretraining models across eras either forgets old usage or becomes too rigid to learn new usage. It contains no temporal reasoning evaluation. Its drift measurements also depend on aligning embedding spaces from different eras through "anchor" words that are assumed to be stable. That assumption is never fully true, and clock time is exactly the kind of stable anchor it lacks.

The broader linguistic-drift literature acts as a constraint on design. Modern part-of-speech taggers fall from 97% to the high 70s on Early Modern English, and spelling normalization recovers only part of the loss (Part-of-Speech Tagging for Historical English). That argues against any pipeline that depends on modern parsers or semantic role labelling, which is how ChronoFact extracts events. Historical question answering with transferred modern reasoning works, but complex architectures overfit small historical sets (Finding Answers from the Word of God). That argues for keeping the learned era-specific component small.

Core Claim

Our hypothesis is that a meaningful share of temporal reasoning errors on pre-1900 text comes from time-anchor words being read with their modern meaning, not from weak reasoning ability. We further hypothesize that era-indexed anchor priors, estimated only from co-occurrence with explicit clock times, recover most of that share. Era-agnostic priors, a publication year placed in the prompt, and spelling normalization should not.

The claim fails in either of two cases: the error gap between explicitly-timed and anchor-only versions of the same passages does not grow with text age, or simply telling the model the publication year closes the gap.

Research Program

Anchor priors. For each 50-year bin $e$ of Chronoberg, plus HathiTrust novels dated by publication year, we collect every window in which a candidate anchor phrase $a$ appears within a few sentences of a parseable clock expression. The candidates are a few hundred phrases about meals, lighting, religious services, mail, travel and work. Collection uses Kim et al.'s 24-class hour setup, including their AM/PM disambiguation. From these windows we estimate a smoothed distribution

$$p_e(h \mid a), \quad h \in \{0,\dots,23\},$$

the probability that an anchor $a$ written in era $e$ refers to hour $h$. We do the same for typical durations: anchors such as "the journey to town" or "a fortnight's post" are paired with explicit elapsed-time statements, giving a log-duration distribution.

Because the reference is the clock and not an embedding space, nothing needs to be assumed stable except clock notation. This avoids the anchor-word problem that affects alignment-based drift measurement. The drift score of an anchor is the divergence between its earliest and latest era distributions. This gives an empirical split between drifted and stable anchors, which the analysis below relies on.

Injecting the priors. We build on TISER's stage that constructs a timeline. Before the model writes its timeline, the system retrieves the anchors present in the passage. It then inserts their era-specific distributions, summarized as a modal hour or duration with an interval, as short notes conditioned on the document's date. When no date is known, it uses an era estimate from a lightweight dating classifier. We then fine-tune Qwen2.5-7B on TISER-style synthetic traces that use these notes.

As a non-generative variant for the wall-clock task, the same priors replace the static emission scores in Kim et al.'s dynamic-programming pass. This tests whether the gain comes from the priors themselves and not from the large language model.

Evaluation set. No existing benchmark tests L2/L3 reasoning over drifted anchors. Here L2 means placing an event at a time and L3 means ordering events relative to each other, following TempReason's levels. We build the evaluation set as follows. We sample passages across bins from Chronoberg and HathiTrust that contain explicit clock times, and write ordering, time-of-day and interval questions about them. We then create a paired "anchor-only" version by deleting the clock expressions while keeping the anchors. Gold answers come from the original clock times. This design isolates exactly the inference a reader must make from anchors. Passages are split by book, so that no book contributes both prior-estimation windows and test items.

A smaller set of passages that never contain explicit times, around a few hundred, is annotated by humans. Kim et al.'s 1.85-hour inter-annotator disagreement on The Great Gatsby serves as the expected noise ceiling. This set checks that the paired construction does not overstate real-world gains.

Baselines

The naive reference point is a constant "noon" prediction for wall-clock hour, together with standard prompting of the same base model for the question-answering tasks. Kim et al. report 6.21 hours of whole-book mean absolute error for the noon baseline. On TempReason L3, TISER reports 40.1% for standard prompting of Qwen2.5-7B. Both numbers anchor our scale.

The strongest existing wall-clock method is Kim et al.'s BERT plus dynamic programming, which reports a 4.23-hour whole-book mean absolute error. We retrain it on our training books with the same data. A win means our prior-augmented dynamic-programming variant lowers error on pre-1880 books by more than the variance across seeds, without hurting post-1950 books.

The strongest existing reasoner is TISER fine-tuning on Qwen2.5-7B, which reports 91.5% on TempReason L3 and a 91.1% macro-average, against 78.5% for GPT-4 with standard prompting. We train it on the same synthetic-trace budget, in tokens and steps, and the same hyperparameter search as our model. The only difference is the anchor notes. A win means higher accuracy on the anchor-only split with the improvement concentrated in pre-1900 bins, and no loss on TempReason L3 itself, so that we are not trading away modern ability.

Three further comparisons separate out the cheap fixes. The first adds the publication year to TISER's prompt. The second runs VARD spelling normalization before TISER. The third uses zero-shot chain-of-thought prompting of a frontier model, given the year. We report ChronoFact-style claim verification only as a secondary check, because its reliance on semantic role labelling is expected to break on archaic syntax.

The decisive ablation keeps everything fixed but replaces the era-specific priors with priors pooled over all eras. A second, harder ablation uses priors deliberately taken from the wrong era. If pooled priors match era-specific ones, the gain comes from "having an anchor dictionary" and not from modelling drift.

Why not just…?

Why not just tell the model the publication year? This is the most serious threat. A large model may already know that Regency dinner was at midday. Prior evidence that scale alone fails is indirect. CHRONOBERG finds a 59% contradiction rate between modern classifiers and historical emotional-tone ground truth, but that concerns tone, not time. GPT-4 trails structured systems on Tiq (P@1 0.236 against 0.462) and on ChronoClaims (macro F1 61.93 against 85.49), but those are modern benchmarks. So we do not assume the answer. The year-in-prompt comparison is run at both 7B and frontier scale. If it closes the drifted-anchor gap at frontier scale, we report the narrower result: priors matter for small deployable models, or for drifted anchors that are rarely discussed. We would not claim a general fix.

Why not just normalize spelling? Eisenstein et al. show normalization helps tagging only modestly: 79.05% with normalization plus their feature-embedding method, against 77.92% without normalization. More fundamentally, "dinner" is spelled the same in both eras. Normalization cannot touch drift in meaning. The VARD baseline measures how much of any gap comes from spelling, and our drifted-versus-stable anchor split should show that normalization does not specifically help drifted anchors.

Why not just fine-tune a temporal reasoner on historical question-answer pairs? No such supervised data exists at scale. CHRONOBERG's era-sequential training also shows forgetting or rigidity, plus a 27–34% perplexity rise on later, unseen eras. Small historical sets reward simple models (Finding Answers from the Word of God). Our priors are an external, era-indexed table, so adding an era requires no retraining and causes no forgetting. We test this by holding out one bin entirely and estimating its priors after training.

Decisive Checks

The primary measurement is the drift gap: accuracy on explicitly-timed passages minus accuracy on their anchor-only twins, plotted against era. For ordering and time-of-day questions it is reported as accuracy; for wall-clock it is reported as mean absolute error in hours.

The proposal succeeds if three things hold. First, the gap for standard TISER grows for older bins. Second, era-specific priors shrink that gap specifically on high-drift anchors and leave low-drift anchors roughly unchanged. Third, both the pooled-prior and wrong-era ablations fail to reproduce the improvement.

A further check is the held-out-era test. Priors estimated for an unseen bin should transfer without retraining.

Finally, the human-annotated subset must show the same direction of effect, so that the paired construction is not overstating gains. If drift in meaning really drives errors, wrong-era priors should do worse than no priors. That is the cleanest single sign that the model is reading the word's era and not just using a lookup table.