Anyone can copy the interface. The scoring record has to be earned in public.
The screen is imitable in a weekend. The asset underneath starts accruing on day one and can only accrue forwards. This page sets out what that asset consists of, what part of it exists in the code today, what part only accumulates with elapsed time, and what is still unbuilt. Where the honest number is zero, the zero is printed. If you would rather see the difference than read about it, run one simulation and watch it delivered two ways.
The state of the record today
Sources, in order: the first public forecasts were issued to the public forecast register on 14 August 2026, so the record was 94 days old on 16 November 2026, the day this release was built; the count is elapsed time, not an estimate. Of 12 editorial forecasts issued, 8 have resolved against their frozen criteria and 4 were withdrawn with their reasons kept; that board is scored publicly and is not evidence for the decision models. Outcomes for the decision models resolve in the visitor’s own browser and are not transmitted, so the model-domain count we can honestly publish is 0. Re-runs happen on other people’s machines with nothing reported back, which is a deliberate architectural choice and therefore permanently uncountable — reporting a number here would mean we had started collecting one. The ten shipped model versions are runway@1.0.0, 1.0.1 and 1.0.2; career@1.0.0, 1.1.0, 1.1.1 and 1.1.2; opportunity@1.0.0, 1.0.1 and 1.0.2, each dated in the changelog.
What the defence consists of
Each item below carries one of three labels. They mean different things and are not interchangeable.
LIVE Working in the shipped code now — you can check it yourself today. COMPOUNDING The mechanism is built; its value is a function of elapsed time and currently near zero. PLANNED Not built. Named here so it cannot later be claimed as though it always existed.
Compounding track record
COMPOUNDINGEvery forecast is written with its resolution criterion frozen before the result is seen, and scored with a Brier score once the criterion resolves. The machinery is live and the register opened with ten frozen forecasts on 14 August 2026; the scored record accumulates only as criteria resolve, and the register page shows its current state.
Why it is hard to copy: a competitor shipping tomorrow also starts at zero resolved forecasts. History is the one input that cannot be bought, hired, funded or accelerated. A later entrant can issue more forecasts per month, but it cannot back-date the ones it did not make.
Reproducibility as a standard
LIVEEvery result carries the model name and version, the frozen inputs, the random seed and a SHA-256 receipt over the canonical bytes of the whole thing. Anyone can re-run it on the same model version and get the identical number, or paste the receipt into the verifier and watch their own browser recompute the hash.
Why it matters even though it is easy to copy: the format is trivially copyable, and that is fine. A result anyone can re-run is easier to cite than one nobody can check. Whether the format becomes something people cite is not ours to claim; it depends on whether they do.
Calibration flywheel
COMPOUNDINGEach resolved outcome on the public register sharpens the calibration LEXUN publishes for that domain, and each outcome a person resolves in their own workspace sharpens the calibration they see — on their own device, where it stays. What does not happen, and cannot with this architecture, is a private decision improving the model for anybody else: saved decisions and outcomes are encrypted locally and never transmitted as records (the optional Copilot receives only what a person types or a filtered result summary), so there is no corpus of customer outcomes here and no flywheel that turns usage into a better global prior. Saying otherwise would be the one claim on this page that the privacy notice forbids. What compounds is the public record and each user’s own history — and if consented, anonymised outcome aggregation is ever built, it will be announced as a change to what leaves your device, not shipped as a side effect. The Accuracy Centre refuses to display an accuracy figure for any domain with fewer than ten resolved outcomes, which today means every domain.
Why it is hard to copy: the improvement loop runs on resolved outcomes, and resolved outcomes only exist after real decisions have played out in real time. Capital shortens almost every other path in software. It does not shorten this one.
Local-first by architecture
LIVEThe simulation runs in the visitor's browser. Decision data is held in that browser's own storage and is never transmitted; the optional Copilot is the one exception, and it receives only what a person deliberately types or a filtered result summary. Running one needs no Key; keeping one needs a LEXUN Key, which is a locally derived passcode rather than an account — there is no server session, no user directory and no server-side copy of any decision to subpoena, leak or sell.
Why it is hard to copy: this is structural, not a policy line in a privacy notice. Anyone running inference server-side cannot retrofit "there is no server copy" without rebuilding the product, and any competitor whose business model depends on holding the data cannot make the claim at all.
Accumulated context, not lock-in
COMPOUNDINGAssumption registers, decision records and audit trails deepen over years in the workspace of whoever keeps them — today one person’s, on one device; shared team workspaces are planned, not live. All of it exports as a .lexun package, and the package is documented rather than obfuscated, so taking it elsewhere costs one click.
Why it is hard to copy: the fact that the exit is free is why a person or a team can afford to put years of real reasoning in. A rival can copy the export format in an afternoon and still not have the years.
Versioned and auditable
LIVEEvery result carries the version of the model that produced it. Audit events are hash-chained, so an entry cannot be edited, reordered or removed in the middle without the chain reporting where. The chain does not prove when an entry was written, and whoever holds the whole record can discard it and start again; it shows the record read is the record written. The engine ships with 1055 automated checks, all currently passing, including tests that deliberately tamper with a sealed record and assert that the tamper is caught.
Why it is hard to copy: buyers with a compliance function commonly ask for this before a number can go into a decision they have to defend. It is a discipline that has to be demonstrated over releases rather than claimed in a sales deck, and the demonstration takes exactly as long as it takes.
Public forecast register
LIVEDated, frozen, publicly scored predictions, each with its resolution source fixed in advance and misses left up permanently. The register opened with ten on 14 August 2026 and a further two on 30 August 2026; it holds 12 — 8 resolved, 0 open, 4 withdrawn with their reasons kept. The counters above read straight from the ledger, misses included, and grow only as new cohorts are issued and criteria resolve.
Why it matters here: this is the mechanism that converts elapsed time into the track record described at the top. It has been publishing since 14 August 2026 — 94 days at this release — and the compounding items on this page are worth exactly what it has resolved so far.
What is not a moat
The mathematics. Monte-Carlo simulation is public, well over half a century old, and implementable by a competent engineer in a day. The mulberry32 generator and the Box–Muller transform this engine uses are textbook. Nothing about the numerical method is proprietary and we do not claim it is.
Nor is the interface, the wording, the colour palette, the animation or the brand. All of it is copyable, some of it in a weekend. The defence is scoring, receipts and elapsed time.
What would break it
Stated plainly, because a defensibility page that lists only strengths is marketing.
- Publishing forecasts and then quietly failing to resolve them. The record only compounds if the losses are scored as diligently as the wins.
- Moving inference to a server for convenience or for a feature that seems worth it. That would forfeit the local-first claim permanently, and no amount of policy language would restore it.
- A competitor starting the same public scoring discipline and being more rigorous about it. This is the honest competitive risk, and there is no technical barrier stopping them.
- Loosening the display gates under commercial pressure — showing an accuracy figure on a handful of resolved outcomes because a customer asked for one. The gates are only worth something while they are inconvenient.
Check any of this yourself
None of the above should be taken on trust, which is the entire argument. Verify a receipt or a package in your own browser. Re-run a published result from its seed and inputs and compare the identifier. Read the honesty rules and their display gates, the dated version history, or the trust page for what is measured and what is not. If you find a claim on this site that is not backed by something you can check, tell us and it comes down.