Decision intelligence glossary: every term explained
LEXUN uses precise language on purpose, because vague language cannot be checked. Each term below is defined in one sentence, then explained, then linked to the part of the product where you can check it — including the terms for what LEXUN does not do, such as signing a receipt or restoring a backup.
Decision intelligence
In one sentence: treating a decision as something with stated inputs, an explicit probability, a success test written in advance and a recorded outcome, so the quality of the reasoning can be judged separately from the luck of the result.
Most tools help you organise a decision. Decision intelligence goes further: it commits to a number before the answer is known, and then goes back afterwards to check whether that number was any good. It is the checking that makes it intelligence rather than presentation. LEXUN is not financial, legal or medical advice, and it never decides for you — it makes your own reasoning explicit enough to be checked against what happened later.
Probability band
In one sentence: a range such as 55–65% rather than a single number, where the width of the range is the honest amount of uncertainty.
A model that tells you "61.4%" is claiming a precision it does not have. LEXUN rounds every probability to 5-point steps and then widens the band by two things: the standard error of the simulation itself, and how much of your input was measured fact rather than assumption. A wide band is not the model being unhelpful. It is the model declining to lie about how much it knows.
Base rate and reference class
In one sentence: how often this outcome happens across the broad class of similar situations, before anything specific to you is considered.
The reference class is the group your situation belongs to; the base rate is what usually happens to that group. Starting from the base rate and adjusting is far more reliable than reasoning from your own case alone, because your own case always feels exceptional and usually is not. When LEXUN cannot find a defensible reference class for a question, that is one of the reasons it will decline to answer.
Seeded Monte-Carlo simulation
In one sentence: running the same model thousands of times with randomly drawn inputs to see the spread of possible futures — but from a fixed starting seed, so the same inputs on the same model version always produce exactly the same answer.
Monte-Carlo simulation is old, public and unremarkable; anyone can implement it. The seed is the part that matters here. Because the randomness is deterministic, a LEXUN result is not a one-off performance you have to take on trust — you can re-run it and get the identical output, byte for byte. Reproduce a result in your browser to see this happen.
Evidence gate
In one sentence: a rule that stops the model running at all when there is not enough real information to run it responsibly.
The gates are specific: the model refuses when core inputs are marked unknown, and it refuses when every single input is an assumption rather than a measured fact. This is deliberately inconvenient. A product that always produces an answer is a product that produces answers it has not earned, and the commercial pressure runs entirely in that direction — which is why the refusal is built into the engine rather than left to judgement.
Insufficient evidence
In one sentence: a real, valid output of the engine, not an error message.
When a gate trips, LEXUN says it cannot answer and tells you which input would change that. Treating "I don't know" as a legitimate result rather than a failure is the difference between a forecasting tool and a confidence generator.
Frozen resolution criterion
In one sentence: the test for whether a forecast came true, written down before the outcome is known and never edited afterwards.
Without this, scoring is meaningless: anyone can be right in retrospect by quietly redefining what they meant. Freezing the criterion at prediction time is what makes a later score a fact about the forecast rather than a story about the forecaster. Every entry in the forecast register carries one.
Calibration
In one sentence: whether the things you called 70% likely actually happen about 70% of the time.
Calibration is not the same as being right a lot. A forecaster who says 90% about everything and is right 90% of the time is well calibrated; one who says 99% and is right 90% of the time is overconfident, even though both were right equally often. Calibration can only be measured from outcomes that have actually resolved, and it needs many of them. Resolved outcomes for the decision models so far: 0, so their calibration is not yet measured at all; the editorial forecast register has eight resolved forecasts, each with its own Brier score, which is far too few to say anything about calibration. The Accuracy Centre prints both counts, separately, rather than hiding either.
Brier score
In one sentence: the average of the squared difference between what you said would happen and what did, where 0 is perfect and 1 is as wrong as it is possible to be.
If you forecast 0.8 and the thing happens, the outcome counts as 1 and that forecast scores (0.8 − 1)² = 0.04. If it does not happen, it scores (0.8 − 0)² = 0.64. Averaging that over many forecasts gives the Brier score. The number worth remembering is 0.25: that is what you get by saying "50%" to everything forever, so any scoring system that cannot beat 0.25 has told you nothing at all.
Decision quality versus outcome quality
In one sentence: a good decision can have a bad outcome, and a bad decision can have a good one, so the two have to be scored separately.
If you take a well-reasoned 80% bet and land in the 20%, you were unlucky, not wrong; if you take a reckless 5% bet and it lands, you were lucky, not clever. Judging decisions purely by how they turned out teaches you the wrong lesson in both cases. When you resolve a decision in LEXUN it asks you to separate the reasoning, the execution and the result, because the fix for each is different.
Material change
In one sentence: something has moved enough since you decided that the decision deserves revisiting.
A decision is made against a set of assumptions on a particular day. Those assumptions expire quietly, and the usual failure mode is not noticing. LEXUN records what the decision assumed, so it can flag when an assumption you relied on no longer holds instead of leaving that to memory.
Trust receipt
In one sentence: the sealed summary of one result — probability, band, input hash, seed, model version, frozen criterion and due date — with a SHA-256 hash over those exact bytes, so any later change to them can be detected.
A receipt is made when a result is frozen, and its seal is checked on the verify page, in your own browser, without an account. Your saved decisions also keep an audit record in which every entry is sealed together with the seal of the entry before it, so an entry edited, removed or reordered in the middle is located, not merely detected. What a receipt shows is that the record you are reading is the record that was written. It is not a digital signature and not a trusted timestamp: it does not show who wrote the record or when, and it cannot stop someone who holds the whole record from discarding it and writing a new one.
LEXUN Key
In one sentence: a passcode you choose that encrypts your saved decisions on your own device, which we never receive and cannot recover.
The passcode is stretched into an encryption key using PBKDF2-HMAC-SHA256 at 600,000 iterations, and the record is encrypted with AES-256-GCM. The consequence is worth stating bluntly: because the key never leaves your device, losing the passcode means losing the saved record, and there is no reset link we could honestly offer you. The iterations make each guess slow, not impossible: anyone who copies the encrypted record from your device can try passcodes against it offline for as long as they like, so a short or common passcode can still be guessed. The minimum is eight characters; a long passcode made of several unrelated words is far stronger than a short clever one.
Local-first
In one sentence: the simulator runs and stores your decisions in your own browser rather than on our servers, and the few optional features that do use the network are named, with their state on this deployment.
Running a decision needs no Key. Saving it to your record does, and so does every part of the dashboard that reads that record. A Key is free, needs no card and is not an account: it is a passcode that encrypts your record on your own device, and the decision itself is still never sent to us. Reproducing a published result and reading the accuracy record need nothing at all. Local-first does not mean nothing ever leaves the device. This is the list, rendered from the same registry as the privacy notice at every release:
- ACTIVE Reading the site — Netlify (hosting and CDN): server request logs: IP address, browser user agent, path and time
- ACTIVE Contact, organisation, pilot and Decision Audit request forms — Netlify Forms, then an email to LEXUN: the fields you type on the organisations request form (/organisations, /packages, /decision-audit, /procurement): name, work email, organisation, role, team size, decision type, preferred next step, message; the fields you type on the pilot request form (/pilot): name, work email, organisation, short decision summary in words, business question, deadline
- ACTIVE The encrypted-sync waitlist (/roadmap#sync) — Netlify Forms: your email address
- ACTIVE The optional email beside a new LEXUN Key — Netlify Forms: your email address, if you give one; whether you ticked the marketing box
- ACTIVE The optional Copilot — Anthropic's API, through a LEXUN serverless function: the text you type into the Copilot; the horizon in months and the model version; the engine-produced result summary: band, probability range, scenario labels and percentages, input names and which are assumptions, the resolution criterion and date
- BUILT · DISABLED Buying Pro or Organisation seats — nothing travels on this deployment until it is configured; when it is: the plan and seat count; the email you give at checkout; card details, entered on Stripe's own page and never seen by LEXUN
- BUILT · DISABLED Recovering a paid plan on another device — nothing travels on this deployment until it is configured; when it is: the email address you type
The privacy notice gives every flow its processor, retention and legal basis under UK GDPR, and the security page sets out the architecture.
Hash
In one sentence: a fixed-length fingerprint computed from the exact bytes of a file or result; change one byte and the fingerprint changes completely.
LEXUN uses SHA-256, which writes the fingerprint as 64 hexadecimal characters. A hash answers one question — are these two things byte-for-byte identical? — and nothing else. Anyone can compute the hash of anything, including an altered copy, so a hash is only as useful as the place you got the hash you compare against: a receipt you kept, the forecast register, a copy someone sent you before the question mattered.
Seal
In one sentence: the hash stored with a receipt or package at the moment it is made, so that recomputing the hash later shows whether anything inside has changed.
Checking a seal means hashing the same bytes again and comparing. A match shows the content is unchanged since it was sealed. A mismatch shows it has changed, and the verify page shows where. A seal covers only what was inside it when it was made: that is why the verify page checks the inputs, the audit record and an independent re-run separately, instead of printing one green tick.
Digital signature
In one sentence: a cryptographic mark made with a private key, which shows who produced a file; LEXUN receipts and packages are not signed.
A signature binds content to an identity: only the holder of the private key can make it, and anyone with the matching public key can check it. LEXUN has no accounts and holds no signing key for your results, so a LEXUN seal can show that a file is unchanged, but never who sealed it. Anyone who tells you a LEXUN receipt proves authorship is claiming more than the receipt carries.
Trusted timestamp
In one sentence: a record from an independent timestamping service that a given hash existed at a given moment; LEXUN results do not carry one.
A date written inside a receipt was written by the software that made the receipt, so it is a statement, not evidence. The public forecast register publishes an anchor hash over its entries; the register page says whether that hash has been anchored outside this site yet, and until it has, the register's dates rest on LEXUN's own publication record.
Reproduction
In one sentence: running the named model version at the named seed on the same inputs and getting the identical result, byte for byte.
Reproduction is what the seed makes possible: the randomness is fixed, so the same model version, seed and inputs must give the same answer on any machine. Every release re-runs 96 frozen results and refuses to ship if one of them differs. You can reproduce a published result in your own browser. A result reproduces only on the model version it names; an old result keeps its old version, and a newer model is never used to re-score it silently.
Verification
In one sentence: checking a receipt or package four separate ways — the seal, the inputs, the audit record and an independent re-run — and reporting each answer separately.
The four checks can disagree, and the disagreement is the useful part: a receipt whose seal matches but whose re-run gives a different number has been sealed over something other than what it claims. The verify page runs all four in your browser; nothing you paste is uploaded.
Backup
In one sentence: a copy you can restore from; LEXUN offers readable copies of your record, not backups, because this version cannot import an export back.
The plain export and the sealed .lexun package preserve everything you saved, and the package can be verified anywhere, but the workspace cannot load either of them back in. So an export is a copy you keep, not a restore point. Encrypted sync across devices is on the roadmap as planned, not live. Until something that can restore ships, the honest advice is to export after each decision you freeze and keep the file somewhere safe.
Archive
In one sentence: a dated copy of something that has since been replaced, kept word for word so anyone can read what it said on that date.
Earlier versions of the privacy notice, the terms and the cookies policy are kept under dated addresses and never edited after they are replaced; every past release is on the changelog. An archive is for reading what was true then. It is not a backup and it is not what is true now.
Current deployment state
In one sentence: what is switched on at lexun.co.uk right now, as opposed to what is built but switched off, planned, or retired.
A capability can exist in the code and still be off. Built, and not switched on at lexun.co.uk today: checkout (buying Pro and Organisation seats online). Every feature carries one of those states in a single registry, and the pages that describe features are rendered from it at each release, so a page cannot call something live that the deployment has switched off. The status page shows the current state of each part.
Model
In one sentence: a released, versioned calculation that turns stated inputs into a probability band; its version changes whenever its arithmetic does.
LEXUN releases three models: business runway (version 1.0.2), career change (version 1.1.2) and major purchase (version 1.0.2). Every result names the model version that produced it, a change to the arithmetic is released as a new version with a changelog entry, and results made on an old version keep that version. A language model never does this arithmetic; it can describe a result, but the number comes from the released model.
Template
In one sentence: a guided start that pre-fills the question, the success test and the inputs a kind of decision usually needs, then runs the same released model as everything else.
A template changes where you start, never the arithmetic. Guided templates open a model with worked example inputs for you to replace; frames pre-fill your question and name the narrower question the model actually answers; written guides explain a method with a reproducible example. The templates page counts its own cards at every release.
Editorial forecast
In one sentence: a forecast about a public statistic, written by LEXUN's operator as a judgement, frozen in the public register with its criterion and date, and scored on its own scoreboard.
Editorial forecasts are not produced by the decision models and are not evidence that the models are calibrated. They exist so that the method — freeze the question, the band, the criterion and the date, then score the outcome and publish the misses — is practised in public. The forecast register lists every one, including the withdrawn ones and the reason each was withdrawn.
Decision-model forecast
In one sentence: a probability produced by one of the released models from a person's own inputs, frozen with its criterion and date, and scored only alongside other outcomes from the same model.
This is what the simulator makes when you freeze a decision. Model-domain outcomes and editorial forecasts are never pooled into one accuracy number, because a good record on one says nothing about the other. The Accuracy Centre keeps the model-domain count separately and publishes no calibration figure for a domain until it has enough resolved outcomes to mean something.
Fact, assumption and prediction labelling
In one sentence: every input carries a label saying whether it is something you measured (a fact), something you guessed (an assumption) or a forecast someone else made (an external forecast).
These three get blurred together in almost every spreadsheet ever built, and once blurred, a guess acquires the authority of a measurement. Keeping them distinct is what allows the model to widen its bands when it is running mostly on assumptions, and to refuse outright when it is running on nothing else.
"Planned"
In one sentence: a feature that does not exist yet, labelled so on the page rather than described in the present tense.
Every unshipped capability on this site is marked as planned wherever it appears, and the roadmap lists them in one place. There is no ship date attached to most of them, because each is released when it passes its reliability gate and inventing a date would be the same failure this glossary exists to prevent.