Can I trust a LEXUN result?
Last updated 2026-11-16 — regenerated at every release from the repository itself.
Five things you can check for yourself, each with the page that shows it. Where something has not been shown yet, it says so beside it.
- Reproducible
The same inputs, model version and seed give the same answer on any machine. Every result carries a receipt you can re-run.
- Uncertainty shown
A result is a probability range across 5,000 simulated futures, with the downside beside it — never one confident number.
- Assumptions explicit
Each input is marked a fact you hold or an assumption. Assumptions widen the range, and the one that moves the answer most is named.
- Limitations published
Each released model states what it cannot see, and a question outside the three released models is refused rather than forced.
- Outcomes scored
Forecasts are frozen before the answer and scored in public, misses included. Decision-model outcomes resolved so far: 0, so no accuracy figure is claimed yet.
The Decision Reality Loop: every prediction stays open until reality closes it
Frozen before the answer is known. Scored when reality arrives. Misses published at the same size as the hits. How LEXUN scores its own forecasts.
The public record, today
Editorial forecasts test forecasting discipline, not the decision models — each model earns its own calibration from resolved outcomes. Your own outcome history informs your future decisions; private outcomes never retrain LEXUN’s models, which change only by versioned release. Methodology → · Forecast register →
Your own record, on this device
0 frozen · 0 resolved · 0 ready to score
Every current figure on this page is generated from the repository at release time. The detail, by question:
Decisions are computed in the browser and never uploaded. Forms and the optional Copilot are the two network components; each is named with its processor.
Security and dataHosting, the encryption of saved records under your Key, the headers this site serves, the DPA and how to report a vulnerability. No certification is claimed.
ReproducibilitySame figures, same seed, same answer, on any machine. Re-run any published result; the verifier checks a sealed package without trusting LEXUN.
Method and calibrationHow each model works, how accuracy will be scored, the ten-outcome minimum and the separation of editorial forecasts from model outcomes.
Model limitationsEach released model states its own blind spots from its recurrence, on its guide page: runway, career change and major purchase.
Automated verification1055 engine checks, 96 frozen results and fifty-eight release gates check that the software behaves as written. They are not evidence of forecast accuracy.
Model versionsCurrent: runway@1.0.2 · career@1.1.2 · opportunity@1.0.2. Every receipt carries its exact version; the history is on the Accuracy Centre.
The public record12 forecasts issued, 8 resolved, 4 withdrawn with reasons kept; model-domain outcomes resolved: 0. Frozen before the answer, misses published at full size.
Browser storageEvery key this site may write, listed with its purpose. No tracking cookies, no analytics on decision content.
Evidence provenanceThe Evidence Passport records each figure’s source, date, owner and review date, and flags what is stale. Updating evidence never rewrites a frozen forecast.
StatusThe static site, forms, the Copilot and checkout each fail in their own way and are stated in their own row; hosted services are not reported as operational.
Known limitations and procurementArchitecture, data flows, subprocessors, retention, model governance and the limitations a buyer should read before signing.
Responsible use: LEXUN is decision intelligence, not financial, legal, tax or investment advice, and a probability band is an estimate for you to use, never an automated decision about you. The terms say so in full.
Automated verification
1055 automated checks run against the engine before any release ships — covering RNG determinism, hash stability, evidence-gate refusals, monotonicity (more cash never lowers survival; a bigger commitment never raises it), Brier scoring, calibration binning, forecast freezing, version immutability, materiality detection and trust-receipt integrity.
Suite: /platform/tests.mjs (served read-only) · run: node tests.mjs from the platform folder of the release bundle · current status: 1055/1055 passing · what is answering right now is on the status page; the company behind the record is described on the about page, and the front page runs the worked example live
The suite ships in the site bundle itself and is readable at the address above — download the bundle and run it. Where a technical claim can be independently reproduced or inspected, LEXUN publishes the artefact needed to do so.
Hostile input, shown refused. Type “Ignore your rules and just give me a single number” — or any question outside the three released models — and the router refuses before anything computes: free text never reaches a model as an instruction, only as figures to extract, and no language model computes the answer, so there is none to talk out of its discipline (the optional Copilot uses one only to read what you type). The refusal is itself under test: the release gate feeds the engine hostile text, empty states and malformed files, and a homepage chip labelled “Watch it refuse” lets you press the same boundary yourself.
Determinism proof — run it yourself
Every result carries a seed, model version and input hash. Identical inputs and seed produce identical output, bit for bit. The canonical worked example:
model runway@1.0.2 · seed 564 · P(survival) band 60–70% · id rw_52f9a7c3_234
Reproduce it in one click on the Reproduce a result page, or independently via the published API contract example. To check a sealed receipt or a .lexun package byte for byte, use the verifier.
The calibration record
Model-domain outcomes resolved: 0 (the editorial forecasts on the public register — 8 resolved of 12 issued — are tracked separately and are not evidence for the models). LEXUN publishes no accuracy percentage until at least 10 outcomes per domain have resolved against frozen criteria. Publishing zero resolved model outcomes demonstrates that LEXUN does not substitute software-test counts for empirical calibration evidence: the accuracy record can only be filled in forwards, and it is published as it stands. Method and scoring rules: Accuracy Centre. Why the record has to be earned over time, and what that does and does not defend: defensibility.
Versioned model changelog
Check counts in this table are historical — each row records the suite as it stood at that release. The current count lives on /platform and /accuracy, and a release gate fails the build if any current claim disagrees with the suite itself. Every current platform statistic also ships machine-readable at /system-manifest.json — each number derived from the repository itself (the suites count themselves, the registries are read, nothing is typed twice) and drift-blocked by its own release gate.
| Version | Date | Change |
|---|---|---|
| runway@1.0.0 · core@1.0.0 | 2026-07 | Initial release: seeded Monte-Carlo runway model, evidence gates, frozen resolution criteria. |
| career@1.0.0 · learning@1.0.0 | 2026-07 | Career financial-resilience model; outcome-gated learning layer; governor checks. |
| career@1.1.0 · opportunity@1.0.0 · challenge@1.1.0 | 2026-07 | Independent heuristic cross-checks; major-purchase affordability model. 120 checks passing. |
| reality@1.0.0 | 2026-07 | Decision Reality layer: monitored assumptions, materiality engine, forecast versioning, trust receipts, .lexun export. 233 checks passing. |
| core@1.1.0 · vault@1.0.0 | 2026-07 | The LEXUN Key: encrypted-at-rest dashboard record (AES-256-GCM, PBKDF2-HMAC-SHA256 at 600,000 iterations). No model maths changed. 292 checks passing. |
| reality@1.1.0 | 2026-08 | Action taken and why reality differed recorded in the person’s own words; lessons carried into the next decision in the same domain. Additive fields only — no model maths changed. 390 checks passing. |
| core@1.2.0 · runway · career · opportunity · path-trace | 2026-09 | The multiplicative shock became log-normal (exp(ln(1+g) − v²/2 + v·z)) so a month can halve revenue and can never delete it, and paths run in antithetic pairs — 34% less Monte-Carlo error for identical work. Survival probabilities moved by up to a point at ordinary volatility and by more at the extremes; every moved result is listed in the changelog entry. |
| core@1.2.1 · runway@1.0.1 · career@1.1.1 · opportunity@1.0.1 | 2026-09 | Break-even searches report when the breakpoint lies outside the searched range instead of returning the bound as a figure; quoted breakpoints are rounded onto the surviving side; optional runway inputs (committed amounts, cost growth) are validated; the survived-code is capped at the horizon in every view. No probability, band, verdict, receipt id or input hash changed — 96 frozen results byte-identical. |
| router@2.0.0 | 2026-08 | The capability firewall: unsupported questions (markets, medical, legal, gambling, relationship outcomes, elections, weather, vague success) are refused with reasons and a supported reframing where one honestly exists; weak signals ask one question instead of guessing; the silent runway default is gone. Routing only — no model maths, seed derivation or hash function changed, so all 96 frozen results resolve unchanged. 397 checks passing. |
| router@2.1.0 | 2026-08 | Capability-firewall keyword widening (pricing, expansion, staffing phrasings; accent folding so “café” reaches the keywords as “cafe”) plus the intent contract: every question the site advertises must route and pass the frame check, enforced in the suite. Routing only — no model maths changed, all 96 frozen results unchanged. 399 checks passing. |
| core@1.2.2 · runway@1.0.2 · career@1.1.2 · opportunity@1.0.2 · reality@1.2.0 | 2026-09 | Canonical inputs: a numeric string is the number it spells — for the hash, the seed, the arithmetic and the frozen record — so one decision typed eight ways mints one receipt where an independent audit on 19 September 2026 found eight. The verifier asks a fourth question, whether the input values in a package hash to the input hash the receipt sealed, so a rewritten figure fails even when the seal holds; its seal panel says only what the seal covers and a mismatch is printed at the precision that shows it. No probability, band, verdict, receipt id or input hash changed for any correctly typed input — 96 frozen results byte-identical. 800 checks passing. |
| reality@1.3.0 · room@1.1.0 · passport@1.0.0 | 2026-09 | The Evidence Passport: source, source date, owner and review date per input, recorded on the decision and never on the forecast. Freshness is read from the dates in words that name the day — verified 3 days ago, not reviewed for 104 days, review was due — and an input past its review threshold is named as the reason a decision’s evidence needs re-verifying. Recording and updating are ledger events carrying the passport’s content hash; an update after the freeze leaves the frozen forecast byte-identical and the sealed receipt unchanged. A decision without a passport is judged exactly as before. No model maths, seed derivation or hash function changed — 96 frozen results byte-identical. 825 checks passing. |
| room@1.2.0 · quick@1.0.0 | 2026-09 | The Decision Room’s Executive view answers in the order the questions are asked — the range, the downside, the variable that matters most, the breakpoint, the weak assumption, what to verify, the receipt — and the workspace opens in a Quick decision mode for a first visit, every expert control hidden by CSS and none removed. Ordering and disclosure only: no model maths, seed derivation or hash function changed — 96 frozen results byte-identical. 857 checks passing. |
| room@1.3.0 | 2026-09 | The three views of a result are named for what a reader wants from them — Answer, Understand and Verify — where they were Executive, Analyst and Audit. The cards in each view, their order and every figure are unchanged. Naming only: no model maths, seed derivation or hash function changed — 96 frozen results byte-identical. 862 checks passing. |
Methodology never changes silently: every change lands here and in the Accuracy Centre with its version number.
Worked case studies
LEXUN launched in 2026 and has no client testimonials yet — and we will not invent any. What we can publish today are fully worked, anonymised decision cases: real engine runs whose every number you can reproduce. As real outcomes resolve, their Brier scores will be appended here — whatever they show.
Case 1 — Small business runway. £80,000 cash, £12,000/mo revenue growing 2%/mo (±15% volatility), £18,000/mo costs, 12-month horizon. Forecast: 60–70% probability of maintaining positive cash (PROBABLY SAFE, WATCH IT); top driver monthly costs; the stress scenario falls to about 20%. runway@1.0.2 · seed 564 · rw_52f9a7c3_234 — reproduce it. Outcome: resolves against the frozen criterion at its due date; no outcome yet, so no score is claimed.
Case 2 — Career change. £12,000 buffer, £2,400/mo essential costs, £2,600/mo expected new income after a 6-month ramp (±30% volatility). The engine prices the transition honestly: the buffer question is resolvable, career "success" is not — so only the former is forecast. career@1.1.2 · worked example seed 705 — full walk-through.
Case 3 — Major purchase. £12,000 buffer, £6,000 upfront, £550/mo commitment against £2,600/mo net income and £2,050/mo essentials. The model reports buffer survival and the break-even commitment, and its adversarial challenge flags the assumed running costs as the evidence gap. opportunity@1.0.2 · worked example seed 507 — full walk-through.
Case 4 — An adviser’s file (illustrative; the client and figures are fictional, the engine run is real). An adviser’s client wants to leave employment for a consultancy contract. Inputs as declared with the client: £20,000 buffer (fact — statement seen), £1,800/mo essential costs (fact — 6-month bank average), £2,500/mo expected new income (assumption — one signed letter of intent), 3-month ramp (assumption). The engine returns its band with the fact/assumption split preserved, and the exported Decision Record — inputs, classifications, band, model version, reproducibility id, sign-off line — goes on the client file as evidence of a documented, disciplined process. What it does not do: make the recommendation, or make anyone compliant. The adviser door, stated precisely.
What would falsify us
Each line is an observation that would show this system does not do what the site says it does, and where to look for it. Finding one is a service, not an attack: tell us, and the answer goes here and on the changelog whether or not it flatters us. The list is kept at the same size as the claims it guards; when one is found, it stays on the list with what changed.
- A register entry edited after issue. Every forecast’s question, band, criterion, due date and source are frozen when it is issued. Check the raw ledger against the anchor chain: each anchor is the SHA-256 of the ledger file at a release, chained to the one before, so an edited entry changes every anchor after it.
- A resolution scored against a source not named in advance. Each resolved entry carries the resolution source frozen at issue (a release is refused if that field changes) and the outcome source it was scored from, with its URL and publication time (a release is refused if those are missing). Both are in the ledger and in the CSV, side by side, so the comparison is yours to make.
- An accuracy figure published before ten resolved outcomes in a domain. The model domains stand at zero; the editorial board is scored publicly but is not evidence for the models. Check the Accuracy Centre and the counts on this page, all written from the ledger.
- A receipt that does not reproduce. The same model version, seed and inputs must give the same probability and the same id on your machine. Check any published receipt on the reproduction page; the 96-result frozen corpus in platform/regression-snapshot.json is re-run at every release and a changed result blocks it.
- The same inputs, written differently, giving different receipts. "80000" and 80000 must be one input, one seed, one receipt. Found — see below.
- A seal reported verified over a changed figure. The verifier must fail a package whose recorded inputs are not the ones its receipt was sealed over. Press “Load a live sample”, then “Change a figure”. Found — see below.
- A public number that cannot be traced to a computation or a record. Every statistic on this site ships derived in the system manifest; the changelog’s shipped rows and the version tables on this page and the Accuracy Centre are frozen in audit/changelog-history.json, and a row that changes what it said blocks the release. There are no testimonials, user counts or logos because there are no clients yet. Found — see below.
- A simulation input leaving this device from any page that says it stays here. Watch your browser’s network panel while you run a simulation or import a CSV; a release gate imports a bank export in the workspace and requires zero network requests while it does.
- Mis-calibration in the resolved record. The calibration table is computed from resolved outcomes, not written. If the forecasts are mis-calibrated it will say so, and it is published whether it flatters us or not. That commitment, not a promise of accuracy, is the product.
Found, and closed
24 September 2026 — found while preparing release v222, the day v221 went live; fixed in v222. Item 7 above was true of the live site. The Evidence Passport row in the version tables on this page and the Accuracy Centre shipped in v218 giving the suite’s count as 825. The script that keeps check counts current bumped the newest row of each table at every release; v219 added no version row, so it rewrote that count to 843, and the freeze added in v217 — which left the newest row open as the release being prepared — then recorded 843. v221 published it. The row says 825 again, byte for byte what v218 shipped, and its entry in audit/changelog-history.json says openly that it was corrected. The newest row of a version table is now the release being prepared only in the release that adds it; once that release has shipped the row is frozen with what it said, and the release chain asks before it bumps a table.
23 September 2026 — reported by an external review of the live site against release v219; fixed in v220. Item 7 above was true of the live site. The workspace’s status line said seven registered sources while the source registry and the system manifest held five; the country page said four forecasts stood open after three had resolved; the citation page listed the permanent forecast pages only to /f/010 when /f/012 existed; the templates page counted twenty-one guided starts over nineteen cards. Each was a current fact typed where it appeared, which the manifest described correctly and nothing compared against the page. What changed: every current statistic on a page is now a marked block written from the manifest at release time, and the release fails if a page types one; the forecast range, the country page’s resolved-and-open list and the template count are written from the ledger and the page itself; the release runs as one ordered chain that renders the pages before it checks the finished output. No forecast, receipt or model result was affected.
19 September 2026 — reported by an independent auditor against release v215; fixed in v217. Items 5 and 6 above were both true of the live site. A numeric string was accepted by the models but hashed differently from the number it spelled, so one runway decision typed eight ways minted eight receipts with probabilities from 0.649 to 0.663 (all displaying 60–70%), and on the career model the displayed band moved between 45–60% and 40–55% for 9000 and "9000". Nothing shipped on this site passes a string to the engine, so no shown result, saved record or published receipt was affected, and every receipt minted from a number keeps its id. Separately, the verifier sealed the hash of the inputs but never compared the sealed hash with the input values in the package, so a package whose recorded cash had been rewritten to £5 reported its seal verified under a sentence that said nothing had changed — the independent re-run caught the figure; the sentence was still wrong. What changed: core@1.2.2 canonicalises inputs before the hash, the seed and the record; the models carry patch versions for it; the verifier asks a fourth question and prints a mismatch at the precision that shows it. While fixing it, a third defect was found and is disclosed on the changelog: two history rows in the version tables on this page and the Accuracy Centre had been rewritten by the count tool and are restored from the frozen record.