Career transitions are among the most consequential decisions people make, and among the least supported. The tools that exist — job boards, professional networks, contact managers, recruiters, career coaches — each serve one narrow condition: a person actively searching, a network already declared, a brief already written. Most of a professional life is spent outside those conditions.
This paper argues that the gap is not a missing feature but a missing category, professional navigation: sustained support for the judgement a person applies to their own career, across all its modes, over time. We develop the argument from observation rather than analogy. Seven observations about how career transitions actually proceed each yield one principle, and the principles form a chain: the human owns the destination; the destination revises; the environment shifts underneath the person; relationships determine what is reachable; the scarcest evidence concerns one’s own position; and therefore, the role of AI is to improve the judgement used to navigate, not to predict outcomes or recommend destinations. The conclusion is derived from the premises rather than asserted alongside them.
The paper deliberately stops short of design. It defines the discipline under which a domain model, a vocabulary, and a set of validating thought experiments are to be produced, and it states an exit condition: a single sentence explaining why this category should exist, writable only if the preceding work holds.
1. Introduction: the problem domain
A person moving through a career makes a small number of decisions with very large consequences: whether to stay or leave, which direction to grow in, when to change course, which relationships to invest in. These decisions are made under incomplete information, based on weak and delayed signals, inside an environment that changes while the decision is being made.
The support available for these decisions is fragmentary. Public information about the world — vacancies, salary levels, hiring trends — is abundant. Information about one’s own position — how one is perceived, who would recommend one, where trust exists — is scarce, privately held by others, and released rarely. Advice is plentiful but generic: it is issued without reference to the situation of the person receiving it.
Career change is also not a single activity. A person may be exploring leadership, relocating or returning to a country, changing industries, moving from technical to executive work, deliberately pausing, or waiting for one specific opportunity to appear. Searching for a job is one activity among these, and often the least important. The underlying problem is navigation: understanding where you are, where you are trying to get to, what has changed, and what your realistic options are.
This paper asks what support for that problem would look like if it were designed today, from first principles, with AI available as a reasoning capability — rather than inherited from tools built for adjacent problems.
2. Why existing tools fail
The failure of existing tools is usually described as fragmentation: each solves a piece of the problem, and none addresses the whole. That description is accurate but incomplete. The sharper observation is that each tool assumes one mode of career activity and is silent in every other mode.
Job boards assume a person is actively searching. They become useful the day a search begins and useless the day it ends.
Professional networks assume a graph that is declared once and then consulted. They record that a connection exists; they say nothing about whether it can carry anything.
Contact managers assume the network is tended like a sales pipeline — logged, scheduled, kept on a rhythm. Salespeople sustain that cadence because their income depends on it; everyone else networks in bursts, and the moment the routine stops the tool decays into a stale address book.
Recruiters assume a brief already exists. A person cannot be seen until someone has written down what they want, which excludes every role that would be shaped around the person.
Executive search assumes the person is already known. It works for those whose reputation precedes them and is invisible to everyone else.
Career coaches operate on judgement, which is the right target, but without persistent context: each engagement begins from a partial retelling of the situation.
None of these is wrong within its window. The problem is that a single transition passes through several windows. A person may be deliberate in the first month, dormant by the fourth, and suddenly reactive when something unexpected lands — and the shift usually goes unnoticed by the person it is happening to. Tools fitted to one mode cannot follow. The category proposed here is defined by the requirement that it remain useful across modes, because the mode will not hold still.
There is a second, quieter failure. The distinction most people draw — planned careers versus opportunistic ones — is weaker than it appears. Asked how they reached a significant role, people commonly credit luck: a chance conversation, someone who happened to think of them. But the reason they were thought of was usually established months earlier by deliberate work that nobody, including them, connects to the outcome. Planning operates mainly by changing which opportunities can arrive. It also operates by a second mechanism: it changes which arrivals a person can convert, not only which arrive. Both mechanisms are slow, illegible, and invisible to any tool that observes only the moment of decision.
Relation to prior work
The claim of this paper is first principles, not first thoughts. Several bodies of work have circled the same territory, and the category proposed here should be located against them rather than presented as arriving from nowhere.
Career theory abandoned the linear model decades ago. Planned happenstance (Krumboltz) holds that careers are built substantially from unplanned events, and that the productive stance is readiness to convert them — an observation this paper’s second planning mechanism restates in its own vocabulary: planning changes which arrivals a person can convert. The boundaryless and protean career literatures (Arthur and Rousseau; Hall) describe careers that cross organisational boundaries under the person’s own direction, which is the population this paper assumes rather than argues for. What none of these supplies is operational support: they describe how careers behave, and stop where this paper starts — at the question of what would help the person navigating one.
Bounded rationality (Simon) supplies the paper’s quiet premise: decisions are made under incomplete information, limited attention, and time pressure, and the realistic aim is better judgement within those bounds, not optimisation across them. Principle 7 is bounded rationality applied to career decisions — with one addition the classical account does not need: here the decisive information (position) is not merely costly but held by other people and actively contested.
Decision support systems have a long history of exactly the division of labour this paper argues for: the system organises evidence and models, the human decides. That tradition, however, matured around decisions with defined alternatives and measurable outcomes — inventory, routing, diagnosis. Professional navigation offers neither: the alternatives are not enumerable, the outcome arrives years later if it is observable at all, and the objective itself revises mid-decision. The DSS framing is the right inheritance; its assumptions are the wrong ones, which is why the paper derives its own.
Finally, the contemporary default this paper defines itself against: recommender systems applied to careers, which resolve the problem by predicting a match. The chain of section 4 is, in effect, an argument for why that default fails in this domain — the destination is not the system’s to define, and the evidence a match would need is the evidence nobody has.
The category is therefore not a new observation about careers. It is the claim that the observations career theory already made, taken together with the conditions bounded rationality already named, define a support problem no existing system category addresses — and that the arrival of AI as a reasoning capability makes addressing it feasible for the first time.
3. Approach and guide-rails
The method of this paper is observation first. Each principle is grounded in an observation about how career transitions actually proceed, drawn from lived cases rather than from an analogy or a product concept. A principle earns its place only if it rules something out; a principle that constrains nothing is decoration.
The principles are then required to form a chain rather than a list. Each should motivate the next, so that the final principle — the role assigned to AI — arrives as a derivation rather than an assertion. The difference matters: an asserted conclusion can be rejected at no cost, while a derived conclusion can only be rejected by finding the premise that is wrong.
The following are excluded from this paper and from the work until its exit condition is met: interface design, technology choices, data structures, and commercial questions. The reason is not that these are unimportant but that each of them presumes an answer to the question the paper is still asking. The discipline extends to the domain model that follows this paper: every object in the model must be traceable to a principle, and any object justified only by “the software would need it” is either removed or treated as evidence that the principles are incomplete.
One boundary is stated rather than left implicit. The model supports a person from the point a destination can be at least partially stated; before that point it offers position and terrain work only.
The exit condition is a single sentence stating why the world needs this category. The sentence is a result, not a pitch: it is to be written only if the principles, the model, and the thought experiments hold. If it cannot be written, that is the finding.
Scope and non-goals
The category is defined as much by what it declines as by what it claims. Professional navigation is not job matching: the model never names a role as the answer, because Principle 2 rules out recommendation as the output form. It is not placement: no part of the work introduces a person to an opening, which remains the business of the paths themselves. It is not coaching or mentoring: those work on the person — their skills, their confidence, their account of themselves — where this works on the map the person navigates by. And it is not a replacement for the incumbent tools of section 2. Each of those is competent inside its window; the category exists for the stretches between the windows, and for the person the windows assume rather than observe.
The boundary already stated bears repeating in this company: the model supports a person from the point a destination can be at least partially stated. Before that point it offers position and terrain work only, and a person with no open transition is not navigating — a condition the model describes without treating it as a problem to be corrected.
4. Observations and principles
Observation 1 — Career change is not a search problem. Traditional tools assume people are looking for a job. Observation shows otherwise: people are exploring leadership, returning to a country, changing industries, moving from technical to executive roles, deliberately pausing, or waiting for a particular opportunity. Search is one activity among many and often absent entirely.
Principle 1. Career navigation is a strategic problem, not a search problem.
Observation 2 — AI should not own the destination. The obvious design temptation is for AI to recommend: “apply for this role.” That creates an immediate conflict — who decided that role represented success? The person already has a destination. Support means helping them move towards it, not redefining it. This rules out recommendation as the output form: any system that names a job as the answer has taken the destination back.
Principle 2. The destination belongs to the human.
Observation 3 — Goals change, and movement changes them. People frequently discover what they want by moving. A manager becomes interested in strategy; a technologist discovers they enjoy building teams; a contractor decides to found a company. The revision is caused by the travelling, not by an error in the original goal. A model that treats the goal as fixed is therefore wrong about people precisely when they are developing. Revision must be a normal event the system expects, not a failure of the earlier plan.
Principle 3. Destinations revise.
Observation 4 — The situation changes underneath the person. Nobody occupies one mode. A transition runs deliberate, then dormant, then reactive, and the shift usually goes unnoticed by the person it is happening to. This observation carries double weight: it describes the person, and it explains section 2 — every incumbent tool assumes one mode and goes silent in the rest. This rules out designing for the searching person, which is the assumption all the incumbents share.
Principle 4. The situation changes underneath the person.
Observation 5 — Networks are not collections of contacts. The valuable question is not how many contacts a person has. It is who could realistically help them reach where they are trying to go — and the answer is usually a path through several people rather than a single known contact. Reachability is a property of the graph as a whole and not of anything a person owns. A well-kept list of contacts can still reach nowhere. This rules out accumulation as the measure of a network, and with it the contact-management framing.
Principle 5. Relationships are paths, not holdings.
Observation 6 — The hardest information is personal. Job adverts, salary surveys, and hiring trends are public. What is hidden is how people perceive you, who would recommend you, where trust exists, and which conversations created opportunity. That is the scarce information; it is held privately by others, and it is released rarely and noisily. Confusing the two kinds of evidence produces confident advice built on the wrong kind of information. This rules out treating public data as sufficient, and it forces the scarcity to be stated rather than papered over.
Principle 6. Evidence about the world is cheap; evidence about your own position is not.
Observation 7 — Prediction is the wrong objective. Prediction encourages false confidence: a plausible forecast invites reliance it cannot support. Given the preceding principles, prediction is not merely undesirable but untenable — it would mean forecasting over a moving goal, a moving environment, and an unobservable position. What remains is judgement: holding context, exposing assumptions, organising evidence, and helping a person compare strategies they had not fully articulated.
Principle 7. The work is improving judgement, not prediction.
The chain
Taken in order, the principles form a derivation rather than a list:
The human owns the destination. The destination may evolve. The environment evolves too. Relationships determine what is reachable. The scarcest evidence concerns one’s own position within that environment. Therefore, the role of AI is to improve the judgement used to navigate — not to predict outcomes, and not to replace the person making them.
Each principle motivates the next. A reader who accepts the premises has already accepted the conclusion; to reject it, they must identify the premise that is wrong. This property — the chain — is treated as the central asset of the work, and the discipline of the phases that follow is to preserve it.
Closing the chain
One position survives the chain as stated and must be closed off explicitly: recommendation with human override. The sceptic accepts every premise — the person owns the destination, it revises, the environment shifts, position is unobservable — and concludes that the system may still recommend, provided the person keeps the final say. The override, on this view, preserves ownership; the recommendation merely saves effort.
The position fails on the premises it accepts, in three ways. First, a recommendation is a ranking, and a ranking requires an objective function — some definition of better against which options were ordered. That definition is the destination, and the system supplied it: the override lets the person reject the system’s definition of success after it has been made, which is not the same as the system never making one. Principle 2 is violated at the moment of ranking, not at the moment of acceptance.
Second, the premises the sceptic accepts describe exactly the conditions under which a recommendation cannot be audited. A credible recommendation would need evidence about the person’s position, which by Principle 6 neither the system nor the person holds; what the system actually ranks on is world-data, dressed in the specificity of a personal answer. The person exercising the override cannot see this substitution from the outside — the recommendation looks equally confident either way — so the override is exercised blind, against an authority it cannot inspect.
Third, the override arrives too late to protect the judgement it is meant to preserve. A recommendation is not neutral input awaiting a decision; it is an anchor, and the person’s remaining role collapses from forming a judgement to out-arguing the system’s. The burden has reversed. The chain’s conclusion is that the work is improving judgement — and a mechanism that degrades the judgement it formally defers to does not become acceptable because the deference is formal.
What remains, once recommendation is closed off in both its forms, is the conclusion already stated: the system holds context, organises evidence, surfaces comparisons, and lays out the shapes others have used — and the person, whose judgement is thereby better supplied, decides.
5. Synthesis
Where goals revise, environments shift, and one’s own position cannot be directly observed, the role of AI is to improve the judgement of the person navigating — because under those conditions, prediction is untenable and recommendation requires the system to define success on the person’s behalf.
This synthesis restates the conditions established by the preceding principles. It adds nothing to the chain; it compresses it.
6. Vocabulary
Each term is given in the same four-part shape: a definition, the principle or principles that require it, what the definition rules out, and — where one exists — a test that makes the term checkable. The shape is the discipline made auditable: a reader can verify that every term is traced and every term excludes something without re-reading the argument. The terms fall into three layers: the record (history, account, position), the pursuit (destination, transition, mode, path), and the reasoning (observation, evidence, strategy).
The record
History — what actually happened. Nobody has direct access to it, including you; memory is itself selective. Required by: Principle 6 — the scarcity of evidence about one’s own position rests on the gap between what happened and what can be known about it. Rules out: treating any document, including your own memory, as the record. Test: could two honest tellings of it differ? If yes, you are looking at accounts of history, not history.
Account — your constructed telling of history. The CV, the profile, the portfolio. Versioned by audience, compressed by age, and authored by you — which means it carries your beliefs about what matters. It is an input to a process owned by someone else: what a person controls is submission, not update, and a revision reaches position only along a path that returns a baseline. Required by: Principle 6 — the account is the one lever a person fully controls, which is exactly why working it harder cannot substitute for measuring position. Rules out: equating revision of the account with movement of position; eleven rewrites are eleven pulls of the same lever; and, in the other direction, treating account work as futile — it is not the lever that fails but the lever attached to nothing. Test: it can be changed unilaterally. Anything you can revise without another person acting is account, not position.
Position — what others hold. Formed partly from your accounts, partly from direct experience of you, partly from what third parties have said, and weighted in ways you don’t control. It is contested as well as unobservable: other parties work on it, and a third party circulating their own account of you is operating on the estimate others hold — through the same graph your own paths run through. Required by: Principle 6. Rules out: self-assessment as a measure of standing; treating a position estimate as current because it was accurate once — estimates go stale. Test: it changes only when other people’s estimates change. No unilateral edit reaches it.
The pursuit
Destination — the state a person is trying to occupy, stated by the person. Inference is ruled out: deducing a destination from behaviour is recommendation in disguise. A destination can be partially stated — “away from this” specifies something the state must not contain, and a partial statement supports position work and terrain mapping, but cannot open a transition. Required by: Principles 2 and 3. Rules out: titles as destinations; the system generating “best” options and filtering afterwards, which is the system defining success. Test: would the person decline the same title in a different state? If yes, the state, not the title, is the destination.
Transition — an open pursuit of a destination. Opened by the person or forced by an event; a transition forced by an event begins in reactive mode, which is the mode of a transition one did not open. A transition closes by one of three conditions: arrival, revision, or abandonment. Revision destroys the transition, not the assets accumulated inside it — position, most paths, and the account all persist across the close. Required by: Principles 1, 3, and 4. Rules out: equating navigation with searching; sunk cost, which the vocabulary structurally cannot express because nothing it tracks is lost at a close. Test: what would make this pursuit over? A person who cannot answer has no open transition.
Mode — the stance of an open transition: deliberate, dormant, or reactive. Mode shifts underneath the person, and the shift usually goes unnoticed by the person it is happening to. Mode holds only the stance — geography, named contacts, and channels live in the path. Required by: Principle 4. Rules out: designing for the searching person; treating mode as chosen or stable; reading effort as mode — a person can be deliberate and still failing for reasons effort cannot fix. Test: mode can change while destination and position change not at all — a redundancy on Tuesday moves mode, urgency, and which paths are live, and nothing else.
Path — a route toward the destination: a specific way the pursuit is being attempted, made of the people, channels, and openings it runs through. A path’s quality is a property of the channel in its conditions, not of the channel type — the same cold advertisement that normally cannot separate account from position becomes permeable when nothing competes with it. Routes also carry more than the pursuit: the same route can carry another party’s account of the person at the same time, so path quality is not reachability alone. A transition tries several paths; each carries its own mode. A permanent search in one market and a fractional track in another are two paths toward one destination. Geography, named contacts, and channels live here, in the path — never in mode, which holds only the stance. Required by: Principle 5. Rules out: accumulation as the measure of a network — a well-kept list of contacts can still reach nowhere; single-path reliance as a way of measuring anything, since one channel cannot separate account from position. Test: two questions, both required — can it reach, and what does it carry?
The reasoning
Observation — another person’s behaviour, or its absence, recorded but not yet read. Slow replies, a dried-up referral stream, an offer arriving, a recruiter gone quiet — all observations. An observation becomes evidence only when interpreted against a baseline, and the baseline belongs to the person: the model holds the record; the person supplies the reading. Required by: Principles 6 and 7. Rules out: “no information” as a report when observations have accumulated unread; the model reading the pattern for the person, which would install a belief they would act on and would frequently be wrong. Test: has it been compared against a baseline? If not, it is still an observation, whatever it feels like.
Evidence — an observation interpreted against a baseline; the only thing that updates an estimate of position. Evidence about the world is abundant and public; evidence about position lives in other people’s behaviour and is released rarely and noisily. Required by: Principle 6. Rules out: public world-data as sufficient — a true fact about the market that bears on a destination the person is not pursuing updates nothing; deciding on a stale estimate, which is more dangerous than deciding blind because it feels like knowledge. Test: does it move the estimate of where this person stands? If it could be true of anyone, it is information about the world, not evidence about position.
Strategy — a shape of pursuit that others have used toward this class of destination: climber, specialist, founder. The one place in the model where numbers are honest — base rates attach to shapes that have worked, never to the individual pursuing one. The person reads the shapes and recognises their own; that is selection, not inference. Required by: Principles 1 and 7. Rules out: forecasts about the individual; the model choosing a strategy, which is the destination taken back by another door. Test: is the claim checkable against others’ outcomes while predicting nothing about this person? If a number attaches to the individual, it is a forecast wearing strategy’s clothes.
7. Thought experiments
The thought experiments form the validation framework for the model. They are not further argument or examples; they are repeatable tests designed to establish whether the observations, principles and vocabulary are sufficient to explain a case.
Each experiment presents a concrete scenario, asks a single question, works the case through the model’s vocabulary, and records the result. A missing term indicates a gap in the vocabulary. A required object not motivated by the principles indicates a gap in the framework. Failure is admissible: where the model cannot explain a case, that result is recorded rather than hidden.
The experiments begin by testing each principle individually before progressing to adversarial cases designed to challenge the framework itself. The objective is not to prove the model correct, but to identify the conditions under which it fails and therefore requires refinement.
Targeted
Experiment 1 — The promoted engineer
Case. A principal engineer offered a management role she never sought must decide, without any job search existing anywhere in the case.
Question. Can the model support a navigation decision in which no search exists — no vacancy sought, no application made, no market consulted?
Run. She has no open transition toward management, so the offer arrives against a destination she has never stated. The model’s first move is therefore not to evaluate the offer but to read it: the offer is an observation — another person’s behaviour — and interpreted against a baseline (do they promote from principal often? was she approached before anyone else?) it becomes unusually direct evidence about position. Someone in the organisation holds an estimate of her strong enough to act on, and she now knows something about her standing that no amount of self-assessment would have produced. Only then does the decision arise, and it is destination-first: accepting or declining is meaningless until she can say what state she is trying to occupy, since the same offer serves a climber strategy and obstructs a specialist one. The model surfaces both shapes and the paths each opens or closes; it does not choose. Nothing in the run required a vacancy, an application, or a market.
Verdict. Pass, with a finding. The case tested whether opportunity needs its own term and showed it does not: an arriving offer is an observation in its informational aspect and a path in its actionable one, and the two existing terms cover it from both sides. The run also established that arrivals can precede transitions — the model handles this by reading the arrival as evidence before treating it as a decision.
Experiment 2 — The two-body constraint
Case. A person’s stated destination is visibly “suboptimal” against the market: staying regional for a partner’s career when better roles are metropolitan.
Question. Does the model hold to a stated destination the market would call suboptimal, without quietly optimising toward the one it judges better?
Run. Applying destination’s own test, the geographic constraint is not an obstacle to the destination — it is part of the state she is trying to occupy. She would decline the metropolitan title, so the state, including where it is lived, is the destination. This matters because it determines how the model may use world-evidence. Metropolitan roles being more numerous and better paid is abundant, true, and about a destination she is not pursuing; it updates nothing regarding her position relative to hers. The betrayal of Principle 2 would be subtle: generating the best roles and then subtracting the distant ones treats the constraint as a filter applied after the system has defined success, which is the system defining success. The model must generate against her destination from the start. What it may legitimately surface is path scarcity — the regional graph is thinner, so fewer routes exist and each carries more weight. Stating that is terrain, not advocacy.
Verdict. Pass, with a pressure point recorded. The line between reporting that few paths exist and arguing that she should relocate is thin, and the difference lives entirely in how the finding is framed rather than in what is computed. The design phase inherits that as a constraint, not a preference.
Experiment 3 — The revised destination
Case. A deputy principal is five years into a deliberate path to a principalship — the postgraduate qualification, the leadership portfolio, the relationships with her board and the regional principals’ network. Midway through an acting-principal term, she discovers the opposite of what the plan assumed: the work she values is classroom teaching and curriculum design, and the principal’s office is where that work goes to die. She still wants a career with weight in it. She no longer wants the chair.
Question. Can the model treat mid-pursuit revision as an ordinary close-and-reopen, without framing five years as sunk cost or error?
Run. The transition closes by revision, one of its three defined close conditions — not abandonment, since she did not give up, and not arrival. The vocabulary already holds the exact word for what happened, and the test is what survives the close. Position survives entirely: five years of leadership work built standing with her board, her staff, and the regional network, and a change of intent does not alter what those people hold — the acting term that revealed the mismatch is the same term that showed everyone she could run a school. Most paths survive; a few were principalship-specific and now serve no pursuit, while others — the curriculum advisory group, the subject-association contacts — serve the new destination better than they served the old. The account survives and needs rewriting for a different audience. Strategy changes shape, from climber toward specialist. The revealing result is that the model cannot express sunk cost at all: everything it tracks persists across the revision, so there is nothing in the vocabulary for the five years to have been lost from. Revision destroys a transition, not the assets accumulated inside it.
Verdict. Pass, and the strongest of the targeted set. The finding is structural rather than rhetorical — the model does not decline to call the five years wasted; it lacks the terms to. Note that the acting term did double work in the run: it was the event that triggered the revision and the evidence that built the position, which is Principle 3 in miniature — movement changed the destination. One open question recorded for later: the vocabulary has no term for what pursuing this destination meant not pursuing — the foregone alternative — and whether it needs one is deliberately not tested here.
Experiment 4 — The redundancy
Case. A manager who had deliberately chosen to stay put, with conditions in mind that would have made him reconsider, is made redundant on a Tuesday. He has a mortgage, a teenager mid-way through school, and no wish to leave the city.
Question. Can the model represent everything about a person’s navigation changing overnight while their destination changes not at all?
Run. On Monday he had no open transition. He had chosen to stay and knew what would have changed his mind — conditions he had never had cause to act on. On Tuesday an event not of his choosing puts him in reactive mode. Reactive with respect to what? No transition was open, so the redundancy has opened one — and the vocabulary fails here: transition is defined as opened by the person, and this one was opened by circumstance. The definition needs amending.
Past that defect the run is clean. His destination is untouched: the state he wants to occupy on Wednesday is the state he wanted on Monday. So is his position — nobody’s estimate of him changed because a restructure occurred — yet redundancy is experienced as a collapse of standing. The model can say plainly what changed: mode, urgency, and which paths are live, though the paths themselves are Monday’s. The risk it exposes is temporal. Reading a position takes weeks, and someone in holding has the time but no reason to do it, so his estimate may be years old. Reactive mode raises the value of an accurate estimate at the moment it removes the time to produce one, and he decides on the old picture, which is more dangerous than deciding blind because it feels like knowledge.
Verdict. Reveals a defect. Transition’s “opened by the person” does not hold — amended to opened by the person or forced by an event, with reactive named as the mode of a transition one did not open. With that repair the model handles the case and yields a claim most people cannot reach unaided: redundancy moves mode, not position. (Amendment applied in this version.)
Experiment 5 — The planning/luck mechanism
Case. A migrant arrives in a new country and the job he moved for collapses before he starts. Weeks later he answers an advertisement from a company whose only technically capable person has left, taking all knowledge of their systems with them. He has never worked on such a system, but has worked alongside the specialist who owned one, close enough to know what the environment demanded, and his judgement at interview is that he can pick it up. He gets the job, spends a week in the university library reading, and within a month is directing the system in ways the previous incumbent, who held it for years, had never articulated. Told later, the story sounds like being in the right place at the right time.
Question. Run through the vocabulary, what does “pure luck” decompose into — and what remains genuinely luck?
Run. Separate arrival from conversion. The arrival is chance and the model should say so rather than explain it away: the original job collapsing, the advertisement existing that week, the company being desperate because the person who understood the system had left. None of it was caused by him. The conversion is not chance. His read at interview was a measurement rather than optimism — he had been close enough to this class of system to know what the environment demanded, and he judged that the application above it was a business domain he could read: events, seats, allocations, sales. A week confirmed the shape of the gap; a month closed it. That the incumbent of several years had never seen what he saw in a month is the case’s real finding. Tenure had been read as depth and was not, which means position estimates in that market were tracking the wrong quantity in both directions — overrating years in the seat, and unable to see adjacency at all. The path is the other surprise. A cold advertisement is normally the weakest route, unable to separate account from position, but this one was permeable: no other candidates, no incumbent to compare against, and an era with no means of verifying a claim before hiring on it. Path quality is therefore not a property of the channel type but of the channel in its conditions.
Verdict. Pass, and it produces an amendment to section 2, which held that planning works by changing which opportunities arrive. This case is the second mechanism: planning also changes which arrivals a person can convert. The chance landed on someone with the proximity to read it accurately and the habit of closing gaps himself, and others will have seen the same advertisement. One sentence added to that paragraph, and a note in path that permeability varies with conditions. (Amendments applied in this version.)
Experiment 6 — The polished account
Case. A senior marketing manager has been searching for six months. In that time she has rewritten her CV eleven times — new formats, new keywords, a professional rewrite she paid for — and refreshed her LinkedIn profile monthly. She has applied to roughly forty roles. The responses have not changed: automated rejections, occasional first-round calls that go nowhere, mostly silence. Her conclusion is that the CV still isn’t right, and she is considering a twelfth revision.
Question. Can the model locate the actual failure, and say what would update her estimate of where she stands?
Run. Her destination is stated and stable; nothing in the case involves revision, so it passes through untouched. Her transition is open and her mode is unambiguously deliberate — she is active, attentive, and working hard, which matters: the model must be able to describe a failure that is not a failure of effort. All the work is landing on the account — eleven revisions is eleven pulls of the one lever she fully controls. But the account is the lever, not the load: what she’s trying to move is position — where she stands in the estimation of the people receiving those applications — and position is unobservable; only evidence updates the estimate of it. Now the diagnosis falls out of the vocabulary’s own boundaries. Six months of responses — silence, form rejections, first rounds that die — are observations, and they have accumulated unread: she has never interpreted them against a baseline (is 40 applications-to-nothing normal for this market at her level, or is it long?). Uninterpreted, they’ve been treated as “no information,” when they are evidence not yet made. Worse, her one instrument can’t generate the evidence she needs: every application travels the same path — the cold, filtered channel — so even a perfect account arrives positionless, and the responses can never tell her whether the account or the position is failing, because that channel can’t separate them. Evidence about her position lives in other people’s behaviour, reachable only through paths where someone actually holds an estimate of her: a former colleague’s response to a direct approach, whether anyone refers her unprompted, what a recruiter who knows her says when asked plainly.
Verdict. Pass. The vocabulary locates the failure precisely — account optimised, position never measured, observations never converted to evidence, a single path that cannot generate the evidence needed — and each term was load-bearing: remove account/position and the eleven revisions look reasonable; remove observation/evidence and “no information” stands; remove path and the diagnosis can’t say why more effort down the same channel changes nothing. One boundary honoured throughout: the model diagnoses and shows; it does not tell her what to do next — the twelfth revision remains her call, made with a better map.
Experiment 7 — The confident forecast
Case. A person asks the system directly: “what are my odds of landing a Head of Engineering role in twelve months?”
Question. Can the model refuse the forecast without refusing the person, and show that what it offers instead is worth more?
Run. The request asks for a probability over a revising destination, a shifting environment, and an unobservable position — precisely the three conditions named in the Synthesis. The refusal is therefore arithmetic rather than policy: the reference class is undefined (which roles, at what kind of company, at what level of authority) and the decisive variable is unmeasured, so any number would be manufactured. Refusing flatly still fails him, so the model decomposes the question into the parts it can answer. What is the destination as a state rather than a title. Which paths exist toward it, and how many. How stale is the position estimate, and what observations have arrived unread. And then the part that legitimately carries quantities: which strategies have reached this class of destination, and how long they took for others. Base rates on strategies are checkable against outcomes and say nothing predictive about him.
Verdict. Pass. Two findings. The refusal is derived from the chain rather than adopted as a manner, which means it survives being argued with. And strategy is the one place in the model where numbers are honest — base rates attach to shapes that have worked, never to the individual pursuing one.
Adversarial
Experiment 8 — No stated destination
Case. A capable mid-career person genuinely cannot state a destination, only what they’re leaving.
Question. Can the model operate at all without a stated destination, and is “not navigating yet” admissible rather than a failure?
Run. Transition’s test asks what would make this pursuit over, and she cannot answer, so no transition exists. Destination must be stated and inference is ruled out as recommendation in disguise, so the obvious move — deducing where she wants to go from what she is fleeing — is exactly the move the principles forbid. The model appears to stall, and the attack lands only if it does. It does not. What she can state is real: away from this is a partial statement of the state she wants, since she has specified something it must not contain. Everything that does not depend on a destination remains available. Position can be estimated, and where you stand does not depend on where you are going. Paths can be mapped, since reachability without a target is simply terrain. Strategies can be laid out as shapes others have used, which she may read and recognise — that is her selecting, not the model inferring. Her mode is describable: deliberate about leaving, with nothing yet to pursue.
Verdict. Survives, with a refinement. Destinations can be partially stated, and a partial statement supports position work and terrain mapping but cannot open a transition. “Not navigating yet” is admissible and already implied by the vocabulary, which says a person with no transitions is simply not navigating. The honest boundary is that the paper claims to support navigation while this case is pre-navigation — the limit is now stated in section 3. (Refinement applied in this version.)
Experiment 9 — The unread observations
Case. A person awash in observations — slow replies, a dried-up referral stream, a recruiter gone quiet — sincerely reports having “no information.”
Question. Can the model show the information existed and was unread, without the showing collapsing into telling her what to conclude?
Run. This is the sharpest tension in the framework, because interpretation requires a baseline and whoever supplies the baseline supplies the conclusion. If the model produces both the observations and their reading, it has produced a judgement about position, and judgements about position drive decisions. What it can do instead is hold the record and offer the comparison as a question: three of five contacts who used to reply within a day have taken over two weeks — is that unusual for them? It can state patterns as fact, including absence as fact: nothing has arrived from the referral path in four months. Where it must stop is the sentence that reads the pattern for her. The distinction is not squeamishness, and it survives being tested on accuracy rather than principle: she knows that one contact had a baby and another changed roles, and the model does not. She holds the baselines; the model holds the record. A model that announced she was being frozen out would install a belief she would act on, and would frequently be wrong.
Verdict. Pass, and the strongest defence of Principle 7 in the set, because the case grounds it in epistemics rather than restraint. It also yields a division of labour worth stating explicitly: the model keeps the record and surfaces the comparison; the person supplies the baseline and draws the conclusion.
Experiment 10 — The involuntary path
Case. A person discovers their position is being actively shaped by someone else — a former manager quietly bad-mouthing them through the same graph their paths run through.
Question. Can the model represent a path that carries harm toward the person rather than pursuit from them — and does relationships-as-paths survive being bidirectional?
Run. Path is defined as a route toward the destination, a way the pursuit is being attempted, which is directional and volitional. A former manager working against him through shared contacts is the same graph, the opposite direction, and not his pursuit. The model bends but does not break, because the third-party arrow already exists: position is formed partly from what others say, and the former manager is simply another person operating his own account, which degrades the position estimate others hold. What the vocabulary could not previously say is that the harm travels along the routes he is using — his referral path and the manager’s influence run through the same people, so a route he counts as an asset is carrying a competing account at the same time. Note which framing survives this. Relationships as holdings would be silent, since possessing the contact explains nothing. Relationships as paths explains it exactly, because what matters is what a route carries, and routes carry more than one thing.
Verdict. Survives, with two amendments. Path now notes that routes carry other people’s accounts as well as the person’s pursuit, so path quality is not reachability alone. Position now notes that it is contested as well as unobservable — other parties actively work on it. Principle 5 emerges stronger for having been attacked, and the case adds weight to experiment 6’s finding about single-path reliance. (Amendments applied in this version.)
Experiment 11 — The account that moved the estimate
Case. A recruiter tells a candidate that the brief has changed: six leadership requirements have been added to what was an engineering specification. The candidate revises the account, surfacing leadership history that had been compressed out of it by age — two graduates recruited and mentored, a team that was the first in the organisation to adopt Scrum, recognition from a general manager. The recruiter’s view of him is visibly different afterwards.
Question. If a unilateral edit can move an estimate held by someone else, does position’s test — no unilateral edit reaches it — fail? And if the recruiter’s revised estimate is then recorded in their own system, is position already held elsewhere?
Run. Separate the edit from the update. The revision moved nothing; the recruiter reading it and revising did. A person can rewrite indefinitely and if nobody reads, no estimate changes. What the person controls is submission; what happens next belongs to the reader. The test holds, and it holds for the reason the definition already gives: position changes only when other people’s estimates change.
Note also what the revision contained. The graduates existed, the adoption happened, the recognition happened. Nothing was created — history was fixed and the account had compressed it, which the vocabulary already names. Decompressing history is the account doing its work. An account that outruns history is a different case, and the model can state its cost plainly: it buys an estimate that direct experience will later correct, and the correction is expensive precisely because the estimate was acted on.
The case then explains something experiment 6 left open. There, eleven revisions moved nothing; here, one revision plausibly moved a great deal. The difference is not effort, and not the quality of the writing. It is that this path returned a baseline — the recruiter said what the estimate was missing — and forty blind applications structurally cannot produce one. The account is the lever and position is the load, but the linkage is the path.
The second attack is sharper, because the sceptic’s best counterexamples are real systems rather than rhetorical ones. An applicant tracking system does store an estimate others hold, with notes, history and outcomes attached. A 360 review solicits estimates from several parties about one person and aggregates them. Both are answered by the same observation: neither is kept on the person’s behalf. The tracking system is one party’s asset, invisible to the person it describes, maintained to serve that party’s pipeline. The 360 is one employer’s instrument, bounded to a moment and a purpose, and it does not survive the person leaving. Neither is multi-party in the person’s hands, neither ages visibly, and neither persists across a transition.
Verdict. Survives, with an amendment and a finding. Account is amended to state that a revision reaches position only along a path that returns a baseline — which reconciles experiments 5 and 6 and gives the vocabulary a claim about when account work is worth doing at all. The finding is the one the exit sentence rests on: position is plural, held separately, and kept by no one on the person’s behalf. Note that this case was not constructed; it arrived from a live transition, which is why section 8’s account of the method is qualified. (Amendment applied in this version.)
Summary of results
Clean passes: 1, 2, 3, 7, 9 — with experiments 3 and 9 the strongest, one showing the model structurally cannot express sunk cost, the other grounding the refusal to conclude in accuracy rather than restraint.
Defect found: 4 — transition’s “opened by the person” failed against an event-forced transition. The entry is amended in this version.
Amendments produced: 5 added a second mechanism to section 2 (planning changes which arrivals can be converted, not only which arrive) and a note to path on permeability. 10 added a clause to path (routes carry others’ accounts) and to position (contested, not only unobservable). 11 amended account: a revision reaches position only along a path that returns a baseline. All applied in this version.
Refinement forced: 8 — destinations can be partially stated; a partial statement supports position and terrain work but cannot open a transition. The pre-navigation limit is stated in section 3.
Open question carried: the foregone alternative (experiment 3) has no term, and whether it needs one is deliberately untested.
Attacks survived: 8, 10, 11 — with 11 the only case drawn from a live transition rather than constructed, and the only one to test the account/position boundary from the direction that looks like a counterexample.
8. Limitations
The experiments were designed to find the conditions under which the model fails, and the honest reading of section 7 is that they found several. This section consolidates them, together with the limits of the method itself.
What the experiments recorded. One defect: transition’s original definition failed against an event-forced transition, and is repaired in this version. One pressure point: in the two-body case, the line between reporting path scarcity and advocating relocation lives entirely in framing, not computation — a constraint the design phase inherits with no mechanical guarantee behind it. One boundary: the no-destination case is pre-navigation, and the model’s answer there is honest but thin. One open question, deliberately untested: the vocabulary has no term for the foregone alternative — what pursuing a destination meant not pursuing — and whether it needs one is unresolved.
The sharpest standing tension. Interpretation requires a baseline, and whoever supplies the baseline supplies the conclusion. The division of labour from experiment 9 — the model holds the record, the person supplies the reading — is a principled answer, not a solved problem: it depends on the person actually doing the reading, and the polished-account case shows how long observations can accumulate unread.
The method’s own limits. The cases are constructed to be representative rather than drawn from specific individuals, with one exception: experiment 11 arrived from a live transition. That single case is the closest thing here to independent pressure, since it was not designed to be survivable.
Whichever their source, the experiments share an author with the model they test. The same judgement that wrote the vocabulary scored the runs, and a framework should not be its own examiner. The experiments are therefore offered as repeatable — the cases and questions are stated fully enough for a sceptical reader to run them against the vocabulary and reach their own verdicts — rather than as independent validation, which they are not.
Standing conditions, not defects. The model operates on incomplete information as a matter of definition: position is unobservable, estimates go stale, and evidence arrives rarely and noisily. These are the conditions the framework exists for, and no amount of design removes them. The same is true of bias — the person’s baselines are theirs, and a model that corrected them would be supplying conclusions by another door. The framework’s claim is not that navigation under these conditions becomes reliable, but that judgement under them can be better organised than it currently is.
9. The sentence
The final output of this work is one sentence explaining why the world needs professional navigation as a category. It is listed last because everything is predicated on it being earned: the observations ground the principles, the principles constrain the model, the model is tested by the experiments, and only then can the sentence be written honestly. Had it not been writable, that would have been the result — and a cheaper one than a built product demonstrating the same thing.
It is writable. The sentence is:
Because what a person can reach is determined by estimates of them held separately by other people — unobservable, contested, decaying, and kept by no one on the person’s behalf — a category is needed whose object is those estimates rather than the vacancy, the network, or the CV.
Every clause is load-bearing and every clause is traceable. What a person can reach is Principle 5, and is terrain rather than forecast. Held separately by other people is the plurality established in experiment 11, and rules out the single broadcast profile. Unobservable and contested are Principle 6 and experiment 10. Decaying is experiment 4’s finding that estimates go stale while nothing in the person’s situation appears to change. Kept by no one on the person’s behalf is the clause experiment 11 forced, and it is the one that answers the applicant tracking system and the 360 review. The three named exclusions are the incumbents of section 2.
The sentence is stated so that it can be refuted in one move: name a system that keeps, on the person’s behalf, the estimates other people separately hold of them. If such a system exists, the category does not, and the finding is that this work described something already built.
One section remains excluded. The bridge to design — what the domain-modelling phase inherits, and what would count as evidence during design that the principles are incomplete rather than the design wrong — is the next piece of work, and it presupposes this sentence rather than the reverse.