Model Is Not Reality

Sync recommends. A named human decides. Execution stays with authorized systems. On the Decision Case and in Recovery, a digital twin, a KPI model, a risk matrix, an RUL curve, a Monte Carlo run, or an LLM summary is a compressed representation. It is not the plant. It is useful only when assumptions, coverage gaps, and refusal conditions are on the record. It is not a diagnosis. It is not root cause. It is not authorization for corrective work. It is not a verified operational outcome. It is not a Decision Case. Treating the model as reality is false clearance. Verification stays open until a verified operational outcome is recorded from the plant, not from the model.

In this essay, model names those six records and no figure of speech. A digital twin is a computed state held in software and synchronized, simulated, or seeded to look like the asset. A KPI model is a formula or score that stands in for an operational result. A risk matrix is a grid of likelihood and consequence that paints a cell. An RUL curve is a remaining-useful-life line projected from a degradation model. A Monte Carlo run is a spread of simulated outcomes from assumed inputs. An LLM summary is a compressed retelling of records the model was given. Reality names the plant: the asset as it is running, the evidence a later reader can inspect, and the outcome that was actually verified. Those are different records. Treating the model as reality stores the compressed representation as if it were the plant.

The meeting puts the digital twin on the screen. The KPI model is green. The risk matrix cell is low. The RUL curve is long. The Monte Carlo band is tight. The LLM summary reads as if the case were already written. Someone says the asset is known, and the room treats the output as clearance. Nobody wrote the assumptions. Nobody named the coverage gaps. Nobody stated the refusal conditions. Nobody attached approved evidence from the plant. Nobody wrote accept, reject, escalate, or return. The Decision Case still has to show the question, the evidence, the recommendation, the named human decision, and the check. A model output is not that record, and it is not the plant.

This essay follows Map Is Not Terrain and stands beside Proxy Is Not Outcome, Assumption Is Not Evidence, Confidence Is Not Evidence, Recommend Is Not Authorize, and Dashboard Is Not Decision. Map keeps a diagram, a CMMS schema, a dashboard, a digital twin, or a process map from being stored as the operating reality. This essay keeps the computed claim — the twin as a model, the KPI model, the risk matrix, the RUL curve, the Monte Carlo run, the LLM summary — from being stored as the plant when assumptions, coverage gaps, and refusal conditions are not explicit. Proxy keeps a KPI, a leading indicator, a model score, a green tile, a closed work-order count, an MTBF estimate, or a sensor health % from being stored as the verified operational outcome. A KPI model is that stand-in when the score is treated as the result. Assumption keeps an unverified model output from being stored as a recorded fact. Confidence keeps a high model score from being stored as a recorded check. Recommend keeps a drafted next action from being stored as authorization. Dashboard keeps the display from being stored as a decision. A precise picture of a model is still not the plant.

Orville Davis states the Decision Case order in Field Manual v0: Question, Evidence, Recommendation, Human decision, Action, Verification, and Learning. The manuals index lives at /manuals. Evidence is step 02. The Evidence chapter keeps a visible split between observed fact, assumption, and what is still missing. A digital twin, a KPI model, a risk matrix, an RUL curve, a Monte Carlo run, or an LLM summary is not a fourth column that records the plant, names the failure mode, authorizes the work, or closes the case. A proposal whose only basis is the model has no evidence from the plant a later reader can inspect. A proposal with no named human decision is not a Decision Case. Honesty boundaries sit beside the spine: recommend is not authorize, evidence is required, and Stage-1 does not treat a model output as approved evidence from the plant or as live plant control. Sync refuses false precision. A number the model printed is not a fact the edition did not hold.

A model is useful only when its limits are explicit

A compressed representation keeps some of the plant and drops the rest. The digital twin keeps the state it was given and drops the state it was not. The KPI model keeps the inputs in the formula and drops the operational result those inputs stand in for. The risk matrix keeps the cell that was painted and drops the basis that was not written. The RUL curve keeps the projected line and drops the degradation modes that were not in the window. The Monte Carlo run keeps the spread of simulated outcomes and drops the inputs that were assumed rather than measured. The LLM summary keeps the sentences it wrote and drops the records it did not cite. Compression is not a fault in the tool. It is the limit of the record. The limit does not become the plant because the output looks finished.

Assumptions are the claims the model needs and does not prove: the failure mode it was built for, the operating regime it expects, the distributions it was fed, the documents the summary was allowed to read. Coverage gaps are the assets, sites, modes, sensors, or time windows the model was not given. Refusal conditions are the states in which the model must say it does not know: missing coverage, unmet assumptions, a disagreement with evidence from the plant, or a precision the evidence cannot support. A model can inform a recommendation to investigate when those three are explicit. Without them, the output is a claim with the limit hidden. Hidden limits are how false clearance gets stored.

The meeting can note that a model was run. Noting it records that a digital twin, a KPI model, a risk matrix, an RUL curve, a Monte Carlo run, or an LLM summary was available. It does not record the plant, and it does not open the Decision Case. The fact stays in the approved evidence a later reader can inspect, or it stays missing. The model stays a compressed representation with named limits, or it stays off the case as an unverified output. Treating the model as reality invents a plant the record does not hold. The room thinks the asset is cleared. The evidence from the plant was never written. That skip is the failure this essay names.

Treating the model as the plant is false clearance

False clearance is a record that says go, defer, or closed when the check is not on the case. The digital twin is nominal, so the room stores nominal as ready. The KPI model is inside band, so the band is stored as the outcome. The risk matrix cell is green, so green is stored as acceptable. The RUL curve is long, so deferral is stored as authorized. The Monte Carlo band is narrow, so the narrowness is stored as certainty. The LLM summary says the bearing is fine, so the sentence is stored as the inspection. Each of those moves is false clearance. The case looks released. The plant was not the source of the release.

False precision is the same failure in a finer unit. A remaining-life line printed to a day the evidence does not support, a probability printed to a tenth the inputs do not support, a risk score printed as if the cell were a measurement, an LLM sentence that sounds specific about a mode nobody cited — those are false precision. Sync refuses false precision. This essay states no OEM limit, no remaining-life figure, no probability, and no risk score. A number the meeting wants so the model will look decisive is not a diagnosis, and this page does not publish one. The case holds the approved evidence that was stored, the assumptions that were named, the coverage gaps that were named, the refusal conditions that were named, and what is still unknown. It does not hold a precision this essay made up.

Evidence from the plant beats the model. When the model and the plant disagree, the case keeps the disagreement in view. It does not overwrite the plant with the twin, the KPI, the cell, the curve, the band, or the summary. It does not overwrite a missing plant record with a confident output. A model that cannot show its assumptions, coverage gaps, and refusal conditions is not ready to sit beside the recommendation as if it were evidence.

The model is not a diagnosis and not root cause

A diagnosis names a failure mode a later reader can inspect: what failed, on what evidence from the plant, and what is still missing. Root cause is a recorded causal check that names why, on that evidence, and leaves the remaining uncertainty in view. A digital twin names neither. A KPI model names neither. A risk matrix names neither. An RUL curve names neither. A Monte Carlo run names neither. An LLM summary names neither, unless the summary is a pointer to approved evidence that already names them — and then the evidence names them, not the summary. The model is not a diagnosis. It is not root cause.

Symptom Is Not Cause keeps a vibration spike, a hot bearing, a pressure drop, an operator report, a CMMS complaint code, a red tile, or any other observed symptom from being stored as the cause. A model that labels the symptom as the mode has not recorded the cause. Threshold Is Not Diagnosis keeps a numeric threshold, a limit, an alarm band, or a set-point from being stored as a diagnosis. A limit inside a KPI model or on an RUL curve is still a limit. Correlation Is Not Causation keeps co-moving tags from being stored as a recorded causal check. A Monte Carlo of co-moving inputs is still co-movement. Trend Is Not Truth keeps a rising or falling line from being stored as a recorded check. An RUL curve is a projected slope. It is not the truth of the asset. Noise Is Not Signal keeps chatter from being stored as an actionable signal. Fitting the chatter does not promote it.

If the case cannot show the asset configuration, the failure and maintenance history, and the human judgment that connects the model to a mode, the output is a number or a paragraph without a trail. Evidence Lineage Is Not Optional already requires that trail. A digital twin does not shorten it. A KPI model does not shorten it. A risk matrix does not shorten it. An RUL curve does not shorten it. A Monte Carlo run does not shorten it. An LLM summary does not shorten it. The summary can cite the trail. Citing is not the same as being the trail.

Assumptions, coverage gaps, and refusal conditions stay on the case

Assumption Is Not Evidence keeps a belief, a habit, a prior shift's story, or an unverified model output from being stored as a recorded fact. The inputs to a Monte Carlo run are assumptions until they are measurements. The regime a digital twin expects is an assumption until the plant confirms it. The documents an LLM summary was allowed to read are a scope, not a proof that the scope was the plant. Naming the assumption keeps it in the assumption column. Hiding it inside the output stores it as fact.

A coverage gap is a part of the population the model did not see. Sample Is Not Population keeps a handful of observed cases from being stored as the fleet. A model fit to that handful is still a model of the handful. Assets that were not instrumented, sites that were not in the window, failure modes that were not in the history, and shifts that were not in the summary are coverage gaps. A gap that is named can inform a recommendation to investigate. A gap that is silent is how the model gets stored as the fleet. The model does not fill the gap by being confident about the rows it had.

A refusal condition is the recorded stop. The model refuses when an assumption is unmet, when a coverage gap is material to the question, when the output would be false precision, or when evidence from the plant contradicts the output. Refusal is not a failure of the tool. It is the honesty boundary on the output. A model that cannot refuse will clear a case it should have left unknown. Unknown stays unknown. The operator must not treat a missing refusal as go.

The model does not authorize corrective work

Authorization is the named act. A drafted next action is a proposal. Recommend is not authorize. Recommend Is Not Authorize keeps that proposal off the decision. A proposal that cites a digital twin, a KPI model, a risk matrix, an RUL curve, a Monte Carlo run, or an LLM summary is still a proposal. The curve can be smooth. The summary can be fluent. Neither the smoothness nor the fluency accepts the operational consequence. The model is not authorization for corrective work.

Corrective work is the change someone intends after a named decision: the repair, the deferral, the inspection, the isolation. A long RUL curve does not authorize deferral. A low cell on a risk matrix does not authorize running on. A Monte Carlo band does not authorize the work the band was used to justify. An LLM summary does not authorize the next sentence as the work order. Sync must not auto-authorize because the model showed a next step. Sync recommends. A named human decides. If the only support is the compressed representation, the proposal is not ready to be stored as authorized.

Recording an ACTION disposition is work intent. It is not plant execution. Action Is Not Execution keeps that intent off the plant. Sync does not write the work order. The systems that already hold execution authority write the work order or the isolation. Execution stays with authorized systems. A model, however precise, does not move execution onto Sync, and it does not replace the named decision that would have to come first.

An alert the model ranked is still an alert. Alert Is Not Decision keeps seeing, acknowledging, or silencing an alert from being stored as a decision. A dashboard of model scores is still a display. Dashboard Is Not Decision keeps that display off the decision. Color on a digital twin does not put it back as authorization, verification of outcome, work-order execution, or plant control.

A model is not a verified operational outcome

Verification records whether the authorized action did what it was supposed to do on the plant. The case stays open until that check is on the case. Verification stays open until a verified operational outcome is recorded. A later model state that says the asset recovered is a later compressed representation. It is not, by itself, the check. The KPI can move. The risk cell can change color. The RUL curve can lengthen. The Monte Carlo band can tighten. The LLM summary can say the work is done. The output does not record what the authorized change did.

Proxy Is Not Outcome keeps the stand-in off the outcome. A KPI model score is a stand-in. A green tile on a digital twin is a stand-in. Model clearance is not the check. Clearing the meeting because the curve looked long, the cell looked low, or the summary sounded finished must not be read as verification of outcome. Sync must not auto-close because the model was current. Sync must not treat model clearance as Learning credit. A closed model view is not a closed case.

Learning is what a later question is allowed to inherit. Learning Requires a Verified Outcome says the later question inherits that verified outcome, not a hope. Learning requires a verified outcome, not a hoped-for state from a digital twin, a KPI model, a risk matrix, an RUL curve, a Monte Carlo run, or an LLM summary. What was believed may include a labeled note that a model was consulted and that its assumptions, coverage gaps, and refusal conditions were explicit. Learning does not inherit the output as if the check had named the cause and confirmed the result on the plant. A later crew that defers the work because "the model said so," without the verification on the prior case, is inheriting a compressed representation. That is the wrong lesson into Learning. The Learning chapter publishes the closed record. It does not publish a model output as a verified operational outcome.

A model is not a Decision Case

A Decision Case opens on a question a named person can decide. The question is recorded before evidence or recommendation. Recording that a digital twin, a KPI model, a risk matrix, an RUL curve, a Monte Carlo run, or an LLM summary was opened does not record that question, and it does not record the decision. The case is the order: Question, Evidence, Recommendation, Human decision, then action, verification, and learning. A compressed representation is not that order.

Treating the model as reality shortcuts the Question → Evidence → Recommendation → Human Decision path. The output jumps to a conclusion. The evidence step is skipped, or it is filled with the model and called evidence from the plant. The recommendation is implied by the curve, the cell, or the summary. The human decision is treated as already made because the twin is nominal or the KPI is green. That shortcut is not a case. A named human records the Decision Case only against approved evidence. Until evidence from the plant is on the case, and until assumptions, coverage gaps, and refusal conditions are explicit, the model remains a compressed representation.

A named person accepts, rejects, escalates, or returns the recommendation. That act is the decision. Human Decision Is Not Optional says until that act is recorded, the case is still a proposal. A model does not replace a named human decision. The digital twin does not sign accept. The KPI model does not sign reject. The risk matrix does not escalate. The RUL curve does not return the proposal. The Monte Carlo run does not record who decided. The LLM summary does not stand in for the name. The signed-in Decision Case keeps who decided visible beside the recommendation. The recorded act is mandatory.

Sync may surface a model when the limits are named

Sync may surface a model beside approved evidence, with the honesty boundary that the model is a compressed representation and not the plant. Surfacing means the digital twin, the KPI model, the risk matrix, the RUL curve, the Monte Carlo run, or the LLM summary can sit beside a question so a person can see the assumptions, the coverage gaps, and the refusal conditions. The label is part of the surface. Without that label, the screen hides the limit and offers false precision. The honesty boundary is not optional.

Honesty Boundary Is Not Optional says Sync states what is simulated, seeded, or out of scope. Naming that limit does not promote a practice model into plant evidence. The Honesty boundaries chapter keeps Stage-1 from claiming a live conclusion the edition does not hold. A digital twin fed by simulated or seeded telemetry is a practice model. A Monte Carlo run on assumed inputs is a practice spread. An LLM summary of practice records is a practice paragraph. None of them is a live operational outcome, and none of them is a substitute for the approved evidence the operator still has to attach.

Sync must not auto-close, auto-authorize, or treat model clearance as Learning credit. Auto-close would store the output as a finished case. Auto-authorize would store the curve or the cell as permission for corrective work. Learning credit would let a later case inherit the summary. None of those acts is available to the system. A named human records the Decision Case only against approved evidence. Verification stays open until a verified operational outcome is recorded. Sync refuses false precision, including a remaining-life figure, a probability, or a risk score this essay does not hold.

Stage-1 evidence is the record held on the case. A live connector that pulls historian or control-system tags sits outside this edition. Simulated or seeded telemetry is a practice record. A practice model is still a compressed representation. It is not live plant evidence, and it is not a substitute for the check the operator still has to record. Sync refuses to treat a model as the plant.

Blank is not zero. Unknown is not clear. Silence is not clearance.

Blank is not zero. Blank Is Not Zero keeps an empty field from being stored as a measured zero. Filling the blank with a nominal value from the digital twin, a default from the KPI model, a midpoint from the risk matrix, or a sentence from the LLM summary collects a model value. It does not collect the missing plant value. In Recovery, economic assumptions that were left blank stay blank. They are not inferred from an RUL curve or a Monte Carlo run. A blank verification field is not a verified outcome of none, and a model that shows "no deviation" does not close it. Storing blank as zero invents false certainty and corrupts Recovery, verification, and Decision Case math.

Unknown is not clear. Unknown Is Not Clear records missing evidence as unknown, not as clear or ready. In the Decision Case and in Recovery, missing evidence stays unknown. Filling the unknown line with a twin state, a KPI, a risk cell, a remaining-life line, a simulated band, or a summary does not change the state to clear, and it does not name root cause. Bad or not-ready evidence stays blocked. A model that looks related to the blocked item does not relabel blocked as ready. The operator must not treat unknown as go because the model was complete. Go would mean the evidence is sufficient for the next recorded step. A compressed representation is not sufficiency. A coverage gap that was not named is still unknown.

Silence is not clearance. Silence Is Not Clearance keeps no reply, no alert, and an empty inbox from being read as approval. No reply is not approval. An empty inbox is not clearance. A quiet model is not controlled. A risk matrix with no red cell is not a decision. An LLM summary that does not mention a hold point does not write accept, reject, escalate, or return. It does not record a check that nobody performed. A nominal twin does not turn the missing response into the named decision. A model that stayed silent on its refusal conditions is still not clearance.

Sync recommends. Humans decide.

Sync may surface a model. Sync may hold the digital twin, the KPI model, the risk matrix, the RUL curve, the Monte Carlo run, or the LLM summary beside a recommendation to gather approved evidence from the plant, with the honesty boundary that the model is a compressed representation and that assumptions, coverage gaps, and refusal conditions have to be explicit. That is a proposal. The person who can accept the operational consequence records the decision. Until that act is on the case, the case is still a proposal. The system does not auto-accept because the model looked finished. Sync recommends. A named human decides.

Confidence Is Not Evidence keeps a strong gut feel, a high model score, or a manager's certainty from being stored as a recorded check. Being sure the twin matches the asset is still not the match. A narrow Monte Carlo band is still not a recorded check. A fluent LLM summary is still not a recorded check. The score on the model is not evidence from the plant.

Execution stays with authorized systems. Coverage of a model is not control of the plant. A Decision Case that records a model as a compressed representation does not put Sync in control of the work, the isolation, or the change. The Verification chapter is the check. The model is not the check. The Human decision chapter is the named act. The output is not the act.

A digital twin, a KPI model, a risk matrix, an RUL curve, a Monte Carlo run, or an LLM summary is a compressed representation, not the plant. It is useful for decisions only when its assumptions, coverage gaps, and refusal conditions are explicit. Treating the model as reality is false clearance. It is not a diagnosis, not root cause, not authorization for corrective work, not a verified operational outcome, and not a Decision Case. Evidence from the plant beats the model. Sync may surface a model when the limits are named. Sync refuses false precision. Sync must not auto-close, auto-authorize, or treat model clearance as Learning credit. A named human still decides. Verification stays open until a verified operational outcome is recorded. Execution stays with authorized systems.

The neighboring boundaries still hold

This boundary is not the neighboring ones. Each of those essays still holds. None of them promotes a digital twin, a KPI model, a risk matrix, an RUL curve, a Monte Carlo run, or an LLM summary to the plant or to a Decision Case.

Map Is Not Terrain is the picture: a diagram, a CMMS schema, a dashboard, a digital twin, or a process map stored as the operating reality. This essay is the computed claim on top of that picture. A twin that is a map is still not the terrain. A twin that is a model is still not the plant, and it is useful only with its limits explicit. Anecdote Is Not Evidence is the story. An LLM summary of a story is still not evidence. Anecdote Is Not Evidence keeps the retelling off the case. Sample Is Not Population is the handful stored as the fleet. A model of that handful is still not the population. Noise Is Not Signal is the chatter stored as an actionable signal. Symptom Is Not Cause is the observed symptom stored as the cause. Threshold Is Not Diagnosis is the line that was crossed. Assumption Is Not Evidence is the unverified model output stored in place of a fact. Confidence Is Not Evidence is the high score stored in place of a check. Correlation Is Not Causation is the chart of co-movement. Trend Is Not Truth is the slope, including an RUL curve read as truth. Proxy Is Not Outcome is the stand-in, including a KPI model read as the result. Verification Is Not Optional is the check that still has to be written. A fluent summary is still not evidence. A tight band is still not a check. A green cell is still not the outcome. A later model state does not close the case. The label stays. The evidence from the plant stays unwritten until it is on the case. The decision stays unwritten until a named person records it.

Unknown, blank, and silence are still the missing record. A model does not interpret them into clearance. On this record, blank is not zero, unknown is not clear, and silence is not the named decision. A case that is unknown, blank, or quiet is not made ready by a nominal twin, a green KPI, a low risk cell, a long RUL curve, a narrow Monte Carlo band, or a finished LLM summary.

The check that closes the case is still verification. The lesson a later case may inherit is still the verified operational outcome, not a hoped-for state from the model. The trail behind a recommendation is still evidence lineage. An alert is still not a decision. A dashboard is still not a decision, and it is still not the plant. Recommend is not authorize. Those records are either on the case or they are not. If they are not, a model does not supply them. Evidence from the plant beats the model.

Where the public statement lives

Field Manual v0 is the public contents of the Decision Case. Start at the manuals index or open Sync Field Manual directly. Evidence is step 02. Observed fact, assumption, and what is still missing stay distinct. A model does not collapse that split into evidence from the plant. A proposal with no basis is not ready for the human decision that follows. Verification is step 06. A model that says the work "already recovered" is not the check. Learning is step 07. A later case inherits the verified outcome, not the digital twin, the KPI model, the risk matrix, the RUL curve, the Monte Carlo run, or the LLM summary the last shift treated as the lesson. Honesty boundaries keep the edition from treating a compressed representation as live evidence or as plant control, and they keep the edition from printing false precision. Later editions can deepen a chapter. The spine stays in this order.

What this article is not claiming

This is an essay about the record, not a customer case study. It names no plant, states no savings figure, and claims no prevented failure. It states no OEM limit and supplies no remaining-life figure, no probability, and no risk score. It does not claim that Sync executes plant work, writes a work order, isolates equipment, or controls a plant. It does not claim that a digital twin, a KPI model, a risk matrix, an RUL curve, a Monte Carlo run, or an LLM summary is the plant, that a compressed representation is evidence from the plant, that a model is a diagnosis, that a model output is root cause, that a curve or a cell is authorization for corrective work, that a green score is a verified operational outcome, or that a model is a Decision Case. It does not claim that self-guided onboarding is a live product path.

Stage-1 readiness means a signed-in user can complete the Decision Case — question, evidence, recommendation, human decision, action, verification, and learning — and Field Manual v0 describes that journey. This edition does not describe plant execute, a live connector tag pull, SMTP invite delivery, or automatic revocation of access on expiry as live. It does not describe Sync writing work orders, isolating equipment, or controlling the plant. Simulated or seeded telemetry and assets are practice records. They are not live plant results, and a practice model is not the plant. Self-guided onboarding is not claimed as a live product path.

The series continues with Forecast Is Not Fact, on why a forecast, a projection, an RUL estimate, a demand plan, a Monte Carlo percentile, or an LLM-generated outlook is a forward-looking compression and not a fact about the plant. Companion reading: Map Is Not Terrain on why a diagram, a CMMS schema, a dashboard, a digital twin, or a process map is a compressed representation and not the operating reality, Proxy Is Not Outcome on why a stand-in measure is not the verified operational outcome, Assumption Is Not Evidence on why an unverified model output is not a recorded fact, Confidence Is Not Evidence on why a high model score is not a recorded check, Recommend Is Not Authorize on why a drafted next action is still a proposal, Dashboard Is Not Decision on why a display is not authorization, Sample Is Not Population on why a handful of observed cases or plants is not the fleet, Noise Is Not Signal on why sensor chatter is not an actionable signal, Symptom Is Not Cause on why an observed symptom is not the cause, Threshold Is Not Diagnosis on why crossing a numeric threshold is not a diagnosis, Correlation Is Not Causation on why co-moving tags are not a recorded causal check, Trend Is Not Truth on why a slope is not a recorded check, Verification Is Not Optional on why the case stays open until the check is recorded, Alert Is Not Decision on why seeing, acknowledging, or silencing an alert is not a decision, Silence Is Not Clearance on why silence, no reply, no alert, and an empty inbox are not authorization, Evidence Lineage Is Not Optional on why a recommendation you cannot reconstruct is not an industrial decision, Learning Requires a Verified Outcome on why a later case inherits the verified outcome, not a hoped-for state from a model, Action Is Not Execution on why work intent is not plant execution, Anecdote Is Not Evidence on why a story is not evidence, Blank Is Not Zero on why an empty field is not a measured zero, Unknown Is Not Clear on why missing evidence is recorded as unknown, not as clear or ready, and Human Decision Is Not Optional on why the named act is mandatory. A Reliability Assessment asks whether the records can support a conclusion. A Strategic Pilot is a governed proof around one operating decision. The Verification chapter is the public statement of the check this essay refuses to replace with a model.