Deep Dive: Normal Distribution Standardisation and Conditional Probability

#CODEX Deep-Dive — The Normal Distribution: Standardisation and Conditional Probability (WST01, spec 6.1/3.2)

Provenance (2026-09-29): Read/checked for this document — research/veridian/WST01-verified-facts.md in full, then re-grepped specifically for "normal distribution", "standardis", "Φ(z)", "percentage points" and "conditional probability" (§1 spec transcription and its verified spec-typo note, §2a formula-booklet contents, §3 source archive, §5 Section 6 trap catalogue, §6 cross-cutting "show that" pattern, §7 topic-richness ranking, and the 2026-09-12 per-lesson audit entry for this exact lesson); the live lesson file content/veridian/curriculum/wst01/normal-distribution-standardisation-and-conditional-probability.ts in full; content/veridian/technique-hub/wst01/codex-formula-and-method-bank.md, forge-exam-technique-guide.md and signal-frequency-and-boundaries.md (grepped for this topic); content/veridian/question-bank/wst01.ts (grepped — no item on this topic exists there yet). No new PDF fetches were made for this document; every specific claim below traces to one of those sources, named inline. All arithmetic in the micro-examples was independently checked (linear-transformation algebra by hand, the two trivial conditional-probability cases by set-containment logic, and the two numerical Normal-CDF checks against the error-function definition of Φ).

#Scope — what this adds, and what it deliberately does not repeat

The live lesson already builds a full four-part worked exemplar (plant-height, $H\sim N(50,8^2)$) covering the subtract-from-1 step, an interval probability, a one-unknown inverse lookup, and one conditional-probability-with-Normal case; a separate worked-chain for the two-unknown simultaneous-equations question type (light-bulb lifetime); a one-unknown inverse chain (puzzle-completion time); a chain-drill closing the non-nested conditional-probability case (cereal-box fill-weight, added in the 2026-09-12 audit); five MCQs; and mechanism prose on why the sign of $z$ has to be reasoned rather than read off, and on "show that" discipline for this topic specifically. The formula bank already tabulates the given-vs-memorise split for every relevant formula and gives a short method template for each Section-6 skill. The forge guide already gives the self-mark checklist and the calculator/table-value convention (§3.3 there). The signal doc already gives this topic's frequency ranking (5 of the 6 fully-reviewed examiner reports; 17 of 75 marks in the single largest question of the Jan 2023 paper). None of that is reproduced here.

This document adds four things not assembled anywhere else in the repo: (1) a derivation, rather than an assertion, of why standardising forces mean 0 and variance 1, by tying spec 5.3's own linear-transformation rules explicitly to the spec 6.1 claim; (2) a derivation of the symmetry identity $\Phi(-z)=1-\Phi(z)$ from spec-required shape/symmetry knowledge, and the observation that the formula booklet's own table domain makes this identity structurally necessary rather than a convenience; (3) a complete four-way structural taxonomy for a conditional-probability question built from two Normal-threshold events, of which the live lesson drills two, with short micro-examples closing the other two and an honest flag on their evidence status; (4) a precise distinction — with a specific evidence-boundary flag — between the two different "hardest skill" claims the facts bank makes about this topic, which the live lesson's own header prose runs together.

#1. The formal model: standardisation and its symmetry, derived rather than asserted

#1.1 What spec 6.1 actually licenses and what it rules out

The Normal model is $X\sim N(\mu,\sigma^2)$: mean $\mu$, variance $\sigma^2$ (spec 6.1, verified facts bank §1, noting the printed spec's own numbering oddity — "6. The Normal distribution" over a content row numbered "5.1" — a confirmed Pearson typo, not a transcription artefact). The guidance box explicitly requires knowledge of shape and symmetry, and explicitly rules out three things: the probability density function itself, any derivation of the mean, variance or CDF, and interpolation within either table (facts bank §1). The standardised variable $Z=(X-\mu)/\sigma$ is one of the handful of relationships the spec's own notation-and-formulae box lists as expected to be known from memory rather than looked up (facts bank §2a) — unlike the general conditional-probability identity, which the formula booklet does print.

The formula booklet (Issue 2, January 2021) supplies exactly two Normal-distribution tables: the Normal Distribution Function table, $\Phi(z)$ for $z=0.00$ to $4.00$, and the Percentage Points table, giving $z$ such that $P(Z>z)=p$ for a fixed list of $p$ values (facts bank §2a). Both are printed for non-negative $z$ only. That detail matters directly for §1.3 below.

#1.2 Deriving $E(Z)=0$ and $\mathrm{Var}(Z)=1$ from spec 5.3, not asserting them

The live lesson states that standardising "converts any Normal random variable into the standard Normal $Z\sim N(0,1)$" without deriving why the mean and variance come out as 0 and 1 specifically. That derivation is available directly from a formula this same course already establishes elsewhere in the same unit — spec 5.3's linear-transformation rules for a random variable, $E(aX+b)=aE(X)+b$ and $\mathrm{Var}(aX+b)=a^2\mathrm{Var}(X)$ (verified facts bank §2a as two of the formulae the spec expects memorised, not looked up). Those rules are stated under the discrete-random-variable spec item, but nothing in their own statement or proof restricts them to discrete variables — they are general properties of the expectation and variance operators, which is exactly why the same shift-and-scale logic works again here for a continuous variable.

Write $Z=(X-\mu)/\sigma$ in the form $Z=aX+b$ with $a=1/\sigma$ and $b=-\mu/\sigma$. Then:

$$E(Z)=aE(X)+b=\frac{\mu}{\sigma}-\frac{\mu}{\sigma}=0,\qquad \mathrm{Var}(Z)=a^2\mathrm{Var}(X)=\frac{1}{\sigma^2}\cdot\sigma^2=1.$$

So the mean-0, variance-1 property of $Z$ follows purely from the general linear-transformation rules — nothing specific to the Normal family has been used yet. What genuinely is specific to the Normal family, and is deliberately left as given rather than derived here, is the further fact that $Z$'s distribution is itself Normal, not merely some distribution that happens to have mean 0 and variance 1 — a closure property of the Normal family that spec 6.1 explicitly does not require a student (or, by the same logic, this document) to derive: derivation of the CDF is named outright as not required.

#1.3 The symmetry identity, derived from spec-required knowledge, and why the table's own shape makes it mandatory

The spec guidance requires "knowledge of the shape/symmetry" of the Normal curve. That single requirement is enough to derive the identity every sign-of-$z$ decision in this topic depends on. The total area under the standard Normal curve is 1; the area to the left of $z$ is $\Phi(z)$ by definition, so the area to the right of $z$ is $1-\Phi(z)$. Because the curve is symmetric about its mean (which sits at $z=0$ once standardised), reflecting the region "to the left of $z$" through 0 produces the region "to the right of $-z$" with identical area, and reflecting "to the right of $z$" produces "to the left of $-z$." Reading the second reflection off:

$$\Phi(-z)=1-\Phi(z).$$

This is not merely a convenient shortcut. Because the printed $\Phi(z)$ table's own domain runs only from $z=0.00$ to $4.00$ (verified, facts bank §2a) — there is no row anywhere in the booklet for a negative $z$ — any probability whose standardised value works out negative has no direct entry to read. The symmetry identity is the only route from a negative $z$ back into the printed table's domain, which is exactly why the live lesson's own sign-of-$z$ mechanism block treats getting the sign wrong as a fully checkable error (an answer that lands outside a stated boundary cannot be right, whatever the working looked like getting there) rather than a matter of taste. Framed this way, the rule is a direct structural consequence of how the table itself was built, not an extra convention layered on top of it.

#1.4 The two lookups, briefly

Everything past this point in a standardisation calculation is mechanical: a forward calculation reads $\Phi(z)$ directly off the Normal Distribution Function table; an inverse calculation (a stated probability, an unknown boundary) uses the Percentage Points table instead, in the opposite direction. This mechanic, and the accuracy convention around it (4 d.p. table values, 3 s.f. final answers), is already fully covered by the formula bank and the live lesson's own worked material and is not repeated here.

#2. A real distinct sub-case: the four-way structural taxonomy for a Normal conditional-probability question

#2.0 The setup

Every conditional-probability-with-Normal question in the reviewed archive reduces to the same underlying shape: a target event $A$ and a conditioning event $B$, each a threshold on the same Normal variable — "$X>\alpha$" or "$X<\alpha$" for some cutoff $\alpha$. The general identity is always $P(A\mid B)=P(A\cap B)/P(B)$ (spec 3.2; the base multiplication-law form is itself printed in the formula booklet — verified facts bank §2a — only the independence-specific simplifications are the memorise items). What decides the shape of $A\cap B$ is fixed entirely by which direction each event points and how the two cutoffs compare — a purely logical question, answerable before any table is opened. There are exactly four possible relationships between two such threshold events, not two:

#2.1 Same direction → nesting is automatic

If $A$ and $B$ point the same way (both "greater than" or both "less than"), one event is always a subset of the other — there is no third possibility.

Case 1 — target nested inside conditioning ($A\subset B$). This is the shape of the live lesson's own central worked example: plant-height $H\sim N(50,8^2)$, target $\{H>60\}$, conditioning $\{H>50\}$; since $60>50$, every value satisfying the target also satisfies the conditioning event, so $A\cap B=A$ and $P(A\mid B)=P(A)/P(B)$. That scenario is explicitly VERIDIAN-original, not a reproduced past-paper question — the facts bank's own Section 6 citations are fragments of real student errors, never full question texts with a stated mean and standard deviation (facts bank §6, and the lesson's own header comment states this directly). The reviewed archive's real citations for "conditional probability not recognised under a Normal model" (Oct 2021 Q6c, Jan 2023 Q5e, and the general Jun 2022 diagnosis) confirm the failure repeatedly, but their own fragments do not give enough of the underlying question text to confirm whether any specific one of them has this exact nested shape rather than the confirmed-overlap shape of Case 2 below — an honest limit of what the archive's fragments establish, not a claim that this shape has never been set.

Case 3 — conditioning nested inside target ($B\subset A$), trivially $P(A\mid B)=1$. Not drilled anywhere in the live lesson: every nested-case item there (the marked-solution part (d), nd-pq-3, nd-mcq-4) uses the target as the smaller, more restrictive event, never the reverse. Fresh micro-example: a coffee machine dispenses volume $V\sim N(200,5^2)$ ml. Target $A=\{V>190\}$; conditioning $B=\{V>205\}$. Since $205>190$, $V>205$ forces $V>190$ automatically, so $B\subset A$, $A\cap B=B$, and $P(A\mid B)=P(B)/P(B)=1$ — no standardisation and no table lookup are needed at all. This direction is a logical consequence of exactly the same "which event is more restrictive" check the live lesson already teaches for Case 1, but no citation in the reviewed archive (6 examiner reports, 14 mark schemes) was found testing it directly — flagged here as a real, necessary completion of the model, not as a confirmed exam pattern.

#2.2 Opposite directions → overlap or disjoint, decided by comparing the cutoffs

If $A$ and $B$ point opposite ways ("greater than" vs "less than"), whether they overlap depends only on which cutoff is larger.

Case 2 — overlap, requiring a genuine interval subtraction. This is the confirmed, real, and richest-evidenced case in the whole facts bank. The audited anchor is Jan 2021 Q3, package weights $W\sim N(18,5.4^2)$: part (b) finds $j$ such that $P(W>j)=0.15$ (giving $j$ above the mean, since only 15% of the distribution exceeds it); part (c) asks, given $W<j$, for $P(W>18\mid W<j)$ — target $\{W>18\}$, conditioning $\{W<j\}$, and since $18<j$ the two events genuinely overlap on the interval $(18,j)$, with neither containing the other. The real mark scheme's credited route (facts bank §6, audit lines 983–1016) is $P(A\cap B)=P(W<j)-P(W\le 18)=0.85-0.5=0.35$ — the $0.5$ coming free from mean-symmetry, no table lookup required for it — giving $P(A\mid B)=0.35/0.85=7/17\approx0.412$, exactly the figure ($7/17$ or awrt $0.412$) the real mark scheme accepts. Before the 2026-09-12 audit, every conditional-probability item in the live lesson used a strictly nested pair; the audit added a chain-drill (cereal-box fill-weight, $Y\sim N(60,8^2)$, target $\{Y>60\}$, conditioning $\{Y<j\}$ with $j$ the upper-10% cutoff) built specifically to drill this non-nested shape with fresh VERIDIAN numbers, mirroring the same structure. Two further real citations confirm this same general skill is tested repeatedly without giving enough structural detail to place them on this taxonomy precisely: Oct 2021 Q6c (candidates misapplying a given percentile ratio outside the specific band the question restricted them to) and Jun 2022 Q6c, which Pearson's own report names as the final and most discriminating part of that series' paper, describing candidates mishandling a double-inequality probability either as a plain subtraction the wrong way round or as a one-sided tail probability alone.

Case 4 — disjoint, trivially $P(A\mid B)=0$. Not drilled anywhere in the live lesson. Fresh micro-example: a delivery time $D\sim N(45,6^2)$ minutes. Target $A=\{D>52\}$; conditioning $B=\{D<38\}$. Since $38<52$, the two threshold zones never meet — being told $D<38$ makes $D>52$ impossible — so $A\cap B=\varnothing$ and $P(A\mid B)=0$ immediately, again with no standardisation needed. As with Case 3, this is logically forced by the same formula and definitions, not confirmed as a tested exam shape in the reviewed archive.

#2.3 Why the trivial cases still belong in a theory reference

Two of the four cases (2 and, less certainly, 1) are richly confirmed real exam content — the single most consistently discriminating sub-skill on the whole paper, per the facts bank's own topic-richness ranking (§7). The other two complete the logical picture the identical formula generates and are worth knowing precisely because they are the fast check that should happen before reaching for a table: a student who checks containment first, and finds the conditioning event already forces the target (or already excludes it), never needs to standardise anything at all. They are presented here as a real, necessary structural completion of the model, not as evidenced exam frequency — no likelihood is claimed for either.

#3. A real distinct sub-case: one unknown vs two unknowns in an inverse-Normal problem, and a precise evidence-boundary flag

#3.1 One unknown — confirmed real, and confirmed the hardest single skill in the archive

Spec 6.1 covers the inverse direction of a standardisation: a given probability, an unknown boundary or parameter, found via the Percentage Points table. The one confirmed real citation for this specific skill in the reviewed archive is Jun 2024 Q5(c), a multi-step inverse-Normal problem whose examiner report is unusually blunt: only the more able candidates scored full marks, and abandoning the problem partway through was common (paraphrased; facts bank §6). The facts bank's own framing of this citation is unambiguous: by its own analysis, this is the single hardest Normal-distribution skill type anywhere in the reviewed archive, by a clear margin. The live lesson's one-unknown chain (puzzle-completion time, $T\sim N(\mu,15^2)$, solving for $\mu$ from a single given tail probability) drills exactly this mechanism and is not repeated here.

#3.2 Two unknowns — real, spec-sanctioned, but not yet confirmed by a specific past-paper anchor

Spec 6.1's guidance names a second question type in advance: "questions may involve solving simultaneous equations." This licenses an entire real question type — both $\mu$ and $\sigma$ unknown, two given probabilities the only route to pinning down both — which the live lesson correctly drills with a worked-chain (light-bulb lifetime, two equations eliminated by subtraction). But across the 6 examiner reports and 14 mark schemes the facts bank reviewed, no citation was found describing a real, sat past-paper question with two unknown Normal parameters solved simultaneously; every citation for this question type traces to the spec's own advance guidance text, not to a confirmed instance from an actual series (facts bank §1, §6). This is not a claim the skill has never been examined — 8 of the unit's 14 mark-scheme series still have no examiner report reviewed at all (facts bank §3) — only an honest statement of what the currently-reviewed archive does and does not confirm.

#3.3 The evidence-boundary flag

The live lesson's two-unknown worked-chain attaches the Jun 2024 Q5(c) examiner-report evidence from §3.1 as its own embedded real-world citation. Read against the facts bank's own description of that same citation — "finding an unknown parameter" (singular) from a given probability, with the sticking point a single sub-calculation inside a multi-step problem — this citation reads naturally as evidence for the one-unknown case of §3.1, not the two-unknown simultaneous system it is attached to in the lesson's worked-chain. Both worked chains remain individually accurate and well-grounded: the light-bulb chain in the spec's own advance guidance (§3.2), the puzzle-time chain in a confirmed real anchor (§3.1). The precise, evidence-faithful claim for each is different, though: "the specification itself names this question type in advance" is the accurate claim for the two-unknown case specifically; "an examiner report singles this out as the hardest skill on the paper" is the accurate claim only for the one-unknown case the citation actually describes. Keeping the two claims separate matters if either is ever repeated to a student as a fact about what Pearson has confirmed.

#4. Synthesis — every real Section 6 exam angle in the reviewed archive, in one place

Honest frequency statement. The Normal distribution is confirmed as tested, with at least one full question devoted to it, in every one of the 5 originally-reviewed examiner reports — Jan 2021, Oct 2021, Jun 2022, Jan 2023, Jun 2024 (facts bank §7). The sixth examiner report the facts bank has since read, June 2019 (a 2026-09-25 spot-check), was read for two unrelated topics (outliers/box plots; discrete random variables) and was not re-examined here for Normal-distribution content, so its status for this specific topic is genuinely unconfirmed, not assumed absent. Eight of the unit's fourteen mark-scheme series still have no examiner report reviewed at all (facts bank §3) — so "5 of 5 fully-reviewed reports" is the honest current denominator for this topic, not "5 of 14," and this synthesis is a lower bound, not a claim of exhaustive coverage.

#5. Open flags

• §2, Cases 3 and 4 (conditioning nested inside the target, giving $P=1$; disjoint threshold events, giving $P=0$) are real, logically necessary completions of the same conditional-probability model this topic already teaches, but no citation in the reviewed archive was found testing either directly.
• §3.3 — the live lesson's two-unknown worked-chain cites Jun 2024 Q5(c) as its real-world evidence, but that citation's own wording (an "unknown parameter," singular) more precisely describes the one-unknown case. Both worked chains are still individually well-grounded (§3.1–3.2); only the specific evidentiary claim attached to each should be kept separate.
• Eight of fourteen mark-scheme series have no examiner report reviewed yet. Any statement above about how many series confirm a given angle is bounded by that limit, matching the facts bank's own standing caveat (§3).

#Sources

research/veridian/WST01-verified-facts.md — §1 (spec transcription and the verified spec-numbering note), §2a (formula-booklet contents, including the two Normal tables' exact domains and the memorise-vs-given split for E(aX+b)/Var(aX+b) and the conditional-probability identity), §3 (source archive — 6 of 14 examiner reports reviewed), §5/§6 (all Section-6 examiner-report citations used above), §7 (topic-richness ranking), and the 2026-09-12 per-lesson audit entry for normal-distribution-standardisation-and-conditional-probability.ts (the Jan 2021 Q3 anchor's full numbers and mark-scheme figures). content/veridian/curriculum/wst01/normal-distribution-standardisation-and-conditional-probability.ts (live lesson, read in full, including its header comment, all worked-chain/chain-drill/MCQ stages, and its embedded-evidence citations). content/veridian/technique-hub/wst01/codex-formula-and-method-bank.md, forge-exam-technique-guide.md, signal-frequency-and-boundaries.md (grepped for this topic). content/veridian/question-bank/wst01.ts (checked — no item on this topic exists there yet).