Before the framework, one finding I did not expect when I started, because it is the clearest example of what happens when you look at compensation through time.
Preservation architecture has to satisfy a constraint the classical designs never faced. It has to survive the career, not merely the holding period. An architecture that preserves beautifully inside one employer and dies at the boundary has failed the actual career even if it never failed a market test.
The tempting conclusion is that portability should simply be maximized. It should be resisted, because portability trades against other things the liability needs. It trades against pooling, since a mortality pool is not a bearer instrument and the worker who exits takes a transfer value rather than the pool's ongoing risk reduction. It trades against commitment, since the same access that lets value cross a boundary lets it leak, and the documented fate of small balances at job change is cash-out. A stack therefore has to distinguish portability of the claim from liquidity of the claim, which are separable by design even though careless architecture merges them.
And it trades against the employer's legitimate alignment interests, which is where the surprise sits.
The Better It Carries, the Harder It Binds
Matching and vesting price retention against ownership, and the evidence on whether they buy the retention they are sold as buying is more conditional than plan design usually acknowledges. Broad-based option grants have been found to reduce turnover only where the market had risen since the grant, with most of the retention effect reversing in the year after full vesting. And when an accounting change accelerated vesting at a set of firms, voluntary chief executive departures rose from five percent to twenty-one.
Handcuffs, on that evidence, hold when the asset has done well and open when it has not, which is close to the opposite of what a retention instrument is sold as doing.
The collision follows from there. The third job of compensation is alignment and the fourth is preservation. Both use duration, so they have historically pulled in roughly the same direction, and no designer has ever had to choose between them.
But an instrument that carries value well becomes, by the very fact of carrying it well, harder to leave. A deferred claim in an asset that preserves or appreciates across the vesting period binds with a force set by the asset rather than by the retention policy. The better a carrier performs the fourth job, the more forcefully it performs the third.
Alignment and preservation have been held together for a century partly by the mediocrity of the carriers. They come apart when a carrier improves.
The design implication is to separate the two jobs deliberately, rather than to let one fund the other without anyone having decided that it should.
The force is also computable, and almost nobody computes it. Call it vesting intensity: the market value of everything an employee would forfeit by resigning, divided by annual pay, expressed in years of pay. The unit has precedent, since pension economics measured the cost of quitting a backloaded plan in exactly those terms. The inputs sit in a vesting table and a payroll file. And essentially no employer produces the number, and no compensation statement displays it.
It belongs in the temporal audit the first paper proposed, and it should be reported in bands rather than as a score, because the honest number is the shadow of an unmeasurable one. What an undiversified, illiquid, unhedgeable position is actually worth to the person holding it cannot be computed.
One further distinction, and it may be the most consequential design concept in the paper. Distinguish portability of the claim from portability of the architecture.
Modern systems have made real progress on the first. A defined contribution balance can generally be rolled from one arrangement to the next. Almost no progress has been made on the second. The defaults that made the worker a saver, the employer contribution, the pooling access, the automation, the liability targeting: none of it travels. At every boundary the architecture dies and has to be rebuilt, or more often is not, and the worker's provisioning quality resets to whatever the next employer happens to supply.
Call the missing property architectural portability. One jurisdiction has begun to build it: Australia's stapling reform attaches the worker's superannuation account to the worker rather than the employer, so at a job change the account, with its provider, settings and insurance, follows the person by default. Stapling is partial, and it has generated its own debates about workers stapled to poor products. But it demonstrates that architectural portability is a policy choice rather than a conceptual impossibility. Everywhere else, it remains a name for an absence.
The Six Tests
Everything in the paper compresses into six tests. Each is derived from a section already argued. Each is phrased as questions rather than scores. And each applies to any preservation architecture, existing or not yet invented.
The Arrival Test. Against the worker's future liability, and not against any index, does the architecture credibly carry value to the horizon at which the liability falls due? What is the target, is it named, is progress measured in liability-relative terms, and would anyone know if the architecture were failing it?
The Assignment Test. Name every risk. Who bears each one, was the assignment chosen or inherited, and where pooling could genuinely reduce a risk, does the architecture reach it?
The Ownership Test. At the value layer and the custody layer separately, what does the worker actually hold? What has to remain true, and about which institutions, for the claims to pay? What could the worker still have after the employer is gone, the provider has changed and the rules have moved? And the question the gentlest jurisdictions never had to ask: what legal and political order has to persist for this property to remain property?
The Journey Test. Does the architecture survive the career the worker will actually have? What crosses an employer boundary, what dies there, what leaks there, and does the architecture itself, not merely the balance, continue?
The Human Test. Three verbs, separately, because architectures fail them separately. Can an ordinary person, busy with a life, begin this architecture without expertise they do not have, persist with it through drawdowns and interruptions and job changes, and convert it at the end without a decision that terrifies them? Required action is a defect. The strongest pass belongs to an architecture whose inaction path succeeds, provided the exit from that path stays open, because an inaction path with no exit is not architecture but conscription.
The Conversion Test. At the end, can accumulated value become the living it was preserved for? Where the liability is longevity-contingent, does a path to pooling exist, and at what price in ownership and liquidity? The difficulty is certified by the field's own elders: Sharpe called decumulation the nastiest, hardest problem in finance, and he was describing exactly this. An architecture with no conversion path has preserved a number rather than a life.
Three notes keep the tests from being misread. They are written for the compensation designer, because that is where the increments are designed. They are ordinal rather than cardinal, and no score of the form "7.2 out of 10" should ever be derived from them. And a test counts as run only when its answer is written down, with workings, because unfalsifiable frameworks get captured by whoever markets them first, and disclosure at least makes capture leave fingerprints.
The Part Where the Framework Refuses
Run different architectures through the tests and the framework shows its character by declining to crown anyone.
A classical defined benefit pension passes Arrival and Conversion brilliantly, passes the Human Test almost by definition, and strains at Ownership and Journey, where everything depends on a sponsor's continuity.
An auto-enrolled target-date defined contribution plan passes the Human Test by engineering, passes Journey adequately in its portable modern form, carries a crude implicit liability logic into Arrival, and strains at Assignment and Conversion.
A self-directed portfolio of directly held property passes Ownership and Journey at their maxima and strains almost everywhere else.
And then there is the one I could not omit without the omission being legible as evasion. Consider a composite collective architecture assembled from features already implemented across leading national systems: mandatory participation and pooled low-fee governance in some, stapled accounts in one, a decumulation covenant in the same one, collective risk-sharing in others. No single system contains every feature, but each feature exists in current practice.
Run it through the tests and the honest report is that it outperforms everything else in the inventory on four of the six. The Human Test by construction, since neither participation nor persistence requires a decision. The Journey Test by mandate, which is the one force that has actually delivered cross-sector pooling and architectural portability. The Assignment Test, because scale purchases the pooling, governance and fee levels the arithmetic says are decisive. And the Conversion Test, where such systems are building the only conversion architecture that currently exists at scale.
That is the strongest scorecard in the paper, and the paper says so aloud.
What remains is not nothing. The collective strains at Ownership, where the worker holds a claim adjustable by legislation rather than property that survives political change, and at the sovereign layer of Arrival. Its failure pattern, in a word, is concentration. Every vulnerability it has sits on a single political balance sheet, and nothing in the worker's statement discloses the accumulation.
Whether four decisive passes outweigh two concentrated strains is exactly the weighting question the framework refuses to decide in the abstract. That answer belongs to the liability, the worker, and the political community.
Four architectures, four different failure patterns. That is the finding. The tests locate the tradeoffs that the liability and the worker have to price.
What the Series Was For
The conclusion of this paper is not that everything we have is failing. The opposite is closer to the truth.
Human societies have built remarkably effective preservation technologies. Pensions that named liabilities and met them for generations. Insurance and annuities that reduce risks no individual can carry. Diversified productive assets with the strongest accumulation record in the data. Indexed bonds, automatic enrollment, behavioral defaults, tax-advantaged wrappers, pooled institutions of every scale.
The problem is not that the old architecture has nothing to teach us. It is that the architecture we inherited was assembled under economic, employment, institutional and monetary conditions that no longer exist in precisely the same form.
The task is recomposition: recombining what works around the worker, and around the economic reality the worker now inhabits.
There is one more paper in this series, and it takes up the question this one deliberately left open. The framework here evaluates architectures against a liability. It says nothing about the unit in which both are measured. That turns out to be its own problem, and it is the last one.