The Geometry of Getting Home

Optionality, Recoverability, and Decision Under Moving Constraints

The world changes after you make the plan. Most itineraries are built as if the difficult part were deciding what to do before that happens.

I arrived in Philadelphia after midnight with the following day already organized: the first useful transit after morning, Independence Hall, Congress Hall, food, architecture, a tour of the Masonic Temple, perhaps a museum, then the airport and a flight home to Chicago. On paper, the day was elegant. Its efficiency was part of the attraction. Air travel had compressed a major American city into something almost improbable: leave one city at night, spend a day inside another history, return home before the day after that had properly begun.

The itinerary represented this as a line.

The body supplied the first correction. The airport yielded almost no sleep. Morning began early. Walking accumulated. Heat accumulated. One planned meal proved disappointing. A scheduled tour moved earlier. By midday, some attractions that had made obvious sense from Chicago no longer made obvious sense from Philadelphia. Their admission prices had not necessarily changed. Their collections had not deteriorated. The streets between them had not moved. Yet the decision had changed because the state around it had changed.

There was also a hard boundary. We needed to reach the airport with enough time to make the flight home. That requirement was not simply the final item on the itinerary. It imposed a constraint backward across everything that preceded it. One more attraction could remain independently worthwhile and become irrational as part of the day. Another hour in the city might purchase experience by spending the margin protecting the return.

The resulting question was no longer simply:

What would be worth doing?

It was:

What is worth doing now, from here, under these conditions, given what doing it will leave possible afterward?

That is a different decision problem.

An itinerary describes intended events. It is useful precisely because it converts a difficult field of possibilities into a tractable sequence: departure, arrival, reservation, attraction, meal, return. But the simplification comes at a cost. The itinerary remembers the decision that was made. It does not necessarily represent whether the assumptions that justified the decision remain true.

A plan can therefore become wrong without having been wrong when it was made.

The proposition of this chapter is that under changing and partially observed conditions, the rational object of travel planning is not a predetermined itinerary alone but a policy over subsequently encountered states; and where a traveler faces a binding return or terminal constraint, a present action should be judged partly by whether it destroys useful future actions without sufficient compensating value.

The last qualification matters. Preserving options cannot itself be the objective. A traveler who refuses commitment in order to preserve every future possibility eventually preserves the possibility of doing everything by doing almost nothing. Tickets must sometimes be bought. Reservations must sometimes become irreversible. Cities must be entered rather than indefinitely contemplated.

The relevant good is more demanding than flexibility.

It is option-preserving judgment: surrendering future possibilities when present value, necessity, or improved reliability justifies their loss, while refusing to collapse useful future action merely because an earlier plan once made the collapse look convenient.

The distinction begins with a problem transportation science has understood for decades.

A Path Is Not a Policy

In 1986, Randolph Hall studied travel through networks in which travel times were both random and time-dependent. Standard shortest-path logic was insufficient. More important for the argument here, Hall showed that the optimal “route choice” need not be a predetermined simple path. The best onward route from an intermediate node can depend on the time at which the traveler actually reaches that node. Because that arrival time is unknown before departure, the better strategy can be to defer part of the routing choice until more of the state has been revealed.¹

The difference is small enough to express in two lines.

A path is a sequence:

A policy is a decision rule:

The notation is elementary. The ontological correction is not. A path specifies where to go if the relevant assumptions survive. A policy specifies what to do from the state that actually occurs.

Tarun Rambha, Stephen Boyles, and S. Travis Waller later formalized adaptive transit routing in a stochastic, time-dependent network as a finite-horizon Markov decision process. Their routing strategies condition decisions on intermediate arrival times and real-time information, and their formulation includes a constraint requiring arrival at the destination within a threshold.² The problem is narrower than the one considered here—the objective is transportation routing, not whether another hour of travel remains worth its cost to an exhausted traveler—but the structural lesson is decisive: under uncertainty, later actions can rationally depend on information that did not exist when the initial plan was made.

ReturnVector therefore does not originate the distinction between a path and an adaptive policy. Nor does it invent stochastic routing, deadline-aware routing, robust optimization, time-dependent graphs, accessibility theory, or the idea that uncertainty can justify preserving flexibility. Any serious account must concede those intellectual territories before claiming its own.

The residual problem appears when routing becomes travel.

A transportation network can tell us that an edge remains traversable while remaining silent about whether that edge is economically tolerable, physically executable, sufficiently evidenced, compatible with a companion, or rational relative to the experience being purchased. The same train can exist for two travelers while belonging to the feasible decision set of only one. The same museum can be valuable to the same traveler at 10:00 a.m. and not valuable enough at 2:00 p.m. The same return route can be chronologically possible and too fragile to support another discretionary stop.

The state variable therefore has to become richer than network position.

Let s_t denote the state of the decision at time t. For present purposes it includes, at minimum, location, time, currently known transportation and inventory conditions, relevant costs and resources, declared traveler constraints, traveler state, and the evidentiary status of the claims on which the decision depends.

This is a conceptual representation, not a description of a calibrated production model. That distinction will matter later.

What matters now is that an itinerary records a path through anticipated states, while an adaptive travel system asks repeatedly what action is warranted from the state that has actually been reached.

The harder question follows immediately.

What, exactly, should such a system preserve?

The Futures We Can Actually See

The language of “preserving optionality” is attractive and dangerously loose.

A future possibility can be written down without being usable. A flight can appear in a timetable with no seat available. A train can run after the traveler has lost the ability to reach it. A transfer can be geometrically imaginable and chronologically impossible. An accessible route can exist on a generic map while failing to meet the needs of the person trying to execute it. A backup can depend on exactly the same point of failure as the plan it supposedly backs up.

The relevant object is therefore not the set of conceivable futures.

It is the set of credible continuations.

Even that phrase requires an epistemic correction. No decision system observes the true set of futures that will remain feasible. Future weather, inventory, delays, traveler condition, operating failures, and other contingencies remain partly unknown. There may in principle exist some actual set,

of continuation policies that would, given the future as it eventually unfolds, permit the traveler to satisfy a declared terminal constraint R.

But that set is not available to the traveler or the model at decision time.

The system can represent only an estimated, evidence-bounded continuation set:

Here R is the declared terminal constraint—arrive at the airport by a chosen time, reach home before a required deadline, preserve an accessible route to lodging, or satisfy another terminal condition. The parameter θ does not denote a discovered constant of nature. It denotes the system’s declared admissibility rule: the standard of evidence and feasibility required before a continuation is allowed to count as presently credible.

A continuation enters the estimated set only if it clears that rule. Chronology must work under the information presently available. Necessary transportation must remain obtainable to the degree required by the decision. Known traveler constraints must not invalidate the route. The evidence supporting the continuation must be sufficiently current and direct for the consequence at stake. Unknown facts remain unknown.

This resolves a central danger in the original formulation. ReturnVector need not pretend to know the future action set. It needs to represent, as honestly as possible, which future actions current evidence still gives us reason to treat as real.

The distinction is not pedantic.

Suppose a screen shows four apparent routes home. The underlying transportation network may indeed contain four paths. Yet one may depend on inventory that has not been checked in several hours. Another may require a transfer that is no longer chronologically credible. A third may be unusable given a declared mobility constraint. The fourth may remain strongly evidenced and executable.

The system should not say that the traveler has four backups.

It should say that the network contains four candidate routes and that, under the current evidence rule, only one presently qualifies as a credible continuation.

The system’s own uncertainty is part of the state.

That makes the chapter’s governing sentence stricter:

A backup is not a backup until it still exists.

The word doing the work is still.

A backup is usually treated as a noun. We have a later flight. Another train. A second airport. A rental car. A different connection. The alternative enters the plan, and its presence creates reassurance.

Still converts the noun into a temporal proposition.

Does the flight still have obtainable inventory? Can the traveler still reach it? Does the transfer still close? Does the fare rule still permit the action contemplated? Does the traveler still possess the money, mobility, physical capacity, documentation, or companion agreement required to execute the route? Does the alternative still survive the disruption threatening the primary plan?

The sentence changes the ontology of recovery. A backup is no longer something found during planning. It is something that must remain true at decision time.

HomeGraph

ReturnVector gives this recovery structure a name: HomeGraph.

The term should not be mistaken for a new graph-theoretic object. It is not. Stochastic transportation networks, time-dependent networks, adaptive policies, correlated link times, and reliability problems long predate it. HomeGraph is instead a traveler-facing representation of a narrower decision problem: which currently evidenced movements can still be assembled into credible continuations satisfying this traveler’s declared terminal constraint?

That limitation strengthens rather than weakens the construct.

In abstract form, let the observed recovery network at time t be

where an edge enters the observed set only if it satisfies the current admissibility rule. Relevant edge attributes may include departure window, observed or estimated availability, price, duration, transfer requirements, evidence age, evidence source, and traveler-specific executability.

The point of HomeGraph is therefore not to draw more routes. It is to refuse to count routes that current evidence does not warrant counting.

A second correction is equally important: route count is not recovery depth.

He Huang and Song Gao study adaptive routing in stochastic time-dependent networks with temporally and spatially correlated link travel times. Their results show, among other things, that correlation changes the value of adaptive routing; dependency among uncertain links is therefore not a peripheral detail of network reliability.³

For ReturnVector the implication is bounded but important. Two routes that look different can share a critical vulnerability. They may depend on the same hub, the same downstream segment, the same transportation mode, the same infrastructure, or another observable failure domain. A system should not infer statistical dependence that it cannot estimate, but neither should it grant full redundancy simply because two routes have different names.

Two written itineraries are not necessarily two independent recoveries.

That distinction produces two forms of margin.

Structural margin asks how many meaningfully differentiated credible continuations remain.

Temporal margin asks how long those continuations remain actionable.

A traveler can have broad but shallow recovery: several apparent ways home, each about to expire. Another traveler can have narrow but deep recovery: one strongly evidenced path with substantial time margin. Raw route count cannot distinguish the two.

The difference becomes clearer in a stylized decision state.

Suppose a traveler in Center City has declared a hard constraint: reach the airport no later than 4:20 p.m. At 1:00 p.m., four apparent return routes are visible.

The first is a rail departure with comfortable schedule margin and currently verified service. The second is a later departure on the same rail system that reaches the airport close to the constraint. The third begins from another station but converges on the same critical rail segment as the first two. The fourth is a road transfer that costs substantially more but avoids the rail dependency.

Now add one traveler constraint: because of pain and accumulated walking, transfers requiring more than a declared walking threshold are no longer acceptable.

The map still shows four routes.

The decision system should not.

The second route has little temporal margin. The third adds less structural recovery than its separate itinerary suggests because it shares the critical dependency. A version of the road transfer that requires excessive walking may fall outside the traveler’s feasible set altogether. What appeared to be four backups may reduce to one robust rail option plus one genuinely differentiated but expensive road fallback.

Now consider the decision to cross town for another attraction.

Its cost is no longer adequately represented by admission plus travel time. The action may consume the comfortable rail departure, leave only the thin-margin rail option, and reduce the road fallback by placing the traveler farther from a usable pickup location. The attraction can therefore be excellent while the move required to reach it is poor.

Nothing probabilistic has to be fabricated to expose the structure.

The system can say: if you make this move, this presently verified continuation disappears; these two remaining routes share a known dependency; this other route remains feasible but costs more; and your declared walking constraint removes the final alternative.

That is already useful.

It is also falsifiable. If representing this structure does not alter decisions, improve recovery, clarify tradeoffs, or outperform a much simpler deadline-and-alternatives display, then the additional architecture has not earned its complexity.

Every Action Edits the Future

The continuation set makes explicit something ordinary planning often hides: an action has both an immediate consequence and a continuation consequence.

Buying a ticket obtains transportation and removes the possibility of spending that money elsewhere. Waiting preserves cash and may lose the seat. Crossing a city purchases experience and consumes return margin. Choosing a cheaper connection preserves money while introducing another point of failure. Choosing a direct flight sacrifices some financial flexibility to purchase operational simplicity.

No single direction is inherently rational.

This is where the rhetoric of flexibility becomes dangerous. The objective cannot be

A traveler who maximizes the number of remaining branches will resist exactly the commitments that make travel possible.

The correct distinction is between option loss and unjustified option loss.

Dimitris Bertsimas and Melvyn Sim’s work on robust optimization is instructive because it formalizes a trade that ordinary planning rhetoric often ignores: one may rationally accept some loss relative to a nominal optimum in exchange for protection against uncertainty. Their “price of robustness” is precisely the recognition that robustness is valuable but not free.⁴ A travel day is not their linear optimization problem, and the mathematics should not be imported by metaphor. The operative structure is narrower: maximizing performance under nominal assumptions can be inferior to accepting some nominal sacrifice in order to remain feasible under variation.

This is why the cheapest route is not always the least costly route and the safest route is not always the best trip.

Too little protection creates brittleness. Too much protection converts travel into defensive logistics.

The rational question is what present value justifies what loss of future maneuverability.

Kenneth Arrow and Anthony Fisher’s analysis of uncertainty and irreversibility provides another bounded mechanism. Their subject is environmental preservation, with stakes that should not be trivialized through equivalence to tourism. Their decision-theoretic point is nonetheless useful: when an action is irreversible and future information may change the decision, retaining the ability to wait can itself possess value.⁵

Travel supplies the missing counterpressure.

Waiting also destroys options.

Inventory disappears. Departure windows close. Weather can worsen. Daylight ends. A body accumulates fatigue. A price can rise. A later connection may become impossible. No traveler can stand outside time until uncertainty resolves.

The question is therefore not whether to preserve optionality.

It is when commitment has become more valuable than the option to defer it.

That is the proper problem for a decision timer.

Waiting Has an Expiration Time

Some decision boundaries are hard.

If the final usable train departs at 3:15, a proposed activity requiring movement until 3:20 is incompatible with that return path. If boarding closes at a known time, the constraint can be propagated backward through required transportation, transfer, and buffer. If a fare expires at midnight, the option expires at midnight.

No prediction is required.

Other boundaries are soft.

A traveler may ask whether to buy a seat now or wait for price movement. Whether to rebook during a developing disruption or see if the original flight recovers. Whether to leave for the airport now or spend another hour in the city. Those decisions depend on beliefs about future states, not just known deadlines.

The original ReturnVector language risked collapsing these different epistemic objects into one timer.

It should not.

A hard decision boundary follows from constraints the system has adequate reason to treat as given.

A soft decision boundary estimates when the value of further waiting has fallen below the value of acting. That is an optimal-stopping-like problem and requires substantially more model maturity.

The distinction matters because a countdown clock looks authoritative regardless of what produced it.

ReturnVector can therefore know that waiting becomes impossible before it has earned the right to say exactly when waiting becomes irrational.

The first statement may follow from schedules, connection requirements, declared buffers, and other direct constraints. The second may require predictions about inventory, delay, price, recovery, or traveler utility.

An interface should display the difference.

“Last departure preserving your 4:20 airport constraint: 3:05” is one kind of claim.

“Recommended decision deadline: 2:17 because waiting longer is expected to reduce your outcome” is another.

The second number must earn far more epistemic authority than the first.

This distinction also prevents a familiar pathology in computational decision systems: converting uncertainty into precision simply because precision is easy to display.

A clock can be exact when the model is not.

The Traveler Is Not Cargo

The network is still incomplete.

Torsten Hägerstrand’s foundational work in time geography insisted that human activity be understood inside the spatial and temporal constraints through which actual persons move. His 1970 address, “What About People in Regional Science?,” resisted abstractions that treated people as interchangeable units in regional models and helped establish a framework attentive to the constraints governing individual action in space and time.⁶

The relevance to adaptive travel is direct but bounded.

A transportation edge is not automatically a human option.

Karst Geurs and Bert van Wee make a related point in the accessibility literature. Their review argues that more theoretically adequate accessibility measures should account for individual and spatial-temporal constraints rather than relying only on simpler mobility measures such as speed.⁷ Accessibility is not reducible to the existence of transportation infrastructure; it concerns the opportunities that can actually be reached under the conditions governing the traveler.

This distinction becomes ethically sharp when the generic network is presented as though it were equally usable by everyone.

Jill Bezyak, Scott Sabella, and Robert Gattis surveyed 4,161 people with disabilities about public transportation and complementary paratransit. Their respondents continued to report substantial physical and attitudinal barriers despite decades of formal accessibility requirements.⁸ The significance for an adaptive decision system is not that disability can be turned into another scalar penalty. It is the opposite: routes that a generic planner calls available may not belong to the practical action set of the person expected to use them.

The traveler model must therefore be allowed to remove an edge that the infrastructure map retains.

A route requiring stairs may not exist for one traveler. A route requiring a particular walking speed may not exist in the traveler’s present state. An itinerary demanding a transfer through inaccessible infrastructure is not “less preferred” if the traveler cannot execute it. It is not a feasible continuation.

This is the difference between network connectivity and traveler-specific executability.

It is also why the system should initially prefer declared constraints to inferred physiology.

Human condition does matter. In a controlled study of chronic sleep restriction and total sleep deprivation, Hans Van Dongen and colleagues found cumulative neurobehavioral performance deficits under repeated sleep restriction.⁹ During a Boston heat wave, Jose Guillermo Cedeño Laurent and colleagues found poorer performance on cognitive tasks among young adults living in hotter non-air-conditioned environments than among peers in air-conditioned buildings.¹⁰

Neither study licenses an application to convert a traveler’s hours of sleep or ambient temperature into a personalized coefficient of rationality.

That temptation should be resisted.

“I am exhausted.” “My knee hurts.” “I will not walk another twenty minutes.” “My companion needs food.” “I am willing to spend fifty dollars to reduce transfer friction.”

These are not primitive data awaiting replacement by a physiological model. They are legitimate declarations of state and preference.

The system’s first obligation is to respect them.

A technically possible route can therefore become humanly infeasible without becoming medically impossible.

That middle category is exactly where many travel decisions occur.

Option Scarcity Is Not Option Destruction

The ethical problem is larger than disability.

Some travelers begin with broader continuation sets because they possess money, schedule flexibility, documentation, physical capacity, social support, or access to transportation modes that others do not.

Karen Lucas’s work on transport and social exclusion traces the relationship among transport disadvantage, poverty, access to essential activities, and wider forms of exclusion.¹¹ The point is not that every inexpensive trip reproduces structural inequality. It is that the resources required to convert disruption into recovery are unevenly distributed. A taxi, replacement flight, unplanned hotel, refundable fare, or missed workday does not carry the same consequence for every traveler.

This requires a distinction that an adaptive system can easily miss:

Option scarcity is not the same thing as option destruction.

Option scarcity concerns futures the traveler did not meaningfully possess.

Option destruction concerns futures that a current action removes.

A traveler unable to afford an unplanned hotel does not make a poor decision merely because the model’s generic recovery graph contains a hotel. A wheelchair user has not “rejected” an alternative that depends on inaccessible infrastructure. A caregiver with a hard return obligation does not possess the same lateness tolerance as a traveler whose return date is discretionary.

If a system treats those differences as deviations from a universal traveler, it will produce a peculiar form of algorithmic moralism: people with fewer feasible choices will appear to be worse optimizers of choices they never had.

The architecture should prohibit that inference.

The traveler-specific continuation set should distinguish, insofar as the data permit, between an action removed by the traveler and an action absent because the relevant resources or conditions never made it feasible.

That distinction gives personalization a defensible boundary.

The system does not become better merely by knowing more intimate facts about a traveler. More data can produce surveillance without producing better judgment.

Personalization earns its place when it prevents the system from presenting theoretical transportation possibilities as actual human options.

The justice question is therefore operational rather than decorative:

Whose theoretical route is being mistaken for a real choice?

Once that question is asked, “accessibility” can no longer mean that the route exists somewhere in the network.

It must mean that the traveler can plausibly use it.

Bounded Judgment

This does not imply that travelers themselves perfectly know what is rational.

Herbert Simon’s critique of idealized rational choice is relevant precisely because travel decisions occur under severe informational and computational limits. His 1955 “Behavioral Model of Rational Choice” challenged models requiring decision makers to possess and process unrealistic amounts of information.¹²

The lazy technological interpretation would be that human boundedness creates a vacancy for algorithmic omniscience.

It does not.

The system is bounded too.

It does not know the true future continuation set. It has incomplete evidence. Its estimates can be stale. Its models can be misspecified. Its training or observation data can fail under new conditions. Its utility representation can omit something the traveler cares about. Its apparent precision can exceed its calibration.

The useful implication of bounded rationality is therefore more modest.

A decision system can reduce the cost of maintaining an adequate picture of the state.

It can remember which alternatives have expired. It can show which remaining routes share a known dependency. It can propagate a hard deadline backward. It can expose that an inventory claim is stale. It can remember a declared mobility constraint when a route search does not. It can show which assumption is carrying the recommendation.

That is already substantial cognitive assistance.

The system need not become the sovereign chooser.

Its legitimate role is to improve the conditions under which the traveler chooses.

The Machine Can Become the New Risk

A hostile reader should now object that the entire architecture is unnecessary.

People already do this.

Travelers check the weather, notice they are tired, abandon attractions, leave early for airports, pay more for direct routes, consult their partners, search alternate flights, and change their minds. “Continuation set,” “HomeGraph,” and “decision timer” may simply turn practical intelligence into technical vocabulary. The result could be worse than redundancy. It could be precision theater.

Suppose an interface announces:

Return Integrity: 82.

The number arrives with an authority that ordinary language does not possess. It appears measured. If rendered as 82.4, it appears measured more carefully. Yet the estimate may depend on stale inventory, weak evidence about disruption, arbitrary utility weights, known-but-unquantified dependency among routes, and an immature traveler model.

Nothing about the typography reveals that.

A system can therefore increase danger by making uncertainty persuasive.

This is not a hypothetical concern in the abstract study of automation. Raja Parasuraman and Victor Riley’s review of human use of automation identifies “misuse” as overreliance that can contribute to monitoring failures and decision bias, while also emphasizing that trust, workload, risk, reliability, and individual differences complicate the prediction of automation use.¹³ Lisanne Bainbridge’s earlier “Ironies of Automation” argues that automation can relocate rather than eliminate human difficulty, leaving operators responsible for abnormal conditions precisely when maintaining the necessary expertise and awareness can be hardest.¹⁴

Neither source proves that travelers will overtrust ReturnVector.

They establish the design burden.

A system that advises under uncertainty must treat the calibration of human reliance as seriously as the calibration of its own forecasts.

This objection changes the architecture in four ways.

First, confidence cannot be hidden inside the score.

Second, heuristic estimates must not be presented as calibrated probabilities.

Third, the evidence supporting a recommendation must remain inspectable enough for disagreement to be informed.

Fourth, the traveler must retain the right to override the recommendation without the product treating disagreement as evidence of irrationality.

The objective is therefore not automated compliance.

It is contestable decision support.

A good recommendation should be capable of explaining why it deserves less trust.

That requirement leads to the most severe part of the architecture.

A Score Is Not a Fact

Numerical output is cheap.

Calibration is not.

Tilmann Gneiting, Fadoua Balabdaoui, and Adrian Raftery distinguish calibration, the statistical consistency between probabilistic forecasts and observations, from sharpness, the concentration of the predictive distributions. Calibration is a relationship between predictions and what subsequently happens; it cannot be inferred from how precise the predictions look.¹⁵ Gneiting and Raftery’s work on strictly proper scoring rules similarly develops methods for evaluating probabilistic forecasts in ways designed to reward truthful and accurate probabilistic assessment.¹⁶

The implication for ReturnVector is non-negotiable.

If the system predicts before reality answers, the prediction must survive long enough to be judged by the answer.

A minimum prospective record has the form

The prediction is frozen before the outcome.

The outcome is observed independently.

If feature definitions, weights, targets, or transformations change materially, the changed system becomes a new model version.

A miss remains a miss.

This is not sophisticated machinery. It is elementary epistemic hygiene.

Without it, retrospective repair can turn a forecasting system into a narrative system. A model can appear perpetually insightful because yesterday’s uncertainty has been rewritten in the vocabulary of today’s outcome.

The right development culture preserves embarrassment.

Only then can the system ask serious questions later:

• Was the model calibrated?
• Was it systematically overconfident?
• Did performance differ across route types or traveler configurations?
• Did a complicated model outperform a simple baseline?
• Did a feature improve discrimination or simply produce more impressive numbers?
• Did the recommendation improve the outcomes travelers actually value?

ReturnVector currently operates at Stage 0 with transparent heuristic models. Its architectural targets, prospective observation discipline, HomeGraph evidence requirements, prediction-outcome separation, and versioning rules can be defined now. Its current outputs should not therefore be described as empirically calibrated probabilities.

The architecture is ahead of the evidence.

That is not an embarrassment.

Pretending otherwise would be.

The same logic requires the system to decompose what ordinary interfaces call confidence.

Evidence freshness asks when the relevant observation was obtained.

Evidence directness asks whether it comes from an operational source, authoritative rule, indirect inference, historical pattern, or user report.

Evidence completeness asks whether material variables needed for the estimate are missing.

Model uncertainty concerns the reliability of the mapping from evidence to prediction.

These are distinct.

A departure status can be recent and direct while downstream connection success remains uncertain. A fare rule can be authoritative yet incomplete for irregular operations. A weather forecast can be high quality while the traveler’s decision threshold remains poorly estimated.

A single confidence number can conceal all of this.

Confidence should therefore describe the epistemic condition of an estimate, not cosmetically strengthen its authority.

This is one of the places where the chapter’s argument should govern product design rather than explain it after the fact.

What ReturnVector Actually Adds

Once the inherited intellectual terrain is cleared, the novelty claim becomes smaller and more defensible.

ReturnVector does not claim to have invented adaptive routing. Hall already makes the path-policy distinction. It does not invent stochastic transit routing; Rambha, Boyles, and Waller formalize it. It does not invent dependent travel-time uncertainty; Huang and Gao explicitly address correlated stochastic networks. It does not invent robustness; operations research has developed that problem extensively. It does not invent accessibility or individual space-time constraints; transport geography and accessibility research already supply those traditions. It does not invent human-automation risk or probabilistic calibration.

The proposed contribution is architectural.

ReturnVector joins seven decision objects that are often treated separately:

• adaptive state-dependent action;
• a traveler-declared terminal constraint;
• an evidence-bounded representation of credible continuation;
• dependency-aware recovery rather than raw route multiplicity;
• traveler-specific executability rather than generic network connectivity;
• commitment timing that distinguishes hard constraint boundaries from model-dependent stopping judgments;
• and prospective accountability that preserves predictions long enough for outcomes to disprove them.

The strongest version of the claim is therefore not:

ReturnVector solves adaptive travel.

It is:

ReturnVector proposes a traveler-facing decision architecture in which the object being maintained is an evidence-bounded set of traveler-executable continuation policies capable of preserving a declared terminal constraint while present utility is pursued.

That is more cumbersome than a product slogan because scholarship and product marketing have different jobs.

The architecture should survive only if the distinction proves consequential.

There are several ways it could fail.

• If a simple itinerary plus deadline alerts performs as well as the richer state representation, the additional machinery is not justified.
• If dependency-aware recovery does not improve decisions over raw alternative counts, that construct should be reduced or removed.
• If traveler-specific executability rarely changes the action set relative to ordinary accessibility settings, its role should be narrowed.
• If evidence provenance does not improve forecast evaluation or human reliance, the interface burden may not be worth carrying.
• If prospective testing shows that heuristic outputs fail to beat simple baselines, the heuristics should not be promoted merely because they are proprietary.
• If users systematically over-rely on the system despite uncertainty displays, the design may be harmful even when its predictions improve.

A serious architecture needs these kill conditions.

Otherwise every failure can be redescribed as a reason for further complexity.

Return Integrity

With those limits established, the five current ReturnVector dimensions can be stated more precisely.

Seat Availability Likelihood concerns the evidence that constrained inventory will be obtainable at the decision window that matters. Until prospective calibration supports probability claims, its output remains a heuristic estimate.

Disruption Risk concerns exposure to operational degradation relevant to the contemplated action. It should distinguish observed conditions from inferred risk wherever possible.

Return Integrity concerns the quality of the currently represented recovery structure if the intended path fails. It is not a synonym for optimism about the primary route.

Trip Utility concerns whether the trip or next action remains worth its full cost to this traveler in this state: monetary cost, time, experience value, friction, route geometry, weather, and declared human constraints.

Confidence concerns the evidentiary and model condition of the estimate and remains separate from the estimate itself.

Return Integrity deserves the most care because it is the bridge between the abstract continuation set and practical recovery.

It should not collapse immediately into one opaque score.

Conceptually it contains at least four separable elements:

• recovery diversity;
• temporal margin;
• evidence quality;
• traveler executability.

Their future empirical relationship is a modeling question. The chapter should not invent weights.

The point is to make visible why a route that appears on a screen may contribute little to actual recoverability.

A low-confidence backup is not equivalent to a verified one.

A highly correlated backup is not equivalent to a diversified one.

A route about to expire is not equivalent to one with substantial margin.

A technically operating route is not equivalent to one the traveler can use.

The architecture earns complexity only where those distinctions alter judgment.

The Return to Philadelphia

By midday in Philadelphia, nothing catastrophic had happened.

That is what made the day useful.

The city had not become inaccessible. Flights had not been canceled around us. No emergency had forced a decision. Reality had simply accumulated enough small changes that the original itinerary no longer represented the decision state particularly well.

We had seen much of what we came to see. We had walked farther than the plan could feel when viewed from home. Heat had become part of the cost function. A scheduled event had shifted. A disappointing meal altered the marginal attraction of another food detour. The airport remained fixed in the future, but its constraint was moving closer through time.

The temptation at such a point is fidelity.

We planned this. We came all this way. It is on the list. We may not be back soon.

These sentences have emotional force because prior intention easily disguises itself as present evidence.

It is not.

The fact that an attraction deserved a place in yesterday’s plan does not establish that it deserves the next hour today.

Adaptive judgment requires a small act of disloyalty to one’s former self.

The state now has evidentiary priority over the plan then.

That does not mean obeying every present aversion. Fatigue can encourage premature retreat. Anxiety can exaggerate risk. A traveler can protect the return so aggressively that the trip collapses into waiting for the return. Human judgment remains fallible in both directions.

The objective is not responsiveness without structure.

It is repeated comparison between current value and current constraint.

The best action will therefore not always maximize immediate pleasure. It will not always minimize expected travel time. It will not always minimize cost. It will not always maximize the number of future options. It will not always maximize return security.

Each objective becomes pathological when allowed to govern alone.

The more defensible object is a policy capable of revising action as evidence, constraints, and traveler state change while preserving enough credible continuation to satisfy what the traveler has declared non-negotiable.

That is the difference between predicting the future and remaining able to answer it.

An itinerary tries, necessarily, to make some future decisions in advance. Good planning should. Commitment creates coordination. Reservations secure scarce goods. Decisions reduce cognitive load. The critique is not a rejection of plans.

It is a rejection of the idea that the plan retains authority after the state that justified it has materially changed.

ReturnVector begins there.

Not as a machine that knows which future will occur.

Not as an optimizer entitled to decide what a traveler should value.

Not as a score generator whose decimals substitute for calibration.

Its legitimate ambition is narrower: maintain a disciplined representation of what current evidence says remains possible, what a contemplated action will close, which alternatives are more correlated than they appear, where hard boundaries lie, where predictions remain weak, and whether a route exists for the person who must actually execute it.

The final decision remains human because the final value judgment remains human.

The architecture helps expose the price of that judgment.

The deepest problem in adaptive travel is not that the future is uncertain. No model needs to discover that.

It is that present actions alter which parts of that uncertain future we will remain capable of answering.

The future does not owe the traveler obedience.

Rational planning begins when the plan stops asking it to.

---

Endnotes

1. Randolph W. Hall, “The Fastest Path through a Network with Random Time-Dependent Travel Times,” Transportation Science 20, no. 3 (1986): 182–188, https://doi.org/10.1287/trsc.20.3.182. Hall’s abstract explicitly distinguishes an optimal adaptive decision rule from a simple predetermined path.

2. Tarun Rambha, Stephen D. Boyles, and S. Travis Waller, “Adaptive Transit Routing in Stochastic Time-Dependent Networks,” Transportation Science 50, no. 3 (2016): 1043–1059, https://doi.org/10.1287/trsc.2015.0613.

3. He Huang and Song Gao, “Trajectory-Adaptive Routing in Dynamic Networks with Dependent Random Link Travel Times,” Transportation Science 52, no. 1 (2017): 102–117, https://doi.org/10.1287/trsc.2016.0691. INFORMS lists the article as published online February 27, 2017 and gives 2017 in its preferred citation, although volume 52, issue 1 is the January–February 2018 issue. The 2017 citation follows the publisher’s stated form.

4. Dimitris Bertsimas and Melvyn Sim, “The Price of Robustness,” Operations Research 52, no. 1 (2004): 35–53, https://doi.org/10.1287/opre.1030.0065.

5. Kenneth J. Arrow and Anthony C. Fisher, “Environmental Preservation, Uncertainty, and Irreversibility,” Quarterly Journal of Economics 88, no. 2 (1974): 312–319, https://doi.org/10.2307/1883074. The chapter borrows only the structure of irreversible commitment under uncertainty and future information, not an equivalence between environmental preservation and travel decisions.

6. Torsten Hägerstrand, “What About People in Regional Science?,” Papers in Regional Science 24, no. 1 (1970), https://doi.org/10.1111/j.1435-5597.1970.tb01464.x.

7. Karst T. Geurs and Bert van Wee, “Accessibility Evaluation of Land-Use and Transport Strategies: Review and Research Directions,” Journal of Transport Geography 12, no. 2 (2004): 127–140, https://doi.org/10.1016/j.jtrangeo.2003.10.005. Their review specifically identifies individual spatial-temporal constraints as an important component of more adequate accessibility measurement.

8. Jill L. Bezyak, Scott A. Sabella, and Robert H. Gattis, “Public Transportation: An Investigation of Barriers for People With Disabilities,” Journal of Disability Policy Studies 28, no. 1 (2017): 52–60, https://doi.org/10.1177/1044207317702070. The study reports an online survey of 4,161 respondents and continued physical and attitudinal barriers affecting public-transport and paratransit use.

9. Hans P. A. Van Dongen, Greg Maislin, Janet M. Mullington, and David F. Dinges, “The Cumulative Cost of Additional Wakefulness: Dose-Response Effects on Neurobehavioral Functions and Sleep Physiology from Chronic Sleep Restriction and Total Sleep Deprivation,” Sleep 26, no. 2 (2003): 117–126, https://doi.org/10.1093/sleep/26.2.117.

10. Jose Guillermo Cedeño Laurent et al., “Reduced Cognitive Function during a Heat Wave among Residents of Non-Air-Conditioned Buildings: An Observational Study of Young Adults in the Summer of 2016,” PLOS Medicine 15, no. 7 (2018): e1002605, https://doi.org/10.1371/journal.pmed.1002605.

11. Karen Lucas, “Transport and Social Exclusion: Where Are We Now?,” Transport Policy 20 (2012): 105–113, https://doi.org/10.1016/j.tranpol.2012.01.013. Lucas reviews the relationship among transport disadvantage, poverty, access to essential services and activities, and social exclusion.

12. Herbert A. Simon, “A Behavioral Model of Rational Choice,” Quarterly Journal of Economics 69, no. 1 (1955): 99–118, https://doi.org/10.2307/1884852.

13. Raja Parasuraman and Victor Riley, “Humans and Automation: Use, Misuse, Disuse, Abuse,” Human Factors 39, no. 2 (1997): 230–253, https://doi.org/10.1518/001872097778543886. Their review defines misuse in part through overreliance on automation and discusses its relationship to monitoring failures and decision bias.

14. Lisanne Bainbridge, “Ironies of Automation,” Automatica 19, no. 6 (1983): 775–779, https://doi.org/10.1016/0005-1098(83)90046-8.

15. Tilmann Gneiting, Fadoua Balabdaoui, and Adrian E. Raftery, “Probabilistic Forecasts, Calibration and Sharpness,” Journal of the Royal Statistical Society: Series B 69, no. 2 (2007): 243–268, https://doi.org/10.1111/j.1467-9868.2007.00587.x.

16. Tilmann Gneiting and Adrian E. Raftery, “Strictly Proper Scoring Rules, Prediction, and Estimation,” Journal of the American Statistical Association 102, no. 477 (2007): 359–378, https://doi.org/10.1198/016214506000001437.

Bibliography

Arrow, Kenneth J., and Anthony C. Fisher. “Environmental Preservation, Uncertainty, and Irreversibility.” Quarterly Journal of Economics 88, no. 2 (1974): 312–319.

Bainbridge, Lisanne. “Ironies of Automation.” Automatica 19, no. 6 (1983): 775–779.

Bertsimas, Dimitris, and Melvyn Sim. “The Price of Robustness.” Operations Research 52, no. 1 (2004): 35–53.

Bezyak, Jill L., Scott A. Sabella, and Robert H. Gattis. “Public Transportation: An Investigation of Barriers for People With Disabilities.” Journal of Disability Policy Studies 28, no. 1 (2017): 52–60.

Cedeño Laurent, Jose Guillermo, Augusta Williams, Youssef Oulhote, Antonella Zanobetti, Joseph G. Allen, and John D. Spengler. “Reduced Cognitive Function during a Heat Wave among Residents of Non-Air-Conditioned Buildings: An Observational Study of Young Adults in the Summer of 2016.” PLOS Medicine 15, no. 7 (2018): e1002605.

Geurs, Karst T., and Bert van Wee. “Accessibility Evaluation of Land-Use and Transport Strategies: Review and Research Directions.” Journal of Transport Geography 12, no. 2 (2004): 127–140.

Gneiting, Tilmann, Fadoua Balabdaoui, and Adrian E. Raftery. “Probabilistic Forecasts, Calibration and Sharpness.” Journal of the Royal Statistical Society: Series B 69, no. 2 (2007): 243–268.

Gneiting, Tilmann, and Adrian E. Raftery. “Strictly Proper Scoring Rules, Prediction, and Estimation.” Journal of the American Statistical Association 102, no. 477 (2007): 359–378.

Hägerstrand, Torsten. “What About People in Regional Science?” Papers in Regional Science 24, no. 1 (1970).

Hall, Randolph W. “The Fastest Path through a Network with Random Time-Dependent Travel Times.” Transportation Science 20, no. 3 (1986): 182–188.

Huang, He, and Song Gao. “Trajectory-Adaptive Routing in Dynamic Networks with Dependent Random Link Travel Times.” Transportation Science 52, no. 1 (2017): 102–117.

Lucas, Karen. “Transport and Social Exclusion: Where Are We Now?” Transport Policy 20 (2012): 105–113.

Parasuraman, Raja, and Victor Riley. “Humans and Automation: Use, Misuse, Disuse, Abuse.” Human Factors 39, no. 2 (1997): 230–253.

Rambha, Tarun, Stephen D. Boyles, and S. Travis Waller. “Adaptive Transit Routing in Stochastic Time-Dependent Networks.” Transportation Science 50, no. 3 (2016): 1043–1059.

Simon, Herbert A. “A Behavioral Model of Rational Choice.” Quarterly Journal of Economics 69, no. 1 (1955): 99–118.

Van Dongen, Hans P. A., Greg Maislin, Janet M. Mullington, and David F. Dinges. “The Cumulative Cost of Additional Wakefulness: Dose-Response Effects on Neurobehavioral Functions and Sleep Physiology from Chronic Sleep Restriction and Total Sleep Deprivation.” Sleep 26, no. 2 (2003): 117–126.