Risks and stage gates
Overview
The main threat to this project is not failure to write code. It is failure to know when the outputs are not yet good enough. This chapter defines the principal risks and the stage gates that govern continuation. Gates are score thresholds on the record defined in scoring-and-resolution.md, not narrative judgments.
Principal risks
1. Earnings histories look reasonable but are wrong in policy-relevant ways
It is possible to match simple age-earnings profiles while still getting mobility, persistence, zero-earnings patterns, or AIME distributions wrong. That would produce a panel that looks plausible but fails at the core Social Security task.
2. Cross-sectional calibration conflicts with longitudinal validation
A model may be able to match annual aggregates and still distort longitudinal transitions. Conversely, a model that respects PSID-style transitions may drift away from administrative totals. Managing that tension is one of the core research problems of the project.
3. Platform work expands faster than policy validation
Now that the project is best understood as making populace longitudinal, there is a new risk: the population platform can become technically interesting without yet being decision-useful for Social Security. That would be real research progress, but it would not by itself justify strong policy claims.
4. Benefit logic works for typical cases but fails at important edges
Social Security policy analysis depends on edge cases, not just median retired workers. Family benefits, disability pathways, and complex claiming patterns matter.
5. Baseline projections drift
Even a strong baseline panel can become untrustworthy when projected forward if drift is not controlled explicitly.
6. Productization happens before validation is mature
A public interface is attractive and visible, but it can also amplify weaknesses if the validation record is not ready.
Stage gate 1: longitudinal Populace quality
The project should advance past stage 1 only if it can show that longitudinal populace is credible on multiple dimensions:
- age-earnings levels
- dispersion and percentiles
- zero-earnings patterns
- AIME distributions
- mobility and persistence
Failure here is not a minor miss. It is a reason to pause and revise the methodology.
Stage gate 2: benefit and family validation
The project should advance past stage 2 only if the synthetic panel produces benefit outcomes that are good enough for exploratory policy analysis:
- beneficiary counts by major type
- average benefits
- selected benefit distributions
- family and auxiliary outcomes that are directionally credible
If the model only works for retired worker benefits and breaks down in family or disability pathways, that should be stated clearly and used to bound later claims.
Stage gate 3: projection quality
The project should advance to public reform analysis only if baseline projections remain stable and interpretable:
- drift is measured and controlled
- near-term projections track published benchmarks reasonably well
- sensitivity to key assumptions is documented
If the model cannot produce a credible baseline projection, it is not ready for strong claims about reform packages.
Stage gate 4: public product readiness
The project should release a public interface only if:
- headline examples are reproducible
- limitations are documented in plain language
- the public product is narrower than the full internal research stack
- the validation record is published alongside the interface
If a gate fails
Failure should still produce useful public outputs. Possible fallback deliverables include:
- a public methods paper on synthetic longitudinal reconstruction
- a validated baseline panel without a public policy interface
- a narrower model focused on a subset of beneficiary types
- static or semi-dynamic tools built on the same validation assets
The point of stage gates is not to threaten cancellation. It is to avoid pretending that an unvalidated model is ready for claims it cannot yet support.
Explicit near-term non-goals
In the first major phase of the project, the model should not claim to:
- replace SSA’s internal models for official estimates
- provide full general-equilibrium feedback analysis
- model every behavioral response in the retirement system
- solve all adjacent lifecycle policy questions at once
Those are legitimate long-run ambitions, but they should not be confused with early validation milestones.