How this was checked

provenance

Answer provenance: worked answers are derived study aids written by working through each question; they are not official answer keys. verified_by_code: true on a solution means a numerical check was run; false means no such check is claimed. Where a scan and its extracted text disagree, the page shows an editorial correction next to the question.

Question bank verification

Output of the bank verifier over the four weekly banks:

Inputs: 488 inventory records; 241 routed; 247 archived.
Week 1: 73 routed sections; IDs unique; each has worked answer and verification marker.
Week 1: 7 topic checklist rows; no blank cells; all topics.json topics present.
Week 2: 65 routed sections; IDs unique; each has worked answer and verification marker.
Week 2: 8 topic checklist rows; no blank cells; all topics.json topics present.
Week 3: 36 routed sections; IDs unique; each has worked answer and verification marker.
Week 3: 8 topic checklist rows; no blank cells; all topics.json topics present.
Week 4: 67 routed sections; IDs unique; each has worked answer and verification marker.
Week 4: 8 topic checklist rows; no blank cells; all topics.json topics present.
Archive: 247 IDs appear in the index and none appear in the four banks.
Images: 147 contextual image links resolve from their bank files.
Corrections: 17 question records have visible editorial notes; 12 shared-context blocks rendered.
Reconciliation: W1 73 + W2 65 + W3 36 + W4 67 + archived 247 = 488.
Topics without direct question: W3 Ways to parameterise a language model (w3-model-parameterisations)
PASS: exam bank structure, coverage, topic checklists, archive routing, and image links verified.

Quality gates

Final summary of the repository gates (links, topic checklists, deterministic plots, saved code blocks, front matter, numeric provenance, JSON integrity, file sizes):

FINAL SUMMARY
PASS G1 routing/bank/index reconciliation
PASS G2 internal links
PASS G3 topic checklists
PASS G4 deterministic plots
PASS G5 saved executable codeblocks
PASS G6 frontmatter and table pipes
PASS G7 numeric answer provenance
PASS G8 JSON integrity
PASS G9 repository hygiene and file sizes

Routing audit

Each of the 488 question records in the 26 papers was routed to a Week 1 to 4 topic or archived with a reason. The routing merge reported:

CHECK 1 inventory coverage: PASS (inventory=488 routed=488 missing=0 extra=0 repeated=0)
CHECK 2 allowed weeks: PASS (invalid=0)
CHECK 3 week-valid topics: PASS (errors=0)
CHECK 4 archive metadata: PASS (errors=0)
CHECK 5 routing overrides: PASS (applied=2 errors=0)

Duplicate detection on the website

The Remove duplicates filter on the question banks hides repeats client-side only. A repeat is an exact duplicate: the same normalised stem text, options, shared context and scan-image bytes. 12 such groups (24 occurrences) were found among the 241 routed occurrences (scripts/dup_overrides.json, rebuilt by scripts/sync_content.py). Nothing is deleted from the banks; the unfiltered view always lists every paper occurrence.

What is not published

The course ebook (a third-party document) and its page images are not part of this site; notes cite ebook page numbers as text only. Course lecture material is not republished either.