Predictions we freeze in public — including the misses.

Client logos can be rented; a public scoring record cannot. Each month this page will carry five LEXUN predictions on fast-resolving public questions — each dated, frozen with a named resolution source and criterion, then scored with the Brier rule when reality answers. Misses are published permanently.

Whose forecasts these are. Every entry names its forecaster and the board it is scored on, and scores are never pooled across boards. The current cohort is operator judgement about public statistics, on the editorial board. None of LEXUN’s three decision models forecasts CPI, Bank Rate, unemployment or insolvency counts, so nothing here is evidence that any of them is calibrated — and a future table showing a good editorial Brier score must never be read as if it were.

The rules every entry follows

• Frozen before the event, with a visible publish date — never edited afterwards.
• A named public resolution source (e.g. ONS release, Bank of England decision) and an unambiguous criterion, fixed at publish time.
• A probability, not a hedge — scored with the Brier rule on resolution.
• Misses stay on this page permanently. Deleting a miss would end this product's reason to exist.
• Withdrawal is bounded: within 24 hours of issue, only for a defect in the forecast's own construction, never after. A withdrawn entry stays in the file as void, is never scored as a hit or a miss, and is never replaced silently — a replacement names what it supersedes. The release gate enforces the window.

Why forecasts are withdrawn, and why that is not editing

Four of the twelve forecasts on this register were withdrawn on the day they were issued, 14 August 2026, and each says so in its own row with the reason on record: two because the reasoning behind the band was wrong on the evidence then available (the July 2026 MPC minutes), two because the due date fell before or exactly at the instant their source publishes, which made them unresolvable as written. Each was superseded by a new entry that names the one it replaced, and the corrections log records that the first wording of two reasons — “withdrawn before publication” — was itself wrong and was corrected. A sceptic can read that as a do-over; the register is built so that the reading is checkable rather than taken on trust: the withdrawn rows keep their sealed receipts, their original bands and their timestamps, they count as void rather than vanish, and after the 24-hour window a forecast cannot be withdrawn at all — it stands and is scored, hit or miss.

Can “frozen” be checked from outside? Partly, today. Ledger fingerprint sha256:86aa04041be0d596588af86ac1bcf263d21195abdbbd988615bcd4dad7213778 (anchor dated 2026-09-16, 12 entries); not yet anchored outside this site — the operator's next step is a public git commit or an OpenTimestamps proof of this hash, and this sentence changes when it is done. The chain of fingerprints is at /ledger-anchors.json: each names the one before it, so a rewritten file would break the chain; to check the current one, take the SHA-256 of the bytes of /forecast-ledger.json and compare.

These editorial forecasts are separate from the three release-gated decision models and are labelled as such — they demonstrate scoring discipline, not model output.

Why publish predictions at all?

Anyone can claim to be well calibrated after the fact, because memory quietly edits the near-misses into hits. The only defence is a record written before the answer is known and left alone afterwards. Publishing five dated predictions a month, with the resolution source fixed in advance, converts a claim about judgement into something a stranger can check. It also imposes a cost: a bad month is visible permanently, and cannot be deleted or reframed.

How Brier scoring works

The Brier score is the squared difference between the probability given and what actually happened, where the outcome counts as 1 if the event occurred and 0 if it did not. Saying 80% about something that happens scores 0.04; saying 80% about something that does not scores 0.64. Lower is better, 0 is perfect, and 0.25 is what you get by saying 50% to everything. Because the penalty grows with the square of the error, the rule punishes confident wrongness far harder than honest uncertainty — which is exactly the incentive a forecasting record should create.

What would make this page dishonest

Three things, all of which are ruled out by construction: selecting questions after seeing which way they were going; editing a probability once published; and quietly dropping an entry that resolved badly. Each entry is frozen with its publish date and its resolution criterion at the moment it goes up. If a criterion turns out to be ambiguous, the ambiguity is recorded and the entry is scored under the reading published first, never under a more flattering one.