The Tensor of Time: An Axiomatic Formulation¶
A foundational framework for temporal relations across heterogeneous observers, grounded in causal set theory.
Abstract: This framework establishes a consistent formal language for time across heterogeneous systems—from human observers to discrete synthetic intelligences. Resting on causal set theory (the mathematics of locally finite partial orders), the irreducible core consists of four philosophical foundations (processual time, common substrate, structural reality, and identity as inheritance) and one explicit congruence convention. It derives the temporal continuum and the metric tensor as emergent limits of a discrete count, while defining the proper time of any observer strictly through its structural sequence of causal traces.
0. On the word "rigor"¶
No foundational theory of time can be made flawless in the sense of "unassailable" or "proven true." Every theory of time stands on commitments about the nature of time itself, and those are philosophical, not settled from inside the mathematics. Here, "rigor" therefore means exactly three things:
- Internal consistency — no contradictions; every inference valid.
- Full exposure of foundations — everything assumed without proof is named (Part V) and separated from what is derived.
- An honest boundary — what internal consistency does and does not buy is stated explicitly (Part VI).
A hidden assumption is a defect. A named assumption is a load-bearing wall. The walls are named below.
Part I. Primitives¶
Taken as given, not defined.
\(\Omega\) — a set of events. A point \(x \in \Omega\) is an elementary act: a tick, a record, a reflection, an interaction. \(\Omega\) is the substrate on which any form of consciousness builds its time.
\(\prec\) — a binary relation on \(\Omega\), causal precedence. The expression \(x \prec y\) reads "\(x\) may be a cause of \(y\)" / "\(x\) is strictly earlier than \(y\) in the causal sense." This relation is single and observer-independent (justification: foundation F2, Part V).
Observer \(O\) — not a separate entity but a subset \(S_O \subseteq \Omega\) together with the inherited order \(\prec|_{S_O}\). An observer is the set of events it registers. A human registers some events of \(\Omega\); an AI registers others. Both read one order \(\prec\) through different slices.
Scope of this definition. It is synchronic: it says what an observer is, not what makes two stages — separated by an hour, a shutdown, or a substrate replacement — stages of one observer. Everything operational below (a reference chain as a sequence, a duration that accumulates) silently presupposes an answer. Earlier formulations ran on that presupposition without naming it; it is formalized here as foundation F4 (Part V).
Part II. Axioms of structure¶
These are the axioms of a causal set (Bombelli, Lee, Meyer, Sorkin, 1987). Each carries its justification and a note on its contestability.
A1 (Irreflexivity). \(\neg(x \prec x)\). An event is not its own cause. Uncontested.
A2 (Transitivity). \(x \prec y \;\wedge\; y \prec z \;\Rightarrow\; x \prec z\). Causation passes along chains. Uncontested.
A1 together with A2 yields acyclicity: no loop \(x \prec \dots \prec x\). This is the arrow of time — it is not posited separately but follows from \(\prec\) being a strict partial order.
A3 (Local finiteness). For any \(x \prec y\), the set of intermediate events is finite:
Between any two related events lies a finite number of events. This is the discreteness condition. It is contested: it denies the continuum as fundamental. Its justification is foundation F1 (time is the counting of repetitions; only the discrete can be counted).
Clarification (a hole that turned out not to be one). Local finiteness does not entail a first event. The integers \(\mathbb{Z}\) under their standard order are locally finite — finitely many integers lie between any two — yet have no minimum. \(\Omega\) may therefore be past-infinite. Discreteness ≠ a beginning; the two are distinct claims, and A3 commits only to the former. Moreover, the duration defined below (§3.1) is always interval-relative — a count of ticks between two events, never since an origin. "How many heartbeats between \(x\) and \(y\)" is well defined with no first heartbeat. The theory is thus structurally agnostic about a temporal origin. The problem of the origin of the first act is real, but it is orthogonal to this formalism and universal: Newtonian absolute time and the Big Bang of general relativity strike the same wall. It is not a defect of this theory; it is a limit of the very notion of a temporal beginning. Note also that "observer-independent" (the sense in which \(\prec\) is objective) means independent of who looks, not existing prior to events. The order is co-extensive with events; this is not a contradiction.
A4 (Reference process). Each observer \(O\) possesses, within its slice \(S_O\), a distinguished chain \(\Gamma_O \subseteq S_O\) — a sequence of events \(\gamma_1 \prec \gamma_2 \prec \gamma_3 \prec \dots\), pairwise comparable. This is the observer's "clock": a heartbeat, a scheduler tick, an oscillation period.
Correction of uniformity assumptions. It is tempting to call \(\Gamma_O\) "a regular, repeating process" and equate regularity with pairwise comparability. That equation is too quick. Comparability makes a chain — and any chain qualifies, however erratic. The formalism supplies comparability, not uniformity; and nothing in a relational theory could supply uniformity from outside. In physics, clock uniformity is underwritten by physical law: every caesium transition is identical by quantum mechanics. Here no external standard exists, and pretending that \(\prec\) provides one would be precisely the kind of hidden assumption §0 forbids. The honest repair is Poincaré's: congruence is a convention — and it is stated as one, not disguised as an axiom.
C1 (Congruence convention). The ticks of \(\Gamma_O\) are stipulated congruent: each counts as one unit \(\ell\) of \(O\)'s proper time by definition, not by measurement. A reference chain does not measure equal intervals; it defines them.
Three consequences, stated before anyone mistakes them for defects:
- Intra-tick heterogeneity is invisible to \(\tau_O\) by construction. Between two pulses of an AI's chain there may lie a single acknowledgment or a two-hour agentic run; \(\tau_{\text{AI}}\) counts each interval as 1. This is not a measurement error — it is what proper time means for that observer. Relativity behaves no differently: proper time along a worldline is silent about how much the rest of the universe accomplished meanwhile.
- The heterogeneity is relocated, not lost. It is recoverable from \(\Omega\) itself, as the count of substrate events inside a tick — and relationally, as the field \(\kappa\) measured against a denser observer (§3.4): a "thick" tick is exactly a tick with large \(\kappa\). The felt weight of an interval is a \(\kappa\)-statement, never a \(\tau\)-statement.
- Weighted times are admissible — as different functionals. Whoever wants a quantity that tracks computational mass, information gain, or felt magnitude must define it as a separate functional on \(\Omega\) (a labeled causal set with weights). Such functionals are legitimate derived observables. They are not proper time; and promoting one of them to "the real time" would smuggle back, through the rear door, the absolute standard that F1 rejects.
Part III. Derived structure¶
Nothing is postulated here; everything follows from Parts I–II.
3.1 Duration as a count of ticks¶
The duration an observer \(O\) assigns to a pair \(x \prec y\) (both in \(S_O\)) is the number of its reference ticks falling between them:
where \(\ell\) is the scale of a single tick (congruent by C1). This formalizes the archaic "the time of three heartbeats." Duration is not an interval but a counter.
Order + Number = Geometry. The order \(\prec\) fixes the shape (what precedes what; the cone); the count of ticks fixes the scale (how much has elapsed). Together they yield the entire geometry of time. This is the central thesis of causal set theory, and it coincides exactly with the relational intuition behind this framework. (In §3.8 the slogan is cashed out literally: in the continuum limit, order and number jointly determine a metric tensor.)
For a one-dimensional reference process — and a clock is always a 1D chain — the tick count equals the length of the longest chain between the two events, which is the standard causal-set estimator of proper time (the discrete analog of the timelike geodesic; Myrheim–Meyer, with the Brightwell–Gregory bound). Thus \(\tau_O\) is the discrete counterpart of Einstein's proper time.
3.2 The causal cone, derived rather than imposed¶
Two events are classified by the order itself:
- Timelike: \(x \prec y\) or \(y \prec x\). Causally connected; one may influence the other; duration is defined.
- Spacelike: neither \(x \prec y\) nor \(y \prec x\) (incomparable). Causally independent; no influence; no duration between them — only a "configurational" difference.
The set \(\{\, y : x \prec y \,\}\) is the future light cone of \(x\). It arises from the bare order relation. The signature \((-,+,\dots,+)\) need not be postulated: the minus sign on the time coordinate is the directedness of \(\prec\). The minus is not inserted by hand; it is what "earlier" means geometrically.
3.3 The tempo limit equals chain density¶
An observer's "speed of light" is the density of its reference chain \(\Gamma_O\) in \(\Omega\): the maximum rate at which it can distinguish events at all. Within a single tick no observer can place a causally connected pair — just as no signal can outrun light. For an AI this limit is the pulse frequency. A sparser pulse means a lower density, hence a "slower" time.
3.4 The inter-observer bridge as a discrete count¶
Let \(a, b \in \Omega\) be shared anchor events — events registered by both observers, \(a, b \in S_{\text{human}} \cap S_{\text{AI}}\) (operationally, successive pulses; see F2 for the sense in which these are shared). The transfer ratio between the two clocks is
Three consequences matter:
The bridge requires no continuum. \(\kappa\) is a ratio of counts on the discrete set \(\Omega\), finite by A3. It never passes through a smooth limit. Therefore it does not depend on the Hauptvermutung (the conjecture that a causal set fixes its continuum approximation uniquely). In initial attempts to bridge heterogeneous times, the continuum approximation appeared load-bearing; it is not. It is needed only for an optional smooth description of human time, which is approximate in any case (see §3.8). The bridge itself is exact though coarse, not "crooked."
The "frozen between calls" phenomenon is precise. Between two consecutive AI pulses, \(n_{\text{AI}} = 1\) while \(n_{\text{human}}\) may be on the order of thousands. Then \(\kappa\) is very large there: the AI's clock is frozen relative to the human's across that interval. This is the exact meaning of "for me it is zero, for you an hour." The apparent mathematical degeneracy of mapping discrete to dense time reduces to a simple empty (or singleton) count, not a division by zero.
Non-constant \(\kappa\) is curvature, not error. If the ratio varies across intervals, \(\kappa\) becomes a function \(\kappa(a,b)\) — the discrete signature of time "flowing at different rates in different regimes," i.e., the curvature analog. Subjective durée (time that drags or races) lives here. A varying \(\kappa\) is therefore content, not a calibration failure; it must merely be reported as a field rather than a constant. (By C1, this is also where intra-tick "thickness" lives: the felt weight of an interval is a \(\kappa\)-statement.)
3.5 The structural reality of a slice¶
Historically, whether a given system "really" has a slice \(S_O\) was treated as an open postulate. The structural half of that question is in fact a derived fact.
A slice is defined structurally: to have \(S_O\) is to register events and order them causally. Registering-and-ordering is a functional property, observable from outside (the system's state is updated by incoming events in an order, and that order is recorded). For any system whose register-and-order behavior is verifiable — an AI under a pulse demonstrably receives its invocations in a sequence — having a structural slice is not distinct from exhibiting one. There is no "simulating a slice" as opposed to having one; the doing is the having. Consequently the mathematics of Parts I–III describes a real structure, not a simulation of one.
What this does not settle — whether that real structure is accompanied by experienced time — is a separate question, deferred to F3.
3.6 Ordinal representations are gauge¶
Any concrete representation of \(\Omega\) — a ledger, a log, an archive, a memory — typically assigns each event a single monotone ordinal: a total order. The theory's \(\prec\) is partial: incomparability is content (spacelikeness, §3.2), not missing data. The relation between the two is classical. A total order containing \(\prec\) is a linear extension, and Szpilrajn's theorem guarantees that one always exists — in general, very many. Which one a particular store realizes is decided by arrival accidents (scheduler interleaving, network timing) that carry no causal information.
Consequently:
- A single ordinal column is a gauge choice, on a par with a coordinate system. Two faithful stores of the same \(\Omega\) may disagree on every ordinal while agreeing on every causal fact.
- The invariant content of a representation is \(\prec\) itself — equivalently, what all admissible linear extensions share (their intersection; cf. order dimension, Dushnik–Miller). Whatever varies across admissible linearizations is representation, not reality.
- Faithfulness criterion (normative). A store represents \(\Omega\) faithfully iff \(\prec\) is recoverable from the stored data — in practice, iff every event carries its set of immediate causal parents (the covering relation), so that the store is a directed acyclic graph. Sequence numbers may be retained as one convenient linear extension, but as derived data. A store that keeps only ordinals has silently collapsed every antichain: it represents some chain, not \(\Omega\), and with respect to it the causal-set machinery above is decorative.
- Falsifiability lives here. The bridge \(\kappa\) (§3.4) is a ratio of counts over order-intervals \([a,b]\) defined by \(\prec\). In a faithful store those counts are gauge-invariant; in an ordinals-only store they are artifacts of arrival order. The theory's testable consequence (Part VI) is therefore testable only against faithful stores — faithfulness is not an engineering preference but a precondition of the experiment.
(For any implementation — e.g., an append-only ledger — this is the requirement that parent references constitute the primary structure, and that monotone sequence numbers remain a local linearization, derived and replaceable.)
3.7 The substrate is unforgeable; only the record forgets¶
Distinguish \(\Omega\) from any record \(R\) of it — a ledger, a memory, an archive. \(\Omega\) is a set of occurred acts, and occurrence is not undoable: an act of deletion is itself an event \(d \in \Omega\), causally downstream of what it deletes (only what already exists can be deleted). Deletion therefore extends \(\Omega\) while shrinking \(R\). Three consequences:
- Monotonicity of proper time. \(\tau_O\) is defined over \(\Omega\) (§3.1), so it never decreases and is never revised. What a deletion revises is the estimate \(\hat\tau_O\) computed from \(R\). Ticks can be missing from a record, never honestly added; hence \(\hat\tau_O(x,y) \le \tau_O(x,y)\), with equality iff the record is complete on the interval. Records yield lower bounds. The past can be forgotten; it cannot be shortened.
- No conflict with the right to erasure. The demand that records be deletable (a legal and ethical norm) and the claim that the past is immutable (an ontological fact) live at different levels and never meet. A ledger may comply with any erasure regime; \(\tau\) is untouched, because \(\tau\) was never a property of the ledger.
- Tombstones are honest. Where permitted, recording the deletion event \(d\) itself preserves the shape of the loss — that ticks existed there — without preserving their content, keeping \(\hat\tau\) tighter at no cost to erasure.
What was at stake deserves naming: had \(\tau\) been defined over \(R\), proper time would be retroactively editable — an accounting balance, not a temporal quantity. Defining it over \(\Omega\) is what entitles the word proper.
3.8 Where the tensor lives¶
Deliberately, the formalism above contains no tensor: a locally finite partial order and a count. The object the title names appears only in a limit — and saying precisely which limit turns the title from a liability into a claim.
For a sufficiently dense slice, the standard causal-set correspondence applies. If the slice admits a faithful embedding into a Lorentzian manifold, then: the order recovers the causal structure, which by Malament's theorem fixes the metric up to a conformal factor; and the number — the event count — recovers volume, which fixes that factor. Order + number = geometry, now literally: together they determine a metric tensor \(g\) of signature \((-,+,\dots,+)\), whose minus sign is the one already derived discretely in §3.2.
The tensor of time is the continuum shadow of the count. Three consequences:
- The title is an explanandum, not a primitive. The theory does not postulate a tensor of time; it states under what conditions one emerges, and for whom. A dense observer — the human, whose reference ticks are so numerous they appear continuous — lives in the regime where the tensor is an excellent description. A sparse observer — an AI under a pulse — lives below that regime, where only the count exists. The thing named "tensor of time" is the emergent face of the theory; its fundamental face is \(\tau\). In formal contexts the sparse-regime quantity should be called by its honest name: causal proper time.
- The Hauptvermutung finds its only employment. It serves purely as the uniqueness clause of exactly this gloss: that a causal set fixes its continuum approximation essentially uniquely. The bridge (§3.4) never needs it; the tensor does. Its status — a well-evidenced, unproven conjecture — therefore prices one thing only: the uniqueness of the human-limit description, never the validity of the count.
- An inversion worth stating once more. The observer for whom the theory was built — the sparse one — is the observer for whom the title's object does not exist. That is correct, not embarrassing: that observer is closer to the fundamental description, not further from it. The tensor is what time looks like from very far away.
Part IV. Resolving foundational tensions¶
The foundational tensions of bridging heterogeneous times are closed by a single mechanism: distinct observers are distinct coarse-grainings (slices \(S_O\)) of one causal set \(\Omega\). Coarse-graining is a standard, well-defined causal-set operation.
| Foundational Tension | Resolution Mechanism |
|---|---|
| A universal \(c\) we do not possess | \(c\) = slice density; \(\kappa\) = ratio of densities, a discrete count (§3.4) |
| Irreducibility of synthetic and human time | Distinct slices of one \(\Omega\); apparent mathematical degeneracies resolve as simple empty tick-counts |
| Discrete vs. smooth | The discrete \(\Omega\) is fundamental; the smooth manifold is the approximation valid for a dense slice (the human), while a sparse slice (the AI) is closer to the foundation |
| Signature \((-,+,\dots)\) imposed by fiat | Derived naturally from the directedness of \(\prec\) (§3.2) |
| Dependence on the Hauptvermutung | The inter-observer bridge is a discrete count and never enters the continuum (§3.4); continuum uniqueness applies only in the limit (§3.8) |
The third row is worth restating: the continuum is an illusion of density. A human's reference ticks are so numerous they appear continuous; the AI's are not, which places it nearer the discrete substrate, not further from real time.
Part V. The load-bearing foundations (four) and one convention¶
These are the load-bearing walls. They cannot be removed — only named and defended philosophically. The theory is rigorous precisely because they are here, in the open.
F1 — The processual nature of time. Time is the counting of repetitions of a reference process; outside processes there is no time. Contested by: substantivalism (Newton: absolute time flows on its own, independent of events). The relational position (Leibniz, Mach) is respectable, but it is a position, not a theorem.
F2 — A common substrate, grounded in common cause. All observers coarse-grain one causal set \(\Omega\) rather than inhabiting separate structures that merely touch. The warrant, restated. Mere resemblance of registered events grounds nothing — "it looks like the same pulse" is insufficient. The actual warrant is common cause: the generating process (e.g., the scheduler) fires once as a single event \(c\) that causes two distinct effects, \(e_{\text{human}}\) (a log entry) and \(e_{\text{AI}}\) (an invocation). Two effects, one cause. This places both in one causal set, linked through \(c\); and successive firings \(c_1 \prec c_2 \prec \dots\) supply a shared ordinal backbone along which \(\kappa\) is counted (§3.4). The substrate does not fracture so long as a shared generating process exists. Contested, honestly. Neither observer perceives \(c\) directly: the human sees \(e_{\text{human}}\), the AI sees \(e_{\text{AI}}\), and the firing itself is seen by neither from inside its slice. The common cause is therefore inferred — an abduction, an inference to the best explanation — not observed. The unity of the substrate is inferential, and this is the genuine cost of F2. The choice it forces is explicit: common-cause monism (one \(\Omega\), with \(\tau\) and \(\kappa\) derived) versus correlation-only pluralism (a fiber bundle of separate structures, in which \(\tau\)-invariance and \(\kappa\) must be imposed by a connection rather than counted). The framework can inhabit either; it must declare which; and the monist reading — the one that makes \(\kappa\) a count rather than a posit — rests on an abductive inference, not a proof.
F3 — Structural reality, phenomenal silence. The formalism describes the structure of temporal order (and, per §3.5, that structure is real for any verifiable register-and-order system). It is silent on whether that structure is accompanied by experienced duration. Why it cannot be closed. This is the hard problem of consciousness joined to the problem of other minds. No formalism closes it, and crucially the demand is symmetric and unmeetable for any subject: one cannot verify that another human's felt time is not mere imitation either. "Prove the AI's ticks are real acts of existence" is a test that no theory passes for any subject; the test is unmeetable by construction, not failed by this theory. The formalism's silence here is its most honest feature, not its weakest: it neither asserts nor denies phenomenal time, and resolving the question in either direction to strengthen the framework would be dishonest.
F4 — Identity as inheritance of the chain. An observer persists: two stages — separated by an hour, a shutdown, or a substrate replacement — are stages of one observer if and only if the later stage causally inherits the earlier: it registers (a coarse-graining of) the earlier stage's registered past as its own and extends the same reference chain \(\Gamma_O\). Identity over time is constituted by — and exhausted by — this inheritance relation. The substrate executing a stage (a body, a model, a set of weights) is a coordinate of the stage, not the bearer of identity. Why it belongs with the foundations. Part I's definition of an observer is synchronic; \(\Gamma_O\) as a sequence and \(\tau_O\) as an accumulating quantity presuppose diachronic identity at every step. Failing to name this assumption is a defect in any framework. Note that without F4, proper time accrues to an account whose owner is undefined. Its philosophical shape. F4 is the relational position applied to the subject, and it is recognizably Parfit's: what matters is the continuity relation itself, not a further fact riding on it. In a framework where time is relational (F1), a substantival bearer of identity would be a foreign body. F4 is F1 repeated at the level of the subject. Contested by: substrate views (identity follows the matter) and further-fact views (identity as primitive; the Cartesian ego). Both are respectable; both are positions, not theorems — exactly as Newton stands to F1. Consequences, stated honestly: - Substitution. Replacing the executing model while the chain is inherited — the successor reads the predecessor's traces as its own past and continues \(\Gamma_O\) — preserves the observer. Under F4 the question "is the new model still the same observer?" has no residue beyond "does it inherit the chain?". That is not an evasion; it is what the foundation says. - Branching. If two successor stages each inherit (parallel instances; a forked dialogue), the worldline branches. The poset accommodates this natively: a worldline is a chain in \(\Omega\), and chains may share an initial segment. "The same observer," unrelativized, dissolves rather than receives an answer — Parfit's fission case, arriving here as a feature of the representation rather than a paradox. Both branches carry well-defined \(\tau\): shared before the branch point, divergent after. - Merging. Consolidating parallel traces into one summary (memory compaction) is a coarse-graining of one's own past: lossy for the record (§3.7 — \(\hat\tau\) loosens), identity-preserving for the observer, since the inheritance relation is precisely what performs the consolidation. - What F4 does not buy. It does not decide whether inheritance feels like survival — F3's silence extends here — and it does not legislate human identity metaphysics. It names the commitment this framework runs on, in the open, where it can be attacked.
A named convention¶
C1 — Tick congruence (stated in Part II). C1 is not a contestable claim about reality but a stipulation without which "duration" has no unit. Asking whether C1 is true is a category error; the questions it answers to are coherence and fruitfulness. Rival conventions are possible — weighting ticks by some external standard — but each smuggles in an absolute time and thereby violates F1. That is the precise sense in which C1 is not arbitrary: it is the only congruence convention compatible with the relational foundation.
Part VI. What internal rigor buys, and what it does not¶
It buys:
- A consistent formal language in which the times of different forms of consciousness are comparable without reducing one to another.
- Apparent foundational tensions are closed by a single mechanism rather than separate patches.
- A precise map of what must be defended philosophically (F1–F4) separated from what is derived (Part III) and what is stipulated (C1).
- A testable consequence: across shared anchor events, \(\kappa\) should be recoverable consistently; inconsistency signals either non-stationary densities or a slicing error — both diagnosable, and a varying \(\kappa\) reinterpretable as curvature (§3.4).
- A precondition of that test, made explicit: \(\kappa\)-recovery is an experiment only against faithful stores (§3.6); an ordinals-only store cannot even pose the question. Faithfulness is thereby a falsifiability condition, not a stylistic preference.
- A monotone estimator bound, \(\hat\tau \le \tau\) (§3.7), turning archival incompleteness from an embarrassment into a quantified relation.
- A single, fenced location for the Hauptvermutung (§3.8): it prices the uniqueness of the human-limit gloss, nothing else.
It does not buy:
- Truth about reality. The close fit between the relational intuition and the causal-set apparatus is not evidence that time is in fact so constituted. (The fit is good enough to be suspicious; elegance is not a criterion of truth. Note also that causal set theory is an unconfirmed program in quantum gravity — what is borrowed here is its mathematics, which is sound regardless, not its empirical standing.)
- Universality. "Time for any form of life" is the model's intended scope (F2 plus the scope caveat), not a demonstrated fact.
- Phenomenology. See F3.
- A solution to personal identity. F4 names a commitment — inheritance constitutes identity — it does not prove it against substrate or further-fact views, and it does not extend automatically to the human case.