Accuracy Centre

The business cash-runway model answers one resolvable question: will cash stay above zero for the stated horizon?

• Simulation: 5,000 monthly cash paths, run in antithetic pairs. Revenue follows a log-normal multiplicative step about the user’s stated growth rate — exp(ln(1+g) − σ²/2 + σ·ε), ε ~ N(0,1) — whose expected value is exactly 1+g at any volatility and which is strictly positive, so no month can drive revenue to zero; costs grow at the stated rate; committed one-off cash events land in their stated months.
• Determinism: the random seed is derived from the inputs (or supplied explicitly), so identical inputs always produce identical results. This is enforced by automated tests.
• Uncertainty display: probabilities are shown as bands rounded to 5-point steps, widened by Monte-Carlo standard error and by input quality. Raw values are stored only for calibration scoring.
• Evidence gates: the model refuses to run when core inputs are marked unknown, or when every input is an assumption. It does not fabricate values.
• Confidence: derived structurally from how many core inputs are facts versus assumptions — never from model enthusiasm.
• Adversarial review: a separate challenge module (different method: straight-line burn heuristic plus a rule library) attacks every result; material disagreement between the two models is recorded and shown.
• Resolution: the outcome criterion and due date are frozen at prediction time and cannot be edited afterwards. Outcomes are scored with the Brier rule, which is defined with the rest of the vocabulary in the decision glossary.

The career-change model answers only the resolvable question: if you make the change, does your financial buffer survive the income transition? Income ramps linearly to your expected level over your stated months, with monthly volatility shocks; essential costs drain the buffer; 5,000 seeded paths. Same disciplines as runway: deterministic seeds, evidence gates, banded display, frozen resolution criteria. Version 1.1 delivers the independent-heuristic cross-check promised at 1.0: a deterministic straight-line drain model (separate code path, no Monte-Carlo) attacks every result, and material disagreement between the two is recorded and shown.

Stated honestly: no defensible base rate exists for career-change success (definitions vary too much to pool), so this model does not offer one. Context figure only: ~2.9m UK workers changed jobs in 2025 (Indeed Hiring Lab, citing ONS). Known weaknesses: partner income, benefits during transition, emergency costs and the return-to-employment option are not modelled.

You will see this model called major purchase in the product and opportunity@1.0.0 in receipts, version tables and the changelog. They are the same model: opportunity is the internal identifier it was released under, and renaming it would break every reproducibility id already issued against it.

The major-purchase model answers only the resolvable question: if you take on this upfront cost and recurring commitment now, does your financial buffer survive the horizon? The upfront cost leaves the buffer at month zero; thereafter income varies with monthly volatility shocks against fixed essential costs plus the commitment; 5,000 seeded paths. It also reports the largest monthly commitment at which your median path still survives, and a waiting analysis that honestly notes it does not model price rises while you wait. Same guarantees as the other models, including a paired independent surplus-arithmetic cross-check.

Stated honestly: no defensible base rate exists for whether a major purchase "works out" (outcomes vary too much to pool), so this model does not offer one. Context only: 13.1 million UK adults (24%) had low financial resilience in May 2024, and 42% could not cover 3+ months of living costs if their main income stopped (FCA Financial Lives 2024). Known weaknesses: price and interest-rate changes while deciding, one-off repair shocks, resale/exit value, and essential-cost inflation are not modelled.

• Calibration by domain (runway, career, opportunity) and by horizon bucket (0–3m, 3–12m, 12m+) — never one universal percentage.
• Reliability tables (mean forecast vs actual frequency per probability bin), sample sizes, and confidence intervals.
• Per model version, so improvements and regressions are visible across versions.
• Retractions and corrections, listed on this page permanently.

An honest forecasting record has to keep three different times apart: when something happened, when the figure about it was published, and when this register learned it. Mixing them is how systems quietly cheat — a “prediction” scored against a revised figure the forecaster could never have seen is not a prediction.

Every ledger entry carries the three clocks as separate frozen fields, and the resolution tool enforces them: issued_at and data_cutoff record what the forecaster could know and when; source_publishes_at records when the deciding figure becomes public — resolving before it is refused outright as “a guess wearing a timestamp”; and the outcome records source_published_at and resolved_at with the figure as first published, so a later revision can never rewrite a score. A worked example from the register: the July 2026 CPI forecast was issued 14 August with a cutoff of the same day, its source published 19 August at 07:00, and it resolved that morning against the first-published 2.9% — the criterion says in terms that later revisions do not change the resolution.

The reference-class figures shown beside results follow the same rule from the other side: each is a bundled snapshot with its publication and retrieval dates displayed, marked as context that does not enter the calculation — and never described as live.

Outcome verification grades

Not all resolutions are equally strong, so every resolved outcome now carries a verification grade, computed from the evidence the ledger actually holds — never assigned by hand. A means automatically resolved from an authoritative source by a connector: impossible in this architecture, because no connectors exist, so no outcome can carry it yet and none does. B means manually resolved against an authoritative published figure — the exact first-published sentence, its URL on the frozen source host and its publication time all sealed in the ledger; all six resolved outcomes on this register are grade B. C and D cover corroborated and unverified user reports — private workspace outcomes, which never enter this public record. E is unresolved or ambiguous. The grades, with each outcome’s grade and the full model coverage and release registry — every model LEXUN ships, every model it merely proposes, and the wall between them — ship machine-readable at /model-registry.json, derived from the model code itself and drift-blocked by its own release gate.

Most forecasting systems keep one history: what happened. An accountable one needs three, kept apart. Reality history is what later evidence says happened — here, every resolved outcome as first published, sealed with its source sentence and URL, never rewritten by revision. Knowledge history is what could genuinely be known at the time — every entry’s data_cutoff, the world-state snapshot frozen into each decision, and the receipt that proves nothing later leaked in. Expectation history is what LEXUN believed would happen next — and it is the one histories quietly lose, because keeping it means leaving your worst calls on display.

This register keeps it. Every probability band ever issued stays published; a change of mind is a new entry that names the old one and the reason, never an edit. The table below is that expectation record, regenerated from the ledger at every release — the corrections with the reasons in the record’s own words:

12 entries ever issued · 4 since withdrawn · 0 deleted · 0 edited. A register that could quietly rewrite its expectation history would have nothing to be scored against.

What this page does not hold: an archive of other forecasters’ expectations — IMF and central-bank projection vintages, professional survey rounds. That is real evidence and a real gap, recorded as such in the gap registry rather than papered over; until it closes, the expectation history here is LEXUN’s own, complete and uncut.

What the engine writes for every run, in the order it writes it. Every field below is live: it is filled by the engine or by you, and nothing on this list is planned or partial. See the exported JSON →

• The question, as you asked itLive
• Every input, classed fact or assumptionLive
• Probability band and confidenceLive
• Resolution criterion and due date, frozen at prediction timeLive
• The outcome, written once when reality answersLive
• Brier score and the domain calibration it updatesLive
• The action you actually took, in your wordsLive
• Why reality differed, in your wordsLive
• Lessons carried into the next predictionLive

The example on the homepage is a real engine run. Anyone entering the same inputs reproduces it exactly:

Run it yourself →

The registry is also shipped machine-readable at /source-registry.json — generated from the same code the engines load (never hand-typed twice), with each source’s authority tier (1 official · 2 recognised research · 3 reputable industry · 4 secondary reporting), licence, publication and retrieval dates, and the register’s resolved-outcome sources. It states plainly that this static site has zero live data connectors: every figure is a bundled, dated snapshot. A release gate blocks any drift between the shipped file and the code. The file also carries a coverage matrix (every figure LEXUN actually holds, with tier, dates and where it is used) and a blind-spot register (what it lacks, classified honestly — from “discoverable and open” to “unavailable in the current architecture” — each with why and the acquisition path).

No source is ever described as "live" unless an actual connection exists and has recently refreshed. Today, none are live; all reference data is bundled snapshots with retrieval dates.

The first Historical Intelligence dataset: the Bank of England’s full official Bank Rate change history — every change since 1975, bundled with its source, licence and retrieval date, and six anchor values cross-checked against independently known history before bundling. It exists to answer one honest question: is today’s rate environment ordinary or unusual by the record’s own standard? No released model takes the rate as an input, so this series is context beside results, never inside them.

LEXUN groups the record into named bands — near-zero (below 1%), low (1–3%), moderate (3–6%), restrictive (6–10%), extreme (10% and above). The thresholds are a naming convention, stated so you can disagree with them; they are not analysis.

Two variables against history — an honest analogue

With two bundled series — the Bank Rate record and the ONS CPI yearly averages (37 years, 1989–2025, shipped at /data/ons-cpi-annual.json, current vintage only and the file says so) — LEXUN can ask a narrow, honest question: which past years most resembled today on these two variables alone? Today’s pair is real on both sides: the 12-month CPI rate as resolved by the public register against its first-published source (July 2026: 2.9%), and the current Bank Rate from the bundled record. One stated mismatch: today’s 12-month rate is compared with calendar-year averages — said here rather than hidden.

Two variables are not a regime. A real regime comparison needs many more series — unemployment, credit, wages, housing, energy — which this static bundle does not hold; the blind-spot register says exactly that. And even a full state vector would not make history a prophecy: the years listed above went on to different futures for reasons these two numbers never carried.

Similarity is not repetition. This series records where the rate has been; it says nothing about where it will go. Two periods that look alike in one variable can end differently for reasons no single series carries — which is exactly why the register freezes forecasts and scores them, rather than reading history as prophecy.

Forecast science starts with a humiliating rule: before any model earns trust, it must beat the do-nothing line. Below, the simplest possible CPI forecast — next year’s average will equal this year’s — is walked forward through the bundled ONS series with no hindsight: each year’s prediction uses only the previous year’s figure, exactly as a forecaster standing in that year could have. Computed in your browser from the same file you can download.

What this exhibit is for. In quiet years the naive line is very hard to beat — and in regime breaks it fails catastrophically, which is the honest shape of the forecasting problem. Any model LEXUN ever promotes for a variable like this must beat this baseline out of sample, and the years in the misses table are why “history as prophecy” is refused everywhere on this site: the past is the best guide available right up until the moment it is the worst one.

• Revenue shocks are modelled as smooth monthly volatility; single lumpy events (losing an anchor client, a tax bill) are only captured if entered as committed cash events.
• Receivables timing (booked vs banked revenue) is not modelled.
• The bundled reference class pools all UK industries and sizes; sector-specific base rates are not yet included, and the 1-year survival rate was not captured in our snapshot, so it is not shown.
• Outcome data is self-reported until verified-evidence attachment ships; self-reporting can bias calibration optimistically.
• No calibration history exists yet — the model's real-world track record is unproven, which is exactly what the outcome loop exists to fix.

LEXUN models business and career decisions. It refuses medical, legal and immigration questions, routes personal-crisis language to human help, and provides decision intelligence — not financial, legal or medical advice.