Karim Eldefrawy

Cryptography, Cybersecurity, Privacy Computer Science

Co-founder & CTO of Confidencial.io
2017–early 2023: SRI (Internet Origin)
2011–2016: HRL (Part of IBM Research)
2006–2010: PhD@UC Irvine

Scientific curiosity

Article 19 of 19 · Draft

Do Not Restart at Every Boundary

Advice for graduate students and young researchers after the institutional pathway fragmented—how to cross academia, industry, public missions, laboratories, and company formation without discarding accumulated research state at every move.

A bright research landscape in which a young researcher carries one glowing thread through a university, an industrial laboratory, a national mission facility, a field test, and a small research company without dropping the notebooks, tools, collaborators, and evidence accumulated along the route.
Conceptual illustration. The institutions are complementary environments, not a prescribed résumé or a promise that every path is equally accessible.

One degree, several institutional destinations

The doctorate no longer feeds one dominant employment pipeline

Among 20,469 U.S. doctorate recipients reporting definite non-postdoc employment commitments for 2024, NCSES reports 40 percent in academia, 40 percent in industry or business, 8 percent in government, 7 percent in nonprofits, and 6 percent in other or unreported sectors. Rounded percentages sum to 101.

These are immediate commitments among respondents with definite U.S. employment plans, not lifetime career outcomes or a measure of research quality. Inspect the source record.

The compounding-path test

Cross a boundary only when it adds a missing capability

A coherent research life is not a sequence of logos. It is a durable question accumulating different kinds of evidence, judgment, infrastructure, and transition power.

  1. Continuity spine

    Keep one durable question

    Define the unresolved problem and the class of evidence that would change your mind. Methods, employers, and applications may change without dissolving the research identity.

    National Academies ...

  2. Deep apprenticeship

    Earn one hard depth first

    Stay long enough with demanding mentors and artifacts to learn how a field detects error, not merely how it presents successful results.

    National Academies ... National Academy of...

  3. Complement test

    Add what the current node cannot supply

    Cross for a real instrument, mission user, systems constraint, manufacturing process, deployment channel, or theoretical lens—not novelty alone.

    U.S. National Scien... U.S. Department of ...

  4. State transfer

    Carry a boundary packet

    Preserve hypotheses, failures, assumptions, code, data rights, collaborators, and next experiments so the new setting starts from accumulated state.

    Kyle Lewis, Donald ... Raghavendra D. Rao ...

  5. Transition option

    Make use testable

    Learn who can adopt, reject, maintain, procure, license, or productize the work. Customer discovery can improve a research question even when no startup should be formed.

    U.S. National Scien...

The thread can remain intellectually continuous even when employers, funding instruments, and application domains change. A boundary that adds none of these capabilities is probably a reset disguised as opportunity.

This article is for the graduate student who has understood the diagnosis in this series and reached an uncomfortable practical question:

If the old path is gone, what am I supposed to do now?

The wrong answer is to reproduce my résumé, or anyone else’s. Ericsson, Lucent Technologies’ Bell Labs, Cisco, a doctorate, AT&T Research, HRL, SRI, DARPA programs, and a research spinout formed one path through a particular historical moment. It contained luck, mentors, immigration and economic circumstances, technical choices, and institutional openings that cannot be reduced to a recipe.

The useful property of that path was not the list of names. It was accumulation without complete restart. Telecommunications supplied systems scale. Academic work supplied formal depth and permission to stay with a hard question. Corporate and mission laboratories supplied demanding users, engineering constraints, facilities, and long projects. Government programs connected distributed teams to explicit capability goals. A spinout made transition, financing, product choice, and organizational survival impossible to treat as somebody else’s problem. M

Each boundary changed the objective function. The important question is whether it discarded the state already accumulated or made that state more capable.

That distinction leads to the central advice:

Do not build a career as a sequence of affiliations. Build a research question that becomes more testable, more integrated, and more useful at every boundary.

The developmental burden moved onto the researcher

An integrated laboratory could once place theorists, experimentalists, systems builders, technicians, product engineers, standards experts, customers, and experienced research managers within one durable institutional graph. No individual needed to become all of them. A young researcher could develop breadth by working near complementary experts while retaining a stable employer, laboratory memory, and technical community.

When that graph fragments, the need for integration does not disappear. Its cost moves.

The graduate student now has to discover which company still supports publication, which government program permits thesis continuity, which laboratory owns the required instrument, which professor will support an external placement, which intellectual-property agreement preserves dissertation rights, and which short appointment is genuine apprenticeship rather than inexpensive labor. Every crossing consumes calendar time and negotiating power before it produces science.

The 2024 Survey of Earned Doctorates shows why a single academic pipeline is no longer an adequate mental model. Of recipients reporting definite U.S. commitments, 39 percent were entering postdoctoral positions and 61 percent had non-postdoc employment commitments. Among the latter, academia and industry or business each accounted for 40 percent; government, nonprofits, and other destinations made up the rest. These are immediate commitments, not lifetime paths, but they establish that doctoral training supplies several institutional systems. D

The National Academies’ graduate-education report accordingly calls for deep specialization joined to broad technical literacy, project-based learning, transparent career outcomes, internships, and exposure to careers across sectors. It also places duties on agencies, universities, faculty, employers, professional societies, and students. That allocation matters: initiative by the student is useful, but the actors controlling money, degree rules, facilities, and employment possess most of the power. D

Choose a spine, not a cage

A continuity spine is a question durable enough to survive several methods and settings. It should be more specific than “AI” or “quantum,” but more durable than one architecture, dataset, sponsor milestone, or product feature.

Examples of the right level are:

  • How can computation remain trustworthy when the operator, hardware, or model is not fully trusted?
  • How can a physical system expose enough evidence to distinguish a real effect from a measurement artifact?
  • How can a distributed system fail safely when communication and identity are contested?
  • How can a biological intervention be made causal, measurable, and manufacturable rather than merely correlated with an outcome?

The question is not a slogan placed above unrelated projects. It must specify what uncertainty remains, what evidence could resolve it, and why the current environment cannot supply every necessary test.

A spine protects continuity, but it must not become a cage. Evidence can reveal that the problem is ill-posed, already solved, socially undesirable, or outside your comparative advantage. Changing the question after such evidence is learning. Changing it at every funding call because nothing portable survived is institutional thrashing.

Earn one hard depth before collecting boundaries

The phrase “interdisciplinary researcher” can hide a dangerous absence: no community has yet trusted the person with its hardest standards.

Before optimizing breadth, learn one craft deeply enough to recognize its failure modes. Write a proof that survives hostile reading. Build a system whose performance is measured outside your own test harness. Calibrate an instrument. Reproduce a result. Maintain code after its authors leave. Follow an experiment through an anomalous result rather than replacing the plot. Sit with a user who rejects the prototype for a reason the paper never represented.

The National Academies’ mentorship synthesis treats mentorship as a learned practice with research, career, psychosocial, sponsorship, and inclusion functions—not as chemistry that can be left to appear spontaneously. One famous adviser rarely supplies all those functions. Assemble a small mentorship portfolio, but name the missing function before adding another person. D

Likewise, a postdoctoral appointment should not be accepted merely because another credential is conventional. The Postdoctoral Experience Revisited distinguishes a defined period of advanced, mentored training from the use of postdoctoral researchers as long-term, low-cost labor. Before accepting, ask what capability will be learned, who is responsible for teaching it, what independence will exist, what the completion condition is, and where prior researchers went. D

Every proposed move should answer one question:

What can this environment add that the current one cannot supply at reasonable cost?

There are only a few strong answers:

  • a physical instrument, fabrication process, dataset, test range, or operating environment;
  • a mentor or team with difficult-to-reconstruct tacit knowledge;
  • a mission user who can state consequential requirements and reject inadequate work;
  • a systems or manufacturing constraint that changes which theory matters;
  • a verification culture capable of exposing errors your home field normalizes;
  • a transition mechanism through standards, procurement, licensing, operations, or product engineering; or
  • a degree of problem-selection freedom not available under the current sponsor.

“It is prestigious” is not enough. “It pays more” may be a completely legitimate life decision, but it is a compensation argument, not a research-compounding argument. “Everyone is working on it” may indicate an important field or merely a crowded incentive surface.

Current programs illustrate pieces of the bridge. NSF INTERN can add a funded nonacademic research placement to some NSF-supported graduate paths. DOE’s SCGSR program allows eligible doctoral students to conduct part of their thesis work with a national-laboratory scientist and facilities unavailable at the university. NIST’s NRC program offers a defined federal-laboratory postdoctoral setting with staff mentorship and measurement infrastructure. NSF I-Corps joins technical, entrepreneurial, and industry-mentor roles around customer discovery and transition. D

These are fragments, not a restored ecosystem. Availability changes. Eligibility may depend on an adviser’s active award, field and mission alignment, citizenship or residency, degree timing, host participation, or competitive selection. A field guide must not turn examples into promises. Verify current terms with the program, your university, and prospective host before making a decision.

Write a boundary contract before crossing

Most boundary loss is predictable. It occurs because everyone discusses the exciting work and postpones ownership, publication, access, time, and return conditions.

Before an internship, laboratory residency, joint appointment, sponsored project, or spinout, write a one-page boundary contract even when the formal legal agreement will be much longer. It should answer:

Boundary question Minimum answer
Research continuity Which question and artifacts may continue across the boundary?
Time What is the protected research fraction, and what events can consume it?
Mentorship Who owes technical review, career guidance, and conflict resolution—and how often?
Evidence access Which data, code, instruments, logs, and collaborators remain available during and after the placement?
Publication Who reviews a manuscript, on what clock, and what happens if review stalls?
IP and confidentiality What is background IP, what may be published in a thesis, and what survives departure?
Credit How are authorship, invention, software, and team contributions recorded?
Reintegration Who funds the return, preserves degree progress, and translates the new capability back into the home group?
Exit What is the completion condition, and which artifacts can leave with the researcher?

This page is not a substitute for university counsel, an employment agreement, immigration advice, or an IP lawyer. Its function is epistemic: reveal incompatible expectations before the least powerful participant has already relocated or delayed a degree.

NSF’s detailed INTERN mechanism is revealing because it requires the academic and host organizations to address mentoring and intellectual-property arrangements in advance. Those are not administrative footnotes. They determine whether the placement compounds the thesis or severs it.

Carry a boundary packet

Institutions maintain memory badly across departures, and individuals remember less than they think. Create a portable, legally permissible boundary packet:

  1. the current research question and its most important competing explanations;
  2. a dependency map of claims, datasets, code, instruments, and people;
  3. failed approaches and the observations that killed them;
  4. unresolved anomalies and the next decisive experiments;
  5. a provenance record for artifacts you are permitted to retain;
  6. explicit restrictions on data, code, export, security, publication, and IP;
  7. the map of who knows what and who can authorize the next step; and
  8. a ninety-day plan for converting the new environment into new evidence.

The packet is not an attempt to take employer property. Protected material must remain protected. Sometimes the portable object is only a clean-room statement of the question, a public bibliography, personal skills, and the names of people you may contact. Record that boundary honestly.

The organizational evidence explains why this discipline matters. Lewis and colleagues found that a group’s shared knowledge of who knows what can support learning transfer, but the effect depends on prior joint experience. Reagans, Argote, and Brooks found that knowing how to work together and knowing teammates’ expertise predicted learning beyond individual experience alone. Rao and Argote showed in a controlled task that structures and routines can buffer some turnover loss. None proves that a personal notebook recreates a laboratory. Together they reject the fiction that hiring the same number of individually qualified people instantly restores the same team state. D

A seven-year path is a portfolio, not a timetable

The following sequence is a diagnostic example, not a norm and not a promise:

Stage Capability objective Evidence that the stage compounded
Years 0–2 Earn depth in one hard method and one correction culture An artifact or result survives independent scrutiny; important failures are understood
Years 2–4 Add one unavailable complement through a funded placement or collaboration The original question changes because of a real instrument, user, constraint, or dataset
Years 3–5 Acquire systems and team responsibility You can explain dependencies, failure containment, labor, schedule, and integration—not only your component
Years 4–6 Test transition without assuming a startup is the answer A user, operator, standards body, licensee, or procurement path provides disconfirming evidence
Years 5–7 Exercise bounded problem-selection and budget authority You make a documented bet, protect verification, allocate resources, and learn from the outcome

Life will not respect these numbers. Caregiving, visa status, disability, financial pressure, field-specific training, failed experiments, geographic constraints, and chance can reorder or prevent stages. The sequence should never become a test of personal worth. Its purpose is to identify which capability the surrounding system has failed to make available.

In the AI age, choose contact with reality

AI will make literature synthesis, code scaffolding, formal presentation, simulation setup, and hypothesis generation cheaper. This is valuable, but it changes the apprenticeship problem.

When plausible form becomes abundant, seek environments that expose you to scarce correction:

  • measurements whose noise cannot be prompted away;
  • production systems with users and adversaries;
  • formal proofs checked independently of rhetorical confidence;
  • hardware, biological, or field experiments with real failure modes;
  • maintainers responsible after the demonstration;
  • reviewers and mentors rewarded for finding the decisive error; and
  • customers permitted to say that the technically impressive result solves the wrong problem.

Do not outsource taste to the model. Use models to widen search and lower mechanical cost, then spend the saved time on problem selection, causal identification, verification, systems integration, and communication with people who bear the consequences.

The strongest AI-era researcher may not be the person who produces the most artifacts. It may be the person who knows which generated artifact deserves scarce physical, mathematical, institutional, or human attention.

What institutions owe the next generation

No student can solve this by personal discipline alone. A serious institutional response would provide:

  • portable bridge awards that belong partly to the researcher and can fund placements, return, and reintegration;
  • paired mentors across the home and host institutions, each with explicit duties and evaluation;
  • continuity agreements covering publication, thesis progress, IP, data, software, and access before work begins;
  • staff-scientist and research-engineer careers so every experienced researcher is not forced into faculty fundraising or temporary training status;
  • transition residencies in standards, procurement, manufacturing, operations, nonprofits, and mission organizations—not only startup accelerators;
  • return rights so an external placement does not silently cost a laboratory position, funding priority, or year of degree progress;
  • open bridge mechanisms for international researchers and those outside well-funded adviser networks, consistent with legitimate security constraints; and
  • pathway accounting that measures who enters, who is excluded, which capabilities compound, and whether participants incur debt, delay, or career penalties.

Universities should not stigmatize nonacademic exposure while celebrating placement statistics. Companies should not advertise research internships that are ordinary feature work with no research time. Government laboratories should distinguish apprenticeship from staffing. Advisers should not veto a placement merely because it reallocates a student’s labor. Funders should pay the indirect work of coordination instead of assuming goodwill will make two institutions interoperate.

The old system often provided continuity accidentally through organizational scale and economic rents. The new system will have to provide it deliberately.

The equity test is part of the capability test

Advice of this kind can become a prestige algorithm: find famous mentors, circulate through elite laboratories, collect scarce fellowships, and call the result merit. That would reproduce the problem.

A national developmental path cannot depend on a student already knowing the correct program manager, laboratory scientist, founder, or professor. It cannot treat citizenship restrictions, unpaid relocation, inaccessible facilities, family obligations, or an adviser’s refusal as evidence of weak commitment. It cannot build the public research workforce by privatizing pathway risk onto people least able to absorb it.

The relevant outcome is not whether a few unusually connected people can still construct an excellent route. It is the number, diversity, cost, and demonstrated capability of paths that ordinary talented entrants can reproduce.

How this advice could be wrong

Falsification matters because a field guide can become autobiography disguised as a universal law. A direct counterexample would be a single durable institution that supplies depth, multiple correction cultures, physical and mission contact, transition, and long-term autonomy more effectively than a sequence of crossings. A direct falsification would be evidence that carefully supported boundary crossings fail to improve retained capability or consequential research outcomes after accounting for who is selected into them.

The compounding-path thesis should weaken if longitudinal evidence shows that:

  • researchers staying within one institution or discipline systematically produce deeper and more consequential work after accounting for selection;
  • boundary crossings add coordination cost without improving verification, transition, judgment, or retained capability;
  • portable records and paired mentoring do not reduce restart time or loss;
  • bridge programs prolong training, increase precarity, or mainly subsidize host labor;
  • hiring and funding systems already reward coherent cross-sector contributions without requiring applicants to translate them into conventional proxies; or
  • AI and distributed collaboration make institutional adjacency sufficiently cheap that physical and organizational crossings add little.

Some people should stay inside one excellent environment for a decade. Some should change fields completely. Some should leave research for work they value more. The article argues neither that motion is virtue nor that a research career is owed to anyone.

It makes a narrower claim: when a consequential question requires capabilities distributed across institutions, the path should be designed so that knowledge compounds instead of resetting at every boundary.

A short field rule

Before the next move, ask five questions:

  1. What durable question will cross with me?
  2. What missing capability will this boundary add?
  3. Which accumulated state may legally and practically survive?
  4. Who is obligated to mentor, verify, and enable reintegration?
  5. What evidence, one year later, would show that this was compounding rather than motion?

If those questions have no answers, the opportunity may still be worthwhile for money, location, family, safety, or joy. Those are real reasons. But do not mistake a good life decision for a research-continuity mechanism.

And do not accept the premise that students must solve the institutional problem alone. The long-run objective is not to train a generation of heroic boundary negotiators. It is to rebuild enough connected research infrastructure that accumulating a research life becomes normal again.

Open question: Which current programs or laboratories genuinely let young researchers cross a boundary, retain accumulated state, and return with more capability—and which only rename the restart?

Argument under pressure

Two rounds of objection—not a ceremonial counterargument

These are simulated steelman exchanges. The second objection responds to the first answer; the conclusion is narrowed where the objection survives.

Claim under test

A young researcher should build a compounding sequence across complementary institutions instead of optimizing every move as an isolated placement.

  1. First-level objection

    This is an argument for breadth at the moment when serious research requires extreme specialization. Boundary tourists acquire vocabulary without earning the depth needed to discover anything important.

  2. Response

    The prescription begins with one hard depth and a durable problem, not generalized exposure. Crossings are justified only by a complement the problem genuinely lacks; unrelated prestige stops, shallow rotations, and method collecting fail that test.

  3. Second-level objection

    Hiring and promotion committees still evaluate legible field-specific outputs. Even coherent cross-sector work can look nonlinear, unfocused, or insufficiently independent.

  4. Bounded conclusion

    Preserve a stable problem statement, name the causal contribution of every crossing, and produce field-legible witnesses at each stage. Institutions should evaluate a contribution portfolio and retained capability—not force applicants to conceal the very transitions the research system needs.

National Academies of... National Center for S...

Claim under test

Researchers can reduce boundary loss by negotiating continuity, keeping portable evidence state, and assembling multiple mentors.

  1. First-level objection

    This converts institutional failure into self-help and asks students with the least authority, money, citizenship security, or professional network to reconstruct Bell Labs for themselves.

  2. Response

    That objection is correct as policy. The individual practices are defensive tools, not an adequate institutional settlement. The corresponding duty falls on universities, funders, laboratories, and employers to finance placements, co-mentoring, portable support, reintegration, and explicit IP and publication agreements.

  3. Second-level objection

    Even funded bridge programs remain selective. Some require an adviser with an active award, alignment with a federal mission, U.S. citizenship or permanent residence, or access to a host scientist willing to collaborate.

  4. Bounded conclusion

    Publish the constraints, measure who cannot use each bridge, fund open alternatives, and never treat lack of access as lack of ambition. A pathway is a national capability only if talented people can enter it without inherited institutional capital.

U.S. National Science... U.S. Department of En... National Institute of... National Academies of...

Revision ledger

Corrections and feedback incorporated

Inspectable claims

Source Ledger

  1. Peer-reviewed laboratory experiment

    Transactive Memory Systems, Learning, and Learning Transfer

    Kyle Lewis, Donald Lange, and Lynette Gillis. Supports: Experimental evidence that an experienced group's shared map of who knows what can support learning transfer, with effects depending on sufficient prior joint experience.

  2. Peer-reviewed organizational study

    Individual Experience and Experience Working Together—Predicting Learning Rates from Knowing Who Knows What and Knowing How to Work Together

    Ray Reagans, Linda Argote, and Daria Brooks. Supports: Evidence that prior experience working together and knowledge of teammates' expertise predict team learning rates beyond individual experience alone.

  3. Peer-reviewed laboratory experiment

    Organizational Learning and Forgetting—The Effects of Turnover and Structure

    Raghavendra D. Rao and Linda Argote. Supports: Experimental evidence that organizational structures and routines can buffer some knowledge loss after personnel turnover, subject to the limits of a controlled task setting.

  4. Federal statistical report

    Doctorate Recipients from U.S. Universities: 2024—Postgraduation Trends

    National Center for Science and Engineering Statistics. Supports: The distribution of definite 2024 U.S. postgraduation commitments between postdoctoral training and employment, and the division of non-postdoc employment commitments among academia, industry, government, nonprofits, and other destinations.

  5. Consensus study report

    Graduate STEM Education for the 21st Century

    National Academies of Sciences, Engineering, and Medicine. Supports: The case for student-centered graduate education, deep specialization with broad technical literacy, project-based learning, transparent career outcomes, and funded exposure to academic and nonacademic pathways.

  6. Consensus study report

    The Science of Effective Mentorship in STEMM

    National Academies of Sciences, Engineering, and Medicine. Supports: Evidence-based treatment of mentorship as a learned and improvable practice requiring institutional support, multiple functions, assessment, and attention to inclusion rather than an informal relationship left to chance.

  7. Consensus study report

    The Postdoctoral Experience Revisited

    National Academy of Sciences, National Academy of Engineering, and Institute of Medicine. Supports: The distinction between a defined period of advanced mentored training and the use of postdoctoral researchers as inexpensive long-term labor, together with recommendations on duration, compensation, career development, and mentoring.

  8. Federal program primary source

    Non-Academic Research Internships for Graduate Students (INTERN)

    U.S. National Science Foundation. Supports: A current mechanism for adding research experience in companies, government laboratories, startups, and nonprofits to an NSF-supported graduate pathway, including mentoring, time, funding, and advance treatment of intellectual-property and publication rights.

  9. Federal program primary source

    Office of Science Graduate Student Research Program

    U.S. Department of Energy Office of Science. Supports: A current bridge allowing eligible doctoral students to conduct three to twelve months of thesis research with a DOE national-laboratory scientist and use facilities unavailable at their universities, subject to mission, citizenship or residency, and program constraints.

  10. Federal program primary source

    NSF I-Corps Teams

    U.S. National Science Foundation. Supports: A structured transition mechanism that joins technical, entrepreneurial, and industry-mentor roles and uses customer discovery to test whether research can become a product, service, license, or better use-directed question.

  11. Federal laboratory program primary source

    Postdoctoral Students—NRC Postdocs

    National Institute of Standards and Technology. Supports: A current example of a defined two-year federal-laboratory postdoctoral research opportunity with facilities and staff mentorship, as well as citizenship and degree-age eligibility constraints that prevent it from being a universal bridge.

Public record

Version history

Published versions are immutable snapshots. Corrections create a new version rather than silently changing the historical record.

No public snapshot has been archived yet. The first snapshot is created when the article is published.

Read the correction, withdrawal, and versioning policy

← Return to the series map

This essay distinguishes firsthand memory, documentary evidence, interviews, and analysis. Published versions remain available even after revision or withdrawal.