What the record keeps — and what it drops.

We put our own historical graph under the microscope: every one of the Clockchain's 6,272 public moments classified into a formal taxonomy, the geometry of eras mapped, the dynamics between branches of history fitted, and — the part we care most about — the recording bias measured. How much of lived human action ever makes it into the historical record? What does the record systematically keep, and what does it drop? Full numbers, methods, and caveats below.

Two maps: what history records, and what humans do

The study runs on two taxonomies — two controlled vocabularies for sorting every moment in the graph by what kind of thing it is. Lens A is the taxonomy of recorded events — 79 categories across six branches (politics & law, conflict, economy, culture & religion, disaster, and — added in v1.1 — science & exploration), induced from the corpus itself and validated against a 120-moment blind holdout (120/120 after the disaster branch was added). It answers: what did the record keep? Lens B is the taxonomy of human action — 71 categories across nine branches of what people actually spend their lives doing: bonding, work, movement, making, knowing, communicating, play, striving, and the passages of the body. It answers: what is it to live? Falling in love, giving birth, shaking hands, cooking dinner — the things a chronicle almost never writes down — all have a home in Lens B.

Both embed in hyperbolic space — a Poincaré disk — because tree-shaped structures need exponential room for their children. The theoretical prediction held on real data, with controls: two hyperbolic dimensions carry the structure with less distortion than two or eight Euclidean ones (events: 0.203 vs 0.267 and 0.275; actions: 0.208 vs 0.274 and 0.247). It is the geometry, not the dimension count.

The recording bias, measured

The two lenses are connected by a small set of typed bridges — formal statements like fighting, at scale, becomes warfare; a birth is recorded, but only when it's royal; eating enters the record only when it fails, as famine. Some actions have no bridge at all: nothing a person does while commuting, conversing, or falling in love produces a recordable event. Because the bridges are formal, they compose into a single mathematical object — a recording operator — that takes any distribution of human action and returns the shadow it would cast on the record.

Here is the question that makes answerable: take an ordinary human life's allocation of action — a stated time-use prior (sleep ~20%, labor ~16%, meals ~8%, conversation ~6%, and so on) — push it through the operator, and compare the result with what the record actually contains.

Between 61% and 69% of lived action casts no shadow on the record at all — and that is an interval across three deliberately different life-priors (modern time-use, subsistence-agrarian, urban-industrial), not a single assumption. A sliver is action the record nominally covers but drops — migration, ordinary travel, falling in love. The bulk is action with no bridge into recorded history whatsoever: sleep, conversation, caregiving, commuting, grooming. The record is not a sample of life. It is a very particular filter over it.

And the filter has a direction:

The bias also grows over time. The divergence between the record and lived action's shadow (Jensen–Shannon divergence; 0 = identical, 0.693 = maximally different) rises monotonically across eras:

In this corpus, the modern record is more action-unrepresentative than the ancient one — not less. More recording capacity has not meant a more faithful record of what life is actually made of. (The trend is computed under the modern prior with uniform propensities; the direction holds across the alternatives.)

The shape of eras

Classifying every moment gives each historical period a profile — its particular mix of coronations, battles, plagues, publications — and profiles can be compared: the distance between two periods is the minimum "effort" it takes to morph one period's event-mix into the other's, where moving weight between nearby categories (battle → siege) is cheap and moving it across the taxonomy (battle → plague) is expensive. The diversity of the recorded world (Shannon entropy of each period's event-menu) bottoms out in 700–799, the most monothematic stretch in the record, with half of everything recorded being religious life — and peaks in 1900–1949, the widest event-menu history has ever run: war, politics, economy, media, and science all live at once.

Measuring the distance between consecutive periods' profiles gives a velocity of history. The three fastest recorded transformations (exact optimal transport — an earlier approximation inflated these ~40%): 1900–1949 → 1950–1999 (Wasserstein 2.76 per century — the corpus pivoting from war and politics to culture and technology), 1550–1599 → 1600–1649 (2.27), and 1700–1749 → 1750–1799 (1.83). The two most distant periods anywhere in the record: the 900s vs the 2000s, farther apart in event-space than any other pair of centuries.

What stirs what

Do branches of history excite each other? A multivariate Hawkes model on annual branch activity (1400–2026), with the baseline adjusted for corpus density so "this century just has more entries" can't masquerade as dynamics, beats the no-excitation null by +574 log-likelihood, with a mean excitation lag of about two years. The strongest fitted couplings read like a plausibility check: conflict excites conflict (wars cluster), conflict stirs politics, and the largest cross-term is economy → culture.

The fit was then validated against something it never saw: the Clockchain's hand-curated causal edges — the graph stores explicit caused-by links between specific moments, added deliberately rather than statistically. The fitted excitation matrix ranks branch pairs concordantly with that curated causality — Spearman ρ = 0.291, permutation p = 0.019 over all pairs. One honest asterisk: restricted to cross-branch pairs only, the concordance (ρ = 0.156, p = 0.085) is not yet significant — much of the signal is history exciting more of itself. Cross-branch causal texture is suggestive, not established, at 73 edges; a larger harvest is queued.

Why a simulation company measures its own record

Timepoint's simulations inherit the Clockchain — a persistent graph of historical moments. That makes the record's bias our bias: a simulation grounded in recorded history will over-see conflict and under-see the everyday economic life most decisions actually live in, unless the engine knows the filter is there. Measuring the distortion is the first step to correcting for it — and publishing the measurement, with its error bars and caveats, is what we mean when we say the honesty is the product.

The record keeps the fighting and drops the working; it keeps the coronation and drops the commute. Most of what it means to be alive has always happened off the page — which is worth knowing, whether you're reading history or simulating it.