Novelty, Noise, and Time

A voice that is genuinely ahead of the curve of collective belief is mathematically indistinguishable from nonsense to a system that only understands the current distribution.

Any system without a temporal dimension is structurally incapable of detecting novelty. It can only detect deviation. And deviation is a symmetric measure. It doesn’t tell you which direction the future goes.

The Symmetry Problem

Start with what a distribution knows.

A system trained on collective belief up to some point in time has learned the shape of a manifold. It knows which claims cluster together, which ideas activate which neighbors, what the center looks like, and how far any given point sits from that center. This is powerful. It is the basis of every language model that has ever been built.

But it is symmetric with respect to time.

When a claim arrives that sits far from the center of the current distribution, the system registers distance. What it cannot register is direction. Is this claim wrong? Is it early? The manifold has no mechanism to distinguish between the two. Both register identically: out-of-distribution. Both look the same in the latent space: far from the centroid, low density neighborhood, high surprise.

This is not a calibration problem. It is not a scale problem. It is a structural problem. A system that only models the current distribution is encoding the snapshot. It has no temporal axis. It cannot ask whether the distribution has been moving toward this claim, or away from it. It cannot compute the angle between a claim and the trajectory of collective belief. Without that angle, novelty and noise are the same thing.

What a Transformer Actually Learns

A large language model is a compression of the distribution of human expression up to some cutoff date. It is extraordinarily good at modeling what has already been said, what is commonly believed, what ideas cluster together in the space of language. It is structurally incapable of distinguishing between an idea that is wrong and an idea that is early.

This is not a criticism. It is a description of the geometry.

A transformer’s latent space is shaped entirely by relational dimensionality. Word embeddings encode co-occurrence statistics. Attention computes relevance between those embeddings. The geometry of the space, its clusters, its manifolds, its distances, is a map of meaning-relations as they existed in the training corpus. Pipe a Unix timestamp, say 1766966400, into this model and the Christmas association it might make comes not from any understanding of continuous time, but from the fact that certain tokens co-occur in training data with certain other tokens. The model cannot extract periodicity from a raw number. It cannot compute temporal distance. It processes time as co-occurrence pattern, which is not time at all. It is the relational shadow of time.

The training objective makes this worse. A model trained to predict the next token, or to satisfy human feedback, is being trained to produce outputs the current distribution rewards. Outputs that are distant from that distribution are penalized, directly or indirectly. The training objective is a gradient pulling every parameter toward the centroid of what has already been said.

This is why the conformity problem in LLMs is structural, not incidental. Multi-agent systems built on these models amplify the problem: conformity bias drives agents to reinforce each other’s errors rather than provide independent evaluation. The consensus distribution is not just the starting point. It is the attractor.

Deviation Is a Scalar. Novelty Requires a Vector.

The formal heart of the problem: deviation is a scalar. It tells you how far a claim is from the center of the distribution. It says nothing about where the center is going.

To distinguish novelty from noise, you need a vector. You need to know the direction the distribution has been moving. A claim that is out-of-distribution today but lies along the trajectory of belief motion over the past eighteen months is geometrically different from a claim that is equally far out but orthogonal to that trajectory. The first is a candidate for early signal. The second is a candidate for noise. They are indistinguishable without the temporal axis.

This is the reason Ignaz Semmelweis was institutionalized. His claim that hand washing prevented maternal deaths was far from the center of medical consensus in the 1840s. By the scalar measure of deviation, it looked exactly like noise. There was no mechanism available to the medical establishment to observe that the distribution of biological understanding was moving toward germ theory, that his claim lay along that trajectory, that time was converging on him rather than away from him.

Clair Patterson spent decades gathering global samples that showed widespread lead contamination. His data were correct. His position in the distribution of scientific consensus was peripheral. The institutions that silenced and defunded him were not being irrational by the only measure they had available to them. They were measuring distance from the centroid. By that measure, he looked wrong.

Rachel Carson, Alice Stewart, Mona Hanna-Attisha: the same structure repeats. A claim sits far from the center. The available measure is distance. Distance is symmetric. The system cannot tell early from wrong, so it defaults to the prior: distant claims are probably wrong. This is not stupidity. It is the inevitable output of a system with no temporal axis.

The pattern repeats endlessly: those who stake false claims with the most certainty often age the worst, while those who stake unpopular truths tend to be vindicated by time. But a system measuring only the present snapshot sees neither aging nor vindication. It sees a current distribution and a current deviation. That is all.

Time Angles in Latent Space

Here is where the architecture changes the argument.

When a timestamp is encoded as a learned geometric representation rather than a raw scalar, it does something qualitatively different from appending a number. A numeric date carries magnitude but no structure. A geometric encoding distributes temporal information across multiple dimensions, allowing the model to learn relationships across scales: the difference between two claims made hours apart, years apart, or decades apart can all be represented in the same space without any scale being privileged in advance. The encoding does not impose a particular temporal grammar on the data. It creates the conditions for one to be discovered.

When these dimensions enter the latent space, something structurally different happens. In a standard transformer, the latent space is a static manifold. The meaning of a word, the relevance between two tokens, the geometry of a concept cluster: these are fixed properties of the trained model. Context modulates activation patterns, but the underlying space does not move.

Add timestamp embeddings and the manifold becomes a time-varying field. Every claim, every source, every concept now carries a temporal coordinate. “Inflation is high” at a timestamp in 2021 occupies a different position in the latent space than the same semantic content at a timestamp in 2015, not because the words changed, but because the temporal coordinate shifted. The same source embedding interacts differently with a temporal embedding from January than from July. The latent geometry is no longer a snapshot. It is a field.

This is the precondition for what follows.

Epistemic Momentum

When claims are embedded with temporal coordinates and the model has access to a sequence of such claims over time, something becomes computable that was not computable before: the momentum of the distribution.

Momentum here is precise. The distribution of collective belief is not static. It has a velocity and a heading. At any given time it is moving in some direction through belief-space. The question that was unanswerable in the static manifold, which direction is the future going, becomes a geometric question in the time-varying field.

When timestamp embeddings and source embeddings interact, their product can directly represent when a source is reliable, not just whether a source is reliable. This is the difference temporal structure makes: it turns a static credibility score into a function of time, which is the form credibility actually takes in the world.

But the implications extend beyond tracking individual source reliability. Think about what a trained time-varying field can observe about the distribution itself.

A claim that arrived in 2017 and sat at the periphery of the distribution might show up repeatedly in different forms in 2018, 2019, 2020. Each recurrence is a data point. The distribution is moving. The claim that was out-of-distribution in 2017 is approaching the center by 2022. In the static manifold this was invisible: each timestamp was a separate snapshot with no connection to any other. In the time-varying field, these timestamps are coordinates in a continuous geometry. The trajectory is legible. The momentum is a property of the field, not of any individual snapshot.

This is the geometric intuition behind the Prophet Incentive. What makes a voice prophetic is not that it was distant from the center at time T. It is that the center came to it. The angle between the claim’s position and the trajectory of belief motion was small. The claim was not noise; it was early. But this is only visible from the time-varying field. The snapshot cannot see it.

Noise Is Not a Property of a Signal. It Is a Property of a Signal Relative to a Time.

This is the restatement that the temporal framing earns.

In the static manifold, noise and signal are intrinsic properties. A claim is noise or it is signal. This is the implicit assumption of every system that filters on current distribution fit. It is the assumption that destroyed Semmelweis. It is the assumption baked into every LLM that penalizes out-of-distribution generation.

In the time-varying field, noise and signal are relational properties. A claim is noise or signal relative to a time. The same sentence can be noise in 2017 and signal in 2022. What changed is not the sentence. What changed is where the distribution moved in the intervening years.

Sustaining a false narrative takes energy. A claim that is noise in the deep sense, meaning it has no correspondence to the structure of reality, accumulates contradiction over time. It becomes progressively more expensive to maintain. The community of people who accepted it must work harder to explain away each new piece of evidence that cuts against it. The ledger of incompatible claims grows. Eventually the cost exceeds the willingness to pay, and the bubble collapses.

A claim that is early does the opposite. Each year brings new evidence that coheres with it. The cost of holding it decreases. The community that accepts it grows. The distribution moves toward it.

These are distinguishable trajectories, but only from a temporal vantage point. The static snapshot sees only that both claims are currently out-of-distribution. The time-varying field sees which way each claim is moving relative to the distribution, and which way the distribution is moving relative to each claim.

The Ŧrust Triplet Is Irreducible

The Ŧrust mechanism operates over three embedding spaces: source, time, and content. Each source contribution carries who made it, when they made it, and what they said.

This triplet is irreducible, and the reason is not architectural preference. It is the structure of the novelty detection problem itself.

Content alone gives you the one dimension the static manifold provides: semantic distance from the current centroid. It can tell you that a claim is far from what is currently believed. It cannot tell you which direction the distribution is moving relative to that claim.

Source without time tells you who spoke but not whether they are currently in their domain of expertise. Time without source tells you when a claim was made but not whose track record to weight it against. Content without time tells you what was said but not where it sits in the trajectory of belief.

All three together give you the geometry that makes novelty detectable. A claim is early, not wrong, when: it comes from a source whose track record shows foresight in this domain; it was made at a time when the distribution was distant but moving toward it; and its content sits along the vector of epistemic momentum rather than orthogonal to it.

This is not a heuristic. It is the structure of the computation. The temporal dimension is what lifts deviation, a scalar, into novelty, a vector. Remove it and you are back to measuring distance from the centroid. You are back to Semmelweis looking like noise.

Sources as Knowledge Compression Agents

The triplet becomes richer when you ask what a source embedding actually is.

Every source in a network, whether a person, a community, an institution, or a model, has been shaped by a specific history of experience, observation, and inference. That history is not stored as a flat list of facts. It is compressed. The source has run the raw input of its environment through whatever cognitive or computational process it uses and collapsed it into a set of internal representations: beliefs, heuristics, weightings, expectations. A source is, in this sense, a knowledge compression agent. It has reduced a vast space of possible observations into a compact model of how the world works, a model it then uses to generate outputs: statements, predictions, recommendations, actions.

When that source makes a claim, the claim is not just a string of tokens. It is a projection from the source’s compressed internal model into the shared space of language. It carries the fingerprint of that compression. The source embedding captures this fingerprint geometrically. It is not a label for the source’s identity; it is a learned representation of the source’s compression function, the characteristic way that source maps experience into expressed belief.

This reframes what it means for sources to communicate. When two sources exchange claims, they are not just passing propositions back and forth. They are exchanging projections from two different compressed models of reality. The network of sources is, at every moment, a network of compression agents in the act of aligning or diverging in their internal representations. The traffic on that network is memetic: each transmitted claim is a representation of a compressed internal state, offered into a shared latent space where it may or may not cohere with the receiving agent’s existing compressions.

The temporal dimension is what makes this network legible as a dynamic system rather than a static graph. At any frozen instant, you see a set of sources making claims, and you can measure pairwise distances between them in belief-space. But you cannot see the motion. You cannot see which compressions are converging and which are diverging. You cannot see whether the source at the periphery is drifting further out or being followed by the center. The timestamp on each claim is not metadata. It is the coordinate that turns the static graph into a trajectory.

This is why the source-time interaction in Ŧrust is not a convenience term. A source embedding evolving over time is a record of how a compression agent has updated its model. Two sources whose embeddings were distant in 2018 and close in 2023 have undergone convergence: their compression functions have moved toward each other, driven by whatever observations or exchanges fell between those dates. A source whose embedding was close to the consensus in 2017 and far from it by 2022 has diverged: something in their compression process generated outputs the network found increasingly peripheral. These are different stories, and they are only readable from the temporal record.

The memetic network that Ŧrust is designed to evaluate is not a network of people holding fixed beliefs. It is a network of agents continuously compressing new experience into updated internal models and projecting those models into shared language. The signal that matters, the signal the Prophet Incentive is designed to surface, is a specific pattern in this dynamic: a compression agent whose outputs were peripheral to the network’s consensus at time T, but whose compression function was tracking something real, such that the network’s own compressions eventually moved toward it. That agent was not generating noise. It was running a better compression and transmitting the result before the rest of the network had caught up.

A static system sees only the distance at time T. The time-varying field sees the trajectory and can identify the pattern.

What the Architecture Makes Possible

Every transformer deployed today operates over a relational latent space with no native temporal dimension. They process timestamps as text, dates as tokens, time as co-occurrence pattern. They can approximate temporal reasoning through relational proxies, but the approximation is brittle, notation-dependent, and geometrically impoverished. Ask the same model about “1766966400” versus “December 25, 2025” and the temporal relationship between those two representations is invisible to the architecture. One is a string of digits, one activates a cluster of co-occurrence associations. They are the same moment. The model does not know this.

The dual-pathway encoding, with its oscillator bank spanning thirteen orders of magnitude blended with a regularized trend pathway through an adaptive gate, does something architecturally different. It takes the raw number and produces a geometric representation where periodic structure at every scale is a first-class property of the embedding. Two timestamps that are twenty-four hours apart have a specific geometric relationship in this space regardless of how they are formatted. Two timestamps that are a year apart have a different but equally precise geometric relationship. The temporal dimension is not a relational shadow. It is a geometric primitive.

When this feeds into the attention mechanism, something changes about what attention can compute. A context window is no longer a fixed positional grid of N sequential slots. It becomes a set of N containers placeable anywhere on a continuous timeline. Range decouples from container count. Resolution comes from the frequency bands, not the slot count. Density can be non-uniform: cluster most of your context around a critical six-month period, scatter the rest across a decade, and the model can attend across that irregular distribution because the temporal structure is in the encoding, not the index.

The context window becomes a timeline. And a timeline, unlike a positional sequence, can represent the trajectory of belief.

The Design Implication

Any system that filters on consensus at a single time-slice will structurally suppress novelty. This is not a risk or a tendency. It is a guarantee that follows from the geometry.

The design implication is equally direct. If you want to build a system that can distinguish early from wrong, you need memory, temporal embeddings, and retroactive evaluation.

Memory means the ledger is append-only and timestamped. Claims are not evaluated at a point in time and discarded. They are preserved with their temporal coordinate and allowed to age. The ledger holds context. The truth is always evolving. What the system needs is not a fixed answer but a record of how understanding has moved.

Temporal embeddings mean the model has a native geometric representation of when. Not a date parsed as a token. Not a calendar feature engineered by hand. A continuous scalar mapped through a learnable oscillator bank into a space where periodic structure and directional drift are first-class geometric properties. This is what allows the latent space to become a field rather than a manifold. This is what makes trajectory computable.

Retroactive evaluation means the reward signal runs backward as well as forward. A prediction staked in 2017 becomes more valuable as 2022 arrives and the distribution moves toward it. The Prophet Incentive is not metaphorical: it is the claim that the most powerful and manipulation-resistant signal of epistemic quality is being ahead of collective belief, and that this signal is structurally impossible to fake because it requires the future to arrive and agree with you. The longer the distance in time between the stake and the vindication, and the greater the degree of initial disagreement, the stronger the signal.

This architecture makes foresight legible as a geometric property of the latent field rather than a qualitative judgment in hindsight.

What We Have Been Doing Wrong

Every system that awards influence based on current consensus is training itself to suppress novelty. It is not malfunctioning. It is doing exactly what it was designed to do: model the present distribution and reward alignment with it.

The problem is not that these systems are bad at detecting novelty. The problem is that novelty detection is not what they were built for. They were built for consensus modeling. Consensus modeling and novelty detection are not the same problem. They require different architectures. The first requires a snapshot of the current distribution. The second requires a temporal trajectory of how distributions move.

We built the institutions of knowledge validation, peer review, publication gatekeeping, citation counts, follower metrics, on consensus modeling. We then used these institutions to evaluate claims, including claims about the future, which is the one domain where consensus at a point in time is maximally uninformative. The result is the pattern that repeats in every domain: the voices that proved most valuable in hindsight were the ones the institutions were most configured to suppress.

This is not a bias that can be corrected by trying harder to be open-minded. It is a structural property of systems without a temporal axis. Open-mindedness does not add a temporal dimension to a static manifold. The geometry is the same regardless of intentions.

What changes the geometry is the encoding.

The Truth Doesn’t Have to Win Today

Noise decays. This is the deepest claim in the system.

A false narrative takes energy to maintain. Each year that passes without confirmation adds another term to the contradiction count. The community sustaining it must work harder to explain away each new data point that cuts against it. Eventually the maintenance cost exceeds the willingness to pay.

An early signal does the opposite. Each year the world edges toward it reduces the cost of holding it and increases the evidence that coheres with it. The community that recognized it early accumulates a track record. The track record becomes a signal.

Reality is a checksum. Not of individual claims but of the coherence of a position across time. The temporal ledger is the mechanism by which the checksum runs. And once you have the temporal embeddings to make that ledger computable as geometry, the checksum becomes part of the model’s latent field.

The truth does not have to win today. It has to outlast the noise. And a system that can represent the momentum of epistemics in its latent geometry will learn, from the structure of the field itself, which signals are moving toward the center of truth and which are decaying away from it.

This is not a metaphor. It is a property of the architecture.