Deep Dive: Mathematical Modelling in Probability and Statistics

#CODEX Deep-Dive — Mathematical Modelling in Probability and Statistics (WST01, spec 1.1)

Provenance (2026-09-29): Read/checked for this document — research/veridian/WST01-verified-facts.md in full, grepped specifically for "model", "assum", "extrapolat", "reliab", "suitab"/"appropriate", "skew" (§1 spec transcription, §2b AO weighting, §5 Section 1/3/4 trap catalogue, §7 topic-richness ranking, §8 flagged issues, and the per-lesson audit entry for mathematical-modelling-in-probability-and-statistics.ts); the live lesson file content/veridian/curriculum/wst01/mathematical-modelling-in-probability-and-statistics.ts in full; content/veridian/technique-hub/wst01/codex-formula-and-method-bank.md §2.1, forge-exam-technique-guide.md item 4, and signal-frequency-and-boundaries.md row 1.1; content/veridian/curriculum/wst01/elementary-probability-and-conditional-notation.ts and regression-gradient-interpretation-extrapolation.ts (grepped for modelling/assumption framing); content/veridian/question-bank/wst01.ts (checked — no item on this topic exists there). Two other papers' verified-facts files were also read for genuine cross-qualification corroboration, not as WST01 evidence: research/veridian/WME01-verified-facts.md (§4, item 1, Mechanics' own spec-1.1) and research/veridian/WMA14-verified-facts.md (AO3 row). No new PDF fetches were made. All arithmetic in the micro-examples below was independently checked with exact-fraction arithmetic (Python's fractions.Fraction) before being written in.

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

Spec 1.1 is the thinnest topic in this entire facts bank: zero standalone questions across all 14 reviewed WST01 series. The live lesson already teaches the core vocabulary (model/assumption/parameter/refinement), the four-step licensing chain, and works it in full through the one real anchor that exists — Jan 2025 Q1(a)–(b), the fair 4-sided dice. The formula bank's §2.1 already gives the same three-step method template in miniature. The signal doc's frequency table already states the "0 of 6 as a standalone question, funded by AO3" fact. None of that is reproduced here.

Given how little dedicated evidence spec 1.1 has, this document's job is different from a normal CODEX deep-dive: rather than deriving one topic's mechanics further, it does the thing no existing file does — it goes hunting, systematically, through the rest of the WST01 archive (and, honestly labelled, two sibling papers) for every real place the same assumption→family→calculation→refinement chain the lesson teaches from one anchor actually shows up in a different guise elsewhere in this unit, plus one place it visibly does not show up. That is the genuine exam-angle synthesis this topic's real evidence supports — not a deeper derivation of the one case already taught in full.

#1. The chain, restated precisely, and one qualification-wide check

The live lesson's own chain — assumption → distributional family → parameters → calculation, with refinement as a fifth step when evidence contradicts the assumption — is exactly what AO3 names: "standard mathematical models... assumptions and refinement" (spec pp. 68–69, abridged per the facts bank's own table, §2b). Two things worth adding precisely:

This AO3 language is not S1-specific. WMA14 (Pure Mathematics 4)'s own AO3 row uses nearly identical wording — "standard mathematical models... assumptions and refinement" — at a flat 10% weighting on that paper (WMA14-verified-facts.md, AO3 row). S1's own 15–20/75 (≈20–27% of the paper, per the facts bank's own comparison, §2b) is markedly higher than that 10% figure and than the Pure units P1–P4's 5–10/5–15 range — confirming, from a second unit's spec table rather than just S1's own, that "assumptions and refinement" is boilerplate AO3 phrasing applied qualification-wide, and S1's unusually high share of it is a real, checkable structural fact about this specific unit, not an artefact of how one spec table happens to be worded.

Why S1 has no standalone question where Mechanics sometimes does. WME01 (Mechanics 1)'s own spec item 1.1 carries a printed vocabulary list — particle, lamina, rigid body, rod, inextensible string, smooth and rough surface... — and its own examiner-report record shows exactly the standalone "critique the model" question type S1 never produces: Oct 2019 Q6's final part, where credit depended on recognising "only modifications to that model received credit" (WME01-verified-facts.md, §4 item 1) rather than a generic critique. WST01's own spec text for 1.1 is a single unadorned sentence with no such list (already noted in the live lesson's header). This is a genuine structural difference, not a difference in how thoroughly each paper was searched: Mechanics' spec gives examiners a closed, nameable set of objects to build a standalone question around; Statistics' spec gives them nothing to anchor one to beyond the general idea itself — which is consistent with, and gives an actual mechanism for, why the WST01 archive search comes back empty every time.

#2. Real distinct sub-cases, each with a short micro-example

#2.1 Case A — the assumption is licensed by the question (already the lesson's own territory; restated in miniature)

The question states the assumption ("fair die") and it's accepted; naming the family it licenses is what earns credit (Jan 2025 Q1(a), B1 for "Discrete uniform"). Fresh micro-example, different numbers from the lesson's own die/spinner scenarios: a fair 12-sided die, numbered 1–12, is rolled once; its score is modelled by $D$, discrete uniform on $\{1,\dots,12\}$. $P(D \text{ is a multiple of } 3) = P(D\in\{3,6,9,12\}) = \tfrac{4}{12} = \tfrac13$.

#2.2 Case B — the reverse failure: an assumption illegitimately self-imposed, not licensed by anything

This is the mirror image of Case A, and it is real, cited, and distinct from the "independence vs mutually exclusive judged by eye" confusion the sibling conditional-probability lesson already owns in depth. Pearson's own stated general diagnosis one series: "candidates often assume independence when an appropriate conditional probability should be used instead" (Jun 2022, general comment) — and, on that same series' Q4(b), a Venn-diagram question, "common errors were to assume independence" rather than compute the genuine conditional probability the diagram required. Framed through spec 1.1's own AO3 language rather than as a probability-notation slip: this is a candidate manufacturing their own unlicensed model — inventing a "the events are independent" assumption the question never granted — rather than reading off the one the question actually licensed. It is the exact opposite failure mode from the lesson's own trap-taxonomy item "assumption treated as a provable fact": there, a real licensed assumption is wrongly doubted; here, no assumption was licensed at all, and one is wrongly asserted anyway.

Fresh micro-example (not the lesson's own two-dice or the sibling deep-dive's card/counter numbers): a bag holds 4 white and 3 black counters; two are drawn without replacement. The unlicensed, self-imposed-independence route gives $P(\text{both white}) = \left(\tfrac47\right)^2 = \tfrac{16}{49} \approx 0.327$. The correct route, using the conditional probability the without-replacement setup actually requires, gives $P(\text{both white}) = \tfrac47 \times \tfrac36 = \tfrac47\times\tfrac12 = \tfrac27 \approx 0.286$ — a different number, from the same starting fractions, because the second draw's sample space genuinely changed. Nothing here was independent; treating it as if it were is precisely the licensing violation Jun 2022's own general comment names.

#2.3 Case C — a fitted model's stated validity boundary, breached: extrapolation as a refinement failure

A least-squares regression line is itself a mathematical model in spec-1.1's sense — fitted under an implicit assumption (the linear relationship holds across the range of $x$ used to fit it) that AO3's "refinement" language covers just as much as a fair-die assumption does, via spec 4.2's own guidance on "the dangers of extrapolation." The real citation is exact and structurally identical to the discipline the modelling lesson's own marked-solution teaches (name the specific evidence, not a generic statement): on a real GDP-vs-population question, the credited reliability judgement was "unreliable as 7 000 000 is much greater than the mean population... for the 9 years" (Jun 2022 Q2(e)) — with the marking note stating outright that "extrapolation so unreliable" on its own is B0, and a bare "value out of range" is also B0; only a comment referencing the actual $t$ or $\bar t$ value earns the mark. This is the identical shape as the fair-spinner marked-solution's own common-wrong-path: a true-sounding, unattached statement about a model's limits earns nothing; naming the specific number that breaches the model's fitted range does.

Fresh micro-example: a line is fitted to data with $x$ (hours of sunlight) ranging from 2 to 10. A prediction at $x=9$ sits inside that range — an interpolation, usable. A prediction at $x=20$ sits outside it — an extrapolation; a credited reliability comment names 20 and the fitted range (2 to 10), not just the word "extrapolation" alone.

#2.4 Case D — the Normal distribution as a model, and an honest gap

Spec 6.1's own guidance requires only "knowledge of the shape/symmetry" of the Normal distribution — which is itself a modelling assumption in spec-1.1's exact sense (assume symmetric, bell-shaped spread; the parameters $\mu,\sigma^2$ then pin down which specific Normal curve). No citation anywhere in the reviewed 14-series archive was found testing a "does the data support treating this as Normal" critique question the way Case C's reliability judgement is tested for regression — this facts bank records no such example, and this document does not manufacture one. Flagged here as a real, checkable absence, not asserted as an untested-but-surely-coming exam angle.

#3. Synthesis — every real angle this chain has actually surfaced from, across the archive

Honest frequency statement, no invented likelihoods. Of the 6 examiner-report series fully reviewed for this facts bank (Jun 2019, Jan 2021, Oct 2021, Jun 2022, Jan 2023, Jun 2024), the self-imposed-independence failure (Case B) is confirmed in 1 of 6 (Jun 2022, corroborated by that series' own general comment) and the extrapolation-reliability failure (Case C) is confirmed in 1 of 6 (Jun 2022, a different question on the same series). Spec 1.1 tested as its own standalone question is confirmed in 0 of 14 series' question papers (the full archive, not just the 6 with a read examiner report). These are lower bounds on a partially-read record (the facts bank's own §8 flags 8 of 14 series as not yet searched question-by-question), not a claim about the other series.

#4. Open flags — where the real evidence currently runs out

• Case D (Normal-model suitability) has no confirmed exam citation — stated as an absence in §2.4 above, not filled with a plausible-sounding invented question.
• AO3 marks are not tagged per mark-scheme line by Pearson. The 15–20/75 figure is a whole-paper spec allocation (§2b of the facts bank); there is no published breakdown saying exactly which marks on any one real question are "the AO3 part" versus AO1 technique, so this document does not attempt to quantify, question-by-question, how many of a series' marks are modelling-interpretation marks specifically — doing so would require inventing a split the source material doesn't supply.
• 8 of the 14 mark-scheme series have no confirmed, independently-verified examiner report read yet (per the facts bank's own §8) — Case B and Case C's "1 of 6" counts above should be read as a lower bound on a partially-searched archive, not as the true total frequency of either failure mode.
• The WME01/WMA14 comparisons in §1 and §3 are cross-paper corroboration, not WST01 evidence — flagged explicitly in the table above so neither is mistaken for a WST01 citation on a future re-read of this document.

#Sources

research/veridian/WST01-verified-facts.md — §1 (spec text), §2b (AO weighting and its own S1-vs-Pure-units comparison), §5 Section 1 (the "0 standalone questions" finding), §5 Section 3 (Jun 2022 Q4(b) and general-comment quotes), §5 Section 4 (Jun 2022 Q2(e) quote), §7 (topic-richness ranking), §8 (flagged gaps), and the per-lesson audit entry for mathematical-modelling-in-probability-and-statistics.ts. content/veridian/curriculum/wst01/mathematical-modelling-in-probability-and-statistics.ts (live lesson, read in full). content/veridian/technique-hub/wst01/codex-formula-and-method-bank.md §2.1, forge-exam-technique-guide.md item 4, signal-frequency-and-boundaries.md row 1.1. content/veridian/curriculum/wst01/elementary-probability-and-conditional-notation.ts, regression-gradient-interpretation-extrapolation.ts (grepped, both read in relevant part). content/veridian/question-bank/wst01.ts (checked, no overlapping item found). Cross-paper corroboration only, not WST01 evidence: research/veridian/WME01-verified-facts.md §4 item 1; research/veridian/WMA14-verified-facts.md AO3 row.