Explanations and the assumption audit
Find out why a plan fails, where the money went in a year, which edit explains a difference between two plans, and what is wrong with your assumptions.
A number you cannot trace is a number you cannot trust. Glideworth’s explanations work from the engine’s own trace and ledger, so each sentence points at the figures behind it rather than at a generated guess. There are four kinds, and one check that runs before any of them.
Why did the plan fail?
After a Monte Carlo run, the failure analysis says when failing trials first run short, whether the early-return quintile of a trial separates success from failure (sequence-of-returns risk), and whether failures cluster before or after retirement. A general “explain” turns findings from the engine trace and ledger into sentences, and can take a question: why did this withdrawal come from that account?
Where did the money go?
Explain a year gives, for a single year or a range, income, outflows, saving, withdrawals by account, investment return and the change in net worth, as sentences plus the top line items. Ask for full detail and you get the ledger and trace evidence behind each one. The year drill-down in the app ends with the reconciliation described in projections and the ledger.
What explains a difference between two plans?
Given a baseline plan and another, explain difference shows what differs, an explanation of the outcome gap and, if you ask for attribution, how much of the change each individual edit explains. It replays each edit alone and reports the interaction left over. For chance of success it runs Monte Carlo as a job. This is the natural partner of branches.
A sensitivity analysis answers a related question: which assumptions move the outcome most, drawn as a tornado. It can use end net worth, lifetime tax, total shortfall or chance of success.
The assumption audit
The audit looks for modelling mistakes rather than market risk. It is deterministic and lists findings most severe first. Examples it checks for:
- housing counted twice;
- today’s money and future money mixed in one plan;
- implausible returns or an implausibly high early withdrawal rate;
- a missing flow, or nothing paying for care costs in your 80s;
- early withdrawals that carry penalties, and cash drag;
- cross-border problems such as a missing exchange rate, an unknown treaty pair, or a state pension that will fall short on qualifying years.
A finding with a fix comes with ready-made plan edits. You can preview them (nothing is stored), propose them as a change set, and apply them only when you agree.
A worked illustration
An illustrative finding, as it might read:
Warning: housing may be counted twice. The plan has a £1,800 a month “Living costs” expense and a £900 a month “Mortgage” expense, but also a mortgage on the home in current finances. Fix: remove the separate mortgage expense. Preview: end net worth rises by £141,000 (today’s money).
The figures are invented. The pattern is real: a finding, why it matters, a fix, and a preview before anything
changes. Through MCP the sequence is audit_plan, then preview_changes with the finding’s patch, then
propose_change_set, and apply_change_set once you say yes. See MCP.
What it does not do
- The audit is not exhaustive. It catches common mistakes, not every one. Garbage in, garbage out.
- Explanations describe the model, not the world. “The plan fails because of weak early returns” is a statement about the simulated trials under your assumptions.
- No completeness score yet. Guided data-quality nudges while you set up are planned, not built.
- Survivor pensions beyond what you enter are not added to survivor explanations, and the scenario says so.
- No language model runs on your data. Explanations are generated by deterministic code, not by an AI service. An AI assistant that you connect over MCP may call these same tools; what it does with the results is up to that assistant and its provider.
Go further
- Projections and the ledger for the evidence these explanations cite.
- Solvers: once you know why, ask what to change.
- Security and privacy for what an assistant can see.
See where your plan lands.
Free to self-host today, with every feature. The hosted service opens later, at $9 a month or $90 a year.