The bond final-year rule: what it merges — and why N never gets +1

Every top-slicing computation starts with one integer: N, the number of complete years the policy has run. Divide the gain by N and you have the annual-equivalent slice that the whole relief is built on (ITTOIA 2005 s.536(1)). So when two readings of “complete years” disagree by one, every figure downstream disagrees too — and there is a specific situation, the final-year straddle, where a widely-repeated misreading adds a year that the statute never grants.

The question a date forces

A bond is fully surrendered a few weeks after its policy anniversary, with both dates falling in the same tax year. Does the stub period after the anniversary count for anything? This is not a hypothetical: in a February 2026 thread on the Big Tent — the Paraplanners' Assembly discussion board — a paraplanner could not reconcile manual chargeable-gain calculations on an onshore bond with the figures from two providers' online calculators (M&G's and Quilter's). The resolution was the final-year rule, and the thread's expert reply noted that the provider's own earlier excess-gain certificate would be superseded by a reissued one. The rule is real, it is confusing, and tools and certificates genuinely diverge around it in practice.

What the statute actually says

For top-slicing relief, N is fixed by IPTM7560 in one sentence, and the sentence forecloses the debate by name:

Pure date arithmetic: count the complete 12-month periods from commencement to the event. No tax-year boundary enters that count. Take a policy commencing 15 June 2017, fully surrendered 20 August 2025 — the case pinned in our engine's regression tests. The eighth anniversary, 15 June 2025, precedes the event; the ninth has not arrived. N is 8.

What s.499(5) merges — and what it doesn't

The misreading has a real rule underneath it, which is why it survives. ITTOIA 2005 s.499(5) says that where the final insurance year would begin and end in the same tax year, it is merged into the previous insurance year. Note the direction: a count of insurance years going down — two years become one — and note the purpose: it stops the part-surrender 5%-allowance schedule double-counting within a single tax year. It governs the gain machinery on part surrenders. It says nothing about N for top-slicing relief, which s.536(1) and IPTM7560 define in 12-month periods, expressly not insurance years. Conflating the two — “the final year merges, so N must move” — is how the phantom +1 gets built.

The merge does have visible consequences, which is exactly what the Big Tent thread shows: an excess gain certified after the last anniversary can be folded into the final-event computation, and the provider reissues the certificate. Where that happens, the working checks against the provider's reissuedchargeable-event certificate — the certificate remains the document of record, and any difference between a tool's figure and the certificate is a question for the provider, not a number to override.

The worked example, on HMRC's own figures

HMRC's published example at IPTM3850 (Example 1) fixes the whole chain: a £50,000 onshore gain over 5 complete years with £45,000 of employment income in 2022/23. The slice is £10,000; the relief is £4,616. Our engine reproduces every figure of that example at 0-pence tolerance on every commit, and the same case is live in the calculator below — change N from 5 to 6 and watch the slice and the relief move, which is the entire point: one year of N is never a rounding detail.

An inflated N cuts the other way from most calculation errors: a larger N means a smaller slice, which tends to keep the notional computation in lower bands and produce morerelief than is due. The +1 misreading fires on any final event whose most recent anniversary falls in the same tax year — roughly ten to eleven months of possible event dates in any policy year — so a tool that carries it does not fail on edge cases; it fails on most full surrenders, quietly, in the taxpayer's favour, which is the kind of error a compliance review eventually finds.

How our engine handles it — including the correction

ParaplanAI's engine computes N as the pass-through of complete 12-month periods, per IPTM7560, and pins it with a property test: for every final event, across 1,000 random date pairs per run, the effective N must equal the complete-years count exactly. The same-tax-year straddle is still detected — it surfaces as an informational flag on the result and the annex, prompting a check of the event dates and the certificate history against the provider — but it is not wired to change any number.

The candid part of this case study: our own engine shipped the +1 reading first. A 2026 calculation-correctness audit flagged it, the derivation was re-done from s.536(1) and IPTM7560, and the decision record ADR-048 (which supersedes the earlier ADR-012) removed the adjustment and inverted the affected regression tests. The corrected rule now runs in every tax-year configuration we serve. We publish that history rather than bury it because it is the strongest argument for the method: a misreading with no statutory basis survives exactly as long as nothing forces the working back to the statute — and a public, HMRC-anchored regression corpus is what does the forcing.

The wider rules for fixing N — including the different clock for part-surrender excess events — are set out in complete years (N): full surrender vs excess; the five-step computation N feeds is in how to calculate top-slicing relief.

IPTM7560 (N — complete 12-month periods) · ITTOIA 2005 s.536 (annual equivalent) · ITTOIA 2005 s.499 (final-insurance-year merge) · IPTM3850 (worked example) · engine reading ADR-048