Atoms of Truth
What the irreducible units of truth are — and why the most important ones go unrecorded
A radiologist examines a chest X-ray and concludes it shows early-stage pneumonia. She documents her finding.
Consider everything that happened in that sentence — carefully. There are four distinct things, and they are not all the same kind of thing.
First: the X-ray was ordered and taken. A physician decided this patient needed imaging. A radiographer positioned the patient, selected the exposure settings, and pressed the button. These are committed decisions made by named authorities at specific moments. They are irrevocably true regardless of what the image contains or what any radiologist concludes from it. The fact that an X-ray was taken, of this patient, at this moment, by this radiographer, under this order — a named authority acted. The act is permanent.
Second: the X-ray image itself. This is not unmediated reality. It is a transduction: tissue absorbed radiation in a pattern determined by density, a sensor captured that pattern, firmware converted it to pixel values, a compression algorithm encoded them into a file. The image is a representation of the patient’s lungs, passed through four transformation steps before it reached the radiologist’s screen. It is high-fidelity and enormously useful — but it is a projection of the underlying physical truth, not the truth itself. The system treats it as its starting point by convention: we designate this image as the observation we will reason from. We could always ask what produced it. We choose to stop here.
Third: the radiologist’s perception and reasoning. Her visual system filters what she attends to. Her pattern recognition — shaped by years of training and the cases seen that morning — selects what seems significant. Her tentative conclusions emerge from a probabilistic process she cannot fully introspect. This internal processing is not ground truth of any kind. It is inference under uncertainty — the inference-as-fact failure mode if it is ever stored as truth. It is real, it is happening, but it leaves no trace in the record and produces no commitment.
Fourth: the radiologist’s documented conclusion. At 09:47 on a specific morning, this specific radiologist made a specific call and recorded it. That documented conclusion is categorically different from everything that preceded it — a committed decision. The act of deciding and the fact of the decision are the same thing. The regress stops here — by nature, not by convention. The reasoning that led to it can be examined, challenged, even shown to be wrong — but the fact that this conclusion was reached and recorded at this moment is permanent.
Six months later, a second radiologist reviews the same image and disagrees. The first radiologist was wrong.
Does that change the first radiologist’s decision? No. The first decision still happened, at that moment, with that conclusion. The second decision is a new fact: a new authority, a new moment, a different conclusion. Both exist in the record. The history is richer, not corrected.
These four things — the act of taking the X-ray, the image produced, the radiologist’s reasoning, and the recorded conclusion — are what the rest of this part is built on distinguishing. Two of them are atoms of truth. Two of them are not. Telling them apart is the whole task.
Three categories, not two
Truth Doesn’t Change introduced a binary: facts versus derivations. That binary is useful but incomplete. Before we can say what a derivation is derived from, we need to be precise about the two kinds of things that are not derivations — and about a third kind of thing that is neither.
Decisions are the hardest kind of ground truth a system can capture. A decision is the moment an authority — human, algorithmic, or institutional — converts its reasoning into a commitment. The decision is not a representation of some more fundamental decision happening elsewhere. There is no more fundamental version. The act of committing and the fact of the commitment are identical. The regress stops here, by nature. The ordering physician decided to image the patient. The radiographer pressed the button. The radiologist documented her conclusion. Each of these is a committed decision made by a named authority at a specific moment — permanent and irrevocable, regardless of whether it was the right decision.
Observations are recordings of the world produced by measurement chains — sensor readings, captured signals, instrument outputs, log entries. They happened. They are real. But they are transductions: the thing recorded is not the phenomenon itself, it is a representation of the phenomenon produced by an apparatus. A temperature reading is not the temperature. A pixel value is not the tissue. The X-ray image is a projection of the patient’s lungs through radiation and sensors and firmware, not the lungs themselves. Every observation invites the question: but what produced this? The regress continues — we choose to stop it at some designated input boundary, by convention, not because the regress has naturally terminated.
Observations belong in the record, but they carry honest metadata about their fidelity: what produced them, what the measurement chain is, what the known error characteristics are. A well-calibrated primary standard measurement and an uncalibrated sensor reading are both observations — but they are not equally trustworthy, and a system that treats them identically will eventually fail in ways that are confusing to diagnose.
Probabilistic inferences are a third kind, different from both. The radiologist’s visual processing, the fraud model’s scoring, the demand forecast — these are computations whose outputs are uncertain even if their inputs are perfect. Information is genuinely lost. The same inputs on a different day might produce a different output. There is no stored trace of the reasoning. A probabilistic inference is not a commitment and not a measurement — but, as the next section shows, it splits in two: the fact that it happened can be recorded as solidly as anything else, while its content stays uncertain until something further is done with it.
Derivations sit outside this triad — they are deterministic computations over decisions and observations. Aggregates, sums, joins, materialisations. They are rebuildable from their inputs, always downstream, never primary.
Of these, only decisions and observations yield atoms of truth about the world — irreducible, primary, the substrate everything else is built on. The atoms are the two the world hands us directly: what an authority committed, and what an instrument measured. Probabilistic inferences are not the truth about the world — but, as the next section shows, the fact that an inference occurred is itself a recordable atom. Derivations are no longer atomic: they sit downstream of everything here.
The act and the content
There is a distinction hiding inside all of this, and once it is visible the awkward cases stop being awkward. Every recorded atom has two coordinates, and they are independent:
- The act — that this happened: who, when, with what force. This is ground truth. It is not up for debate.
- The content — the claim the act carries: certain, likely, or barely a guess. Its strength is a separate question, settled on the layers above.
A decision shows both: that the radiologist committed at 09:47 is the act (non-debatable); whether the diagnosis is correct is the content (defeasible). A measurement shows both: that the thermometer reported 38.4°C is the act; whether the patient truly runs that temperature is the content.
Even this is subtler than it looks: what a system stores is only an agent’s expression of a commitment — a signature or keystroke that can be coerced, mistyped, or forged. How a thing is captured decides how close to the truth the record reaches; the next articles turn on exactly this.
This is exactly what an inference needs. The content of an inference is uncertain. But the act — that this agent produced this output at this moment — is as solid as any other event. We can record it without pretending its content is true.
When we do, we are not deciding to act on the inference. We are making a weaker, honest commitment: a conjecture — I put this forward; I stand behind having inferred it; I have not committed to it being true, or to acting on it. This is the author who writes a thought down: a commitment to the idea, not to the deed. It is the LLM harness that judges its output “good enough” and sends the response: a commitment to having produced and transmitted this text, not a claim that the text is true.
So there are two commitment boundaries, not one:
- The assertion boundary — committing the inference as a conjecture: the act is recorded at full strength, the content enters the record at a low tier.
- The action boundary — committing to act on the content: a separate decision, usually by a different authority, often later.
A conjecture that never crosses the action boundary is not a loose end. It is precisely the right way to hold an uncertain thought: visible, attributed, dated, and clearly marked as not-yet-acted-upon. This is also the honest fix for the inference-as-fact failure mode — the error is never that an inference was recorded; it is recording its content as substrate instead of as a conjecture at its true tier.
How a system classifies the acts that enter its permanent log — and how that classification is derived — is taken up later in this part, once we have named the Real beneath the record and the substrate the record can hold.
Computations are ephemeral. Decisions are permanent.
A computation runs, produces a result, and is discarded. Run it again with the same inputs and you get the same result. It is reproducible, deterministic, stateless. It has no history because it leaves no trace.
A decision is different. A decision is the moment an authority converts a computation’s output into a commitment. It happens once, at a specific moment, by a specific entity. It cannot be un-happened. It can be superseded by a new decision, but the original decision remains in the record as a permanent fact about what that authority concluded at that time.
The computation that led to a decision is interesting — it is the reasoning behind the decision — but it is not the decision itself. The decision is the commitment. The computation is the evidence.
In event-driven systems, this distinction determines what belongs in the permanent record — but “the computation does not belong there” is too crude. The run of a computation — the transient, in-memory process — is reproducible and need not be stored step by step. What must be stored is the decision together with precise references to everything it stood on: the specific inputs it consumed (by identity and version), the method, model, or rule version that processed them, the assumptions taken as given, and any inference acts it relied on. The computation is reproducible; that provenance is not, and it is captured at the moment of the decision.
This is not bookkeeping — it is what tells us when a decision is still valid. A decision is a commitment made against a particular set of inputs and methods. When an input it referenced is later corrected, or a method it used is superseded, we can find every decision that rested on that input and ask whether it still holds — and, crucially, what change in inputs requires a new decision. A conclusion recorded without those references is an orphan: visibly concluded, but impossible to re-evaluate when the world beneath it moves. So what enters the record is the conjecture (the act of inferring, its content at an honest tier) or, once an authority acts on it, the decision — and in either case, bound to the inputs, methods, and assumptions it rests on.
The loyalty tier example
A customer accumulates purchases. Each purchase is an event, and each one triggers the loyalty rule engine: it re-evaluates the customer’s total spend over the trailing twelve months against the VIP threshold. On most purchases nothing changes. Then one purchase tips the trailing total past the threshold, and on that event the engine determines the customer now qualifies for VIP status.
What here is ground truth?
The purchase events are atomic decisions — choices made by the customer, captured at the moment they occurred. The spend computation that runs on each purchase is ephemeral: a calculation producing a number, leaving no permanent trace. The threshold comparison is also ephemeral — and on most purchases it produces nothing worth recording. But on the purchase that crosses the threshold, something happens: the rule engine concludes that this customer qualifies for VIP status, and acts on that conclusion.
That conclusion is a decision. It was made by an authority — the rule engine, operating under policies set by humans — at a specific moment, based on specific evidence. The fact that this decision was made is as irreducible as any other fact in the system. It joins the permanent record alongside the purchase decisions that preceded it.
This event is not a derivation materialised for convenience. It is the record of an atomic decision — what was concluded, by what authority, at what moment, based on what evidence.
When the loyalty rules change — new thresholds, new tiers — old upgrade decisions remain ground truth. The customer was upgraded to VIP on that date under those rules. That is what happened. New rules produce new decisions going forward. The past is not invalidated.
The most important events are the ones nobody thought to record
Every system has obvious events: OrderPlaced, PaymentReceived, UserRegistered. These get recorded because they are the explicit business transactions the system was built to handle.
But systems are full of decisions that nobody thought to record, because nobody thought of them as decisions at all.
A fraud detection model runs and scores a transaction as low-risk. No event. The score is used and discarded.
A recommendation algorithm selects which products to show a customer. No event. The selection happens and is forgotten.
A pricing engine applies a surge multiplier to a ride. No event. The price is displayed and the reasoning is gone.
A content moderation system decides not to flag a post. No event. The non-decision is invisible.
These are decisions made by authorities — algorithms, models, rule engines — with real consequences for real people. When something goes wrong — a transaction that should have been flagged, a recommendation that caused harm, a pricing decision that turns out to be discriminatory, content that should have been moderated — the question is always: what was decided, by what system, on what basis?
If the decision was never recorded, the question cannot be answered.
The accountability surface
There is a deeper reason why decisions must be captured, beyond operational convenience.
When a decision causes harm — to a customer, to a business, to a third party — accountability requires being able to answer: who decided this? On what basis? With what information available at the time?
If the decision was not recorded, accountability cannot be assigned. Not because nobody is responsible — someone always is — but because the decision cannot be traced to its source. The harm happened but its origin is invisible.
This is not merely a legal or regulatory concern, though it is those things too. It is an engineering concern. Systems that cannot account for their own decisions are systems that cannot learn from their mistakes. Every unrecorded decision is a lost opportunity to understand what the system actually did and why.
The principle is simple: if something in your system made a decision that affected anything outside itself, that decision should be in the permanent record — with the authority that made it, the evidence it acted on, and the moment it happened.
Not as a log line. Not as a metric. As a first-class, named, immutable fact.
The test for atomicity
How do you know whether something is a decision worth capturing as a permanent fact?
Ask two questions.
Could this have happened differently? If a different outcome was genuinely possible — the fraud model could have scored high instead of low, the loyalty engine could have not upgraded, the deployment pipeline could have aborted — then a decision was made. Record it.
Did this affect anything outside the component that made it? If the outcome influenced the state of the world — a customer saw a recommendation, a price was applied, a user was blocked — then the decision had consequences. Record it.
If both answers are yes, the decision is an atom. It belongs in the record, with its authority, its evidence, and its timestamp.
What this prepares
This article has established four categories — decisions, observations, probabilistic inferences, and derivations — and singled out decisions as the hardest, purest ground truth a system can hold.
Two of these four sit beneath everything else, as the substrate the rest is built on: decisions, because an authority committed; observations, because the world was measured. The next articles name what lies beneath that substrate (The Real), what the log can actually hold (Captured Evidence), and how agents think before they commit (Thinking by Writing).