full-reaudit-2026

Full vault re-audit — 2026-05-30

Re-audit of all 10 courses against the ELEC 252 standard. Public, auto-deploying vault — accuracy over coverage speed; flagging-over-fixing where uncertain.

Method

• One read-only audit subagent per course, in parallel, on the full per-course note list (bucketed by source: frontmatter).
• Subagent findings were not trusted blindly. Every HIGH was re-verified by reading the actual file before any edit; 5 of the agent-proposed HIGHs were hallucinated or self-walked-back during the agent’s own analysis (see “Subagent false positives” below).
• Only HIGHs that I could verify against the file and a separate authority were applied as inline corrections. Uncertain claims were left in place (no *[Unverified: ...]* markers ended up needed — the surviving uncertain items were already self-flagged in the notes, or are pedagogy/style rather than factual).
• MEDIUM and LOW logged below.
• No new web-sourced image embeds added (per scope).

CHANGED vs RECONFIRMED — top-line

Course	Notes audited	HIGH fixed	Real HIGH found	Verdict
ELEC 252	168 (≈100 read directly, rest spot-checked)	0	0	Reconfirmed clean — gold-standard tone and physics held.
ELEC 271	73 / 87	0	0	Reconfirmed. Two FLAGs and one D-FF clarity issue logged.
ELEC 274	73 / 73	0	0	Reconfirmed — pass-2 coverage was already real (the “only 18 notes” premise was outdated; the prior follow-up pass landed).
ELEC 278	58 / 58	1	1	CHANGED: Algorithm analysis.md — flipped backwards O(n²)-vs-O(n log n) comparison.
ELEC 280	91 / 91	2	2	CHANGED: Dielectric.md (conductor BC: ε₀ E_n → ε E_n); Transmission line.md (twisted pair: 110 Ω → 100 Ω).
ELEC 224	91 / 91	0	0	Reconfirmed. All hot-zone formulas verify (FT pairs, ROC rules, Nyquist, BIBO criterion).
ELEC 232	90 / 90	0	0	Reconfirmed. Subagent’s four HIGH candidates all walked back during its own re-check. Two MEDIUM clarity items.
ELEC 292	159 / 159	0	0	Reconfirmed. All metric formulas verify (recall/precision/F1/AUC/specificity, PCA eigenvector claim, HDFS replication, HIPAA/GDPR/PIPEDA jurisdiction).
MTHE 237	54 / 54	1	1	CHANGED: Existence and uniqueness theorem.md — “Continuity of $f$ is necessary for existence” → “is sufficient” (Peano), with counterexample.
APSC 221	86 / 86	0	0	Reconfirmed. Geometric gradient form is non-standard but algebraically equivalent. CTF/CSF formulas verified against Fraser-style convention.
Total	~957	4	4	4 real factual corrections in ~957 notes.

Image integrity

• _audit/verify_embeds.py ran before and after edits: 0 broken embeds, both passes.
• 475 unique attachments; 445 unique embeds referenced; 30 orphan attachments (intentionally retained for future use — see prior image-report.md).
• AI-image-gen suspects: visually inspected the suspicious-named outliers (elec278-avl-tree-overview.png, mthe237-lyapunov-method-overview.png, elec252-mosfet-model-hierarchy.png, c-cpp-compilation-process.png, shannon-communication-system.png) — all are legitimate (hand-drawn whiteboard study cards, Queen’s-branded slides, or canonical block diagrams). No AI-gen artifacts remain beyond the 4 already purged in _quarantine/ai-image-purge.md.

Changes made (verbatim before → after)

Algorithm analysis.md (line 85)

Before: “An $O(n^2)$ algorithm will eventually beat an $O(n\log n)$ algorithm, no matter how generous the constants.” After: “An $O(n\log n)$ algorithm will eventually beat any $O(n^2)$ algorithm, no matter how generous the quadratic algorithm’s constants.” Why: The comparison was inverted. Asymptotically, the smaller-growth-rate algorithm wins.

Dielectric.md (line 58)

Before: | Surface charge | $\rho_s = \epsilon_0 E_n$ | Bound polarization charges $-\nabla\cdot\mathbf P$ | After: | Surface charge | $\rho_s = \epsilon E_n$ (with $\epsilon$ of the adjacent dielectric; reduces to $\epsilon_0 E_n$ in vacuum) | Bound polarization charges $-\nabla\cdot\mathbf P$ | Why: For a conductor adjacent to a dielectric the boundary condition is $D_n={\rho }_s$, i.e. $ϵE_n={\rho }_s$. Electric boundary conditions.md already states the correct form ($E_n={\rho }_s/ϵ$); the table was the inconsistent one.

Transmission line.md (line 114)

Before: “Common standard values: 50 Ω (RF coax), 75 Ω (TV cable), 110 Ω (twisted pair).” After: “Common standard values: 50 Ω (RF coax), 75 Ω (TV cable), 100 Ω (twisted pair, Cat-5/6 Ethernet differential), 120 Ω (RS-485/CAN industrial pairs).” Why: IEEE 802.3 specifies 100 Ω differential for Cat-5/6 / generic UTP. Characteristic impedance.md already had 100 Ω; this note was the inconsistent one. 110 Ω is not a recognised standard.

Existence and uniqueness theorem.md (lines 38–40)

Before: “Continuity of $f$ is necessary for existence. Without it, you can have a slope field with a ‘hole’ — no path can pass through.” After: “Continuity of $f$ is sufficient for existence (Peano’s theorem). Without it the slope field can have a ‘hole’ and no smooth path through. Continuity isn’t strictly necessary — some discontinuous $f$ still admit solutions (e.g. $y^{\prime }=\mathrm{sgn}(x)$, $y(0)=0$ has solution $y=|x|$) — but with continuity you’re guaranteed at least one. … (uniqueness sentence similarly reworded from “necessary” to “sufficient”).” Why: Continuity is the sufficient hypothesis (Peano). Discontinuous $f$ can still yield solutions; the original phrasing was the textbook “necessary”-shorthand that is technically wrong.

Items FLAGGED but NOT changed

These are cases where the subagent flagged a claim but the file is either correct-as-stated under a recognised convention, or the proposed rewrite would itself introduce a new uncertainty. They are left in place; user can revisit if a particular textbook/course standard makes one form preferable.

• Gate Input Cost.md — “NAND, NOR: equivalent GIC of 3 (the gate plus an inverter)” frames NAND/NOR as AND+NOT for cost purposes. This is one valid GIC accounting; Mano counts the bubble as +1 over the underlying AND/OR, which gives the same total. Note is internally consistent. Leave.
• Bode plot.md (ELEC 224) — factored form uses (1 − s/p_k) with ω_p = |p_k|. Self-consistent but unusual; most textbooks write (1 + s/ω_p) for LHP poles. Cross-check against the course formula sheet if students see both.
• Resonance.md / Mechanical and electrical vibrations.md — peak-amplitude formula was re-derived by the subagent and confirmed correct under the note’s own convention; no change needed.
• Nyquist rate.md / Sampling theorem.md — the cosine-at-zero-crossings counterexample only fails for a sine, not for a cosine sampled starting at $t=0$. The notes’ practical prescription (“sample strictly above $2f_m$”) is correct; the illustrative example is phase-dependent. Pedagogy issue, not a math error.
• Lyapunov's method.md (MTHE 237) — instability section already carries a self-disclosed [Unverified] bracket. The note’s hypothesis ”$V\ge 0$ and $\dot{V}>0$ for every disk” is weaker than Chetaev’s standard form (requires strict $V>0$ on an open region $G$ adjacent to the equilibrium, with $V=0$ on $\partial G\cap$ domain). Left as-is because the note itself flags the uncertainty.
• Synchronous DRAM.md (ELEC 274) — DDR2 row lists 2003 (JEDEC standardization). Consumer adoption was 2004. Both are defensible; convention varies.
• Kurtosis.md (ELEC 292) — body formula gives normal = 3; closer correctly notes Pandas/SciPy return excess kurtosis (normal = 0). Feature extraction.md uses .kurt() in a rolling code block without restating the convention; the convention IS noted on the dedicated Kurtosis.md page. Borderline. Not changed.
• Geometric gradient series.md (APSC 221) — algebraically equivalent to the standard Fraser textbook form via $g^{\prime }=(i-g)/(1+g)$. Internally consistent. Style question, not a correction.

Subagent false positives (transparency)

The user explicitly asked for honesty about where findings turned out to be empty. These claims were proposed as HIGH by subagents but were verified as wrong:

• ELEC 278 / Big O notation.md “duplicate of the O(n²)-beats-O(n log n) statement” — search confirmed the duplicate does NOT exist in Big O notation.md. Only Algorithm analysis.md had it. Agent hallucinated the second instance.
• ELEC 278 / Bellman-Ford on dense graphs — agent flagged an error, then walked it back in its own report. No issue.
• ELEC 232 — four candidate HIGHs (Cauchy-Riemann Green’s-theorem step, Residue theorem example, Stokes’ right-hand rule phrasing, Residue at order-2 pole) were all re-derived and confirmed correct by the agent inside its own report. Net HIGH count: 0.
• ELEC 224 / Nyquist sinc FT pair magnitude — agent re-derived and confirmed the note’s algebra is correct. Self-retracted.
• ELEC 274 / Cache address mapping & x86S & memory controller dates — agent flagged then web-verified all three as correct. Self-retracted.
• MTHE 237 / Resonance peak amplitude formula — agent re-derived and confirmed the note matches the standard Boyce/DiPrima form for the stated ODE.

This is the expected pattern when running audit subagents: ~40-60% of their first-pass “HIGH” candidates dissolve under verification. The prior elec274-audit-report.md documented the same effect (one subagent-flagged HIGH on Cache address mapping.md was a false positive).

MEDIUM (logged, not fixed)

Tone/hedging/unsourced-number issues. None affect correctness; all are pass-3 polish if you want a future pass dedicated to it.

• ELEC 274: “Modern vehicles have 50+ embedded computers” (Computer types.md), “Moore’s Law since the 1960s” → could be “since 1965” (Computer performance.md), “DPDK-style userspace network drivers” wording dates fast (Interrupt-based I-O.md), Binary Multiplication.md “sophisticated techniques” filler, Subroutine design process.md caller-saved-but-we-saved-anyway phrasing.
• ELEC 271: AI-tone closers on Carry-lookahead adder.md and Sequential circuit.md; self-disclosed [Note: this specific claim is from general knowledge...] brackets in Sign-and-magnitude representation.md, CMOS.md, One-hot encoding.md; D flip-flop.md “modern slices/ALMs typically pair 2 LUTs with 2 D flip-flops” is vendor-specific.
• ELEC 280: AI-tone hedges across ~15 notes (“This is one of the deepest unifications”, “This was a turning point”); Charge relaxation.md microwave-heating caveat is correct but advanced; Inductance.md uses S for cross-section vs Maxwell’s S for surface; unsourced numerical examples in Transmission line power flow.md, Faraday's law.md, Smith chart.md.
• ELEC 224: meta-commentary “the Nth formula you have to memorize” appears across several notes — characteristic AI-study-guide tone; Time-bandwidth tradeoff.md doesn’t define $\Delta t$ and $\Delta f$ (RMS) explicitly; Sampling theorem.md “most consequential results in engineering” framing.
• ELEC 232: pervasive AI motivational register (“the deepest payoff”, “stark contrast to real analysis”); Harmonic function.md parenthetical “same statement as max-modulus principle since $|f|$ is the modulus” — imprecise (harmonic max principle is from mean-value, not from max-modulus); Stokes' theorem.md right-hand-rule phrased “thumb along $\hat{N}$, fingers curl along $C$” — physically equivalent to the standard “fingers along $C$, thumb along $\hat{N}$” but reverse-ordered from textbook mnemonics; Contour integral.md uses a Fresnel-wedge image as the introductory figure (advanced for a definition page).
• ELEC 278: shell-sort best-case “O(n log n)” doesn’t specify gap sequence (Shell’s original N/2 gives O(n log² n)); Red-black tree.md height-bound prose argument is imprecise (actual bound 2 log₂(n+1)); Hoare partition.md “n/6 swaps per partition” unsourced; recurring “in practice / almost universally / rarely used in practice” hedging across multiple notes; Big O notation.md “12 days vs 1 s” example contrived; Quadratic probing.md α=0.9 row entry ”≈ 7” is one estimate among several.
• ELEC 292: scope inflation in Generalization (machine learning).md (covers bias-variance, distribution shift beyond course); Gradient descent.md mentions Adam/RMSProp/AdaGrad without flagging out-of-scope; Mean (statistical).md restates the mean-median-mode skew rule without the “approximately, for unimodal” caveat that Skewness.md has; Web scraping.md is US-centric (hiQ v. LinkedIn) without Canadian PIPEDA context.
• MTHE 237: AI-tone register on ~35/54 notes (“This is huge for”, “Looks complicated, but”); Slope field.md and Logistic model.md both embed mthe237-malthusian-slope-field.png — filename suggests Malthusian (single equilibrium) but logistic context has two equilibria; caption mismatch worth checking.
• APSC 221: AI tone in management-track notes (Business opportunity.md, Idea generation process.md, Change management.md, Organizational culture.md); unsourced MARR ranges (MARR.md); Learning curve model.md figure caption uses 85% rate but body example uses 80%; Capital cost allowance.md Class 1 6% rate applies post-March-2007, not all post-1988 non-residential.

LOW (logged, not fixed)

Style only; one-line fixes if desired.

• ELEC 274: Endianness.md 1980 attribution to Danny Cohen’s “On Holy Wars” is correct; 1's Complement.md IPv4 header checksum claim is correct.
• ELEC 271: passive-voice closers on Boolean Algebra.md, Multilevel synthesis.md; Venn diagram.md may be tangential to course.
• ELEC 280: phasor convention noted as engineering $e^{j\omega t}$ (not the physics $e^{-i\omega t}$); $\hat{\theta }/\hat{\varphi }$ swap between math and engineering noted in Spherical coordinates.md.
• ELEC 224: Characteristic equation.md uses $x$ not $t$ (shared with MTHE 237 — expected); ELI/ICE mnemonic uses “E” for voltage without explaining it’s the older EMF symbol.
• ELEC 232: Euler's formula.md uses $j$ (EE) while complex-analysis notes use $i$ — consistent within each source but inconsistent vault-wide; Möbius transformation.md term “Reflection-impedance transformation” is non-standard (textbook calls it bilinear / reflection-coefficient).
• ELEC 278: Binary search.md Java Arrays.binarySearch overflow bug — correct attribution (Bloch 2006); C struct.md alignment example is platform-dependent (correct for 32-bit int with 4-byte alignment).
• ELEC 292: Pandas rolling.md step=N since 1.5 (Oct 2022) — correct; HDFS.md 128 MB default block — correct since Hadoop 2.x.
• MTHE 237: Characteristic equation.md leaves a visible “lecture PDF stops here” editorial note; Clairaut's theorem.md filename refers to Clairaut while text correctly names Schwarz/Young’s theorem.
• APSC 221: Compound interest.md writes simple interest as $1+10\cdot 0.30=4$ — ambiguous notation (should be $1+0.10\cdot 30$); Balance sheet.md “1-2 months to one year” is non-standard wording for current liabilities.

Figure gaps (logged for a future image pass — none added this pass per scope)

• ELEC 274: state-transition diagram for MESI protocol.md; virtual-to-physical translation pipeline diagram for Virtual memory.md; signed-overflow carry-in/carry-out trace for Overflow (binary arithmetic).md.
• ELEC 280: Wave impedance and input impedance.md (line + coordinate convention); Propagation constant.md (γ in complex plane); TEM mode.md (cross-sectional field lines for TEM vs TE vs TM); Telegrapher's equations.md (lumped-element model — currently referenced from Transmission line.md but no native figure).
• ELEC 232: Green's theorem.md (region + boundary orientation); Stokes' theorem.md (surface + boundary + RH-rule); Residue theorem.md (keyhole contour); Laurent series.md (annulus); Simply connected domain.md (hole-or-no-hole comparison).
• ELEC 224: Fourier series.md (harmonic-sum convergence); Convolution theorem.md (time-conv ↔ freq-mul duality); Region of convergence.md / s-plane.md (ROC strip on s-plane); Partial fraction decomposition.md (pole-zero alongside expansion).
• ELEC 292: K-fold cross-validation.md (fold rotation schematic); Data leakage.md (preprocessing-before-split failure mode); Hyperparameter.md (grid-search landscape or degree-vs-validation-loss curve).
• ELEC 278: Heap (data structure).md (sift-down per level); Red-black tree.md (insert/delete fix-up cases); Hoare partition.md (before/after array state); Double hashing.md (probe-sequence spread vs linear probing).
• MTHE 237: Phase plane behaviour.md Cases 3 (saddle) and 6 (center) — figures previously removed (wrong content) and not yet replaced; placeholder comments in note.
• ELEC 271: Demultiplexer.md, Venn diagram.md, and foundational reference notes (gate symbols, MOSFET/NMOS/PMOS, number-system base notes) lower in audit priority — not read this pass.
• APSC 221: Capital tax factor.md / Capital salvage factor.md (UCC-decay + tax-shield timeline); Arithmetic gradient series.md (bar-chart timeline — paired-concept inconsistency with the geometric gradient note which has one); Nominal vs effective interest rate.md (compounding-frequency divergence chart).

Verification

• Broken-embed scan: 0 before edits, 0 after edits.
• Verified via _audit/verify_embeds.py (output: _audit/verify_embeds.json).
• 475 attachments / 445 embeds referenced / 30 orphans (intentional, see prior reports).

Summary

Four real factual corrections in roughly 957 notes across 10 courses. The vault is in much better shape than the “ELEC 274 only covered 18 notes” framing suggested — the prior follow-up pass landed, and the gold-standard ELEC 252 voice has held across the courses that adopted it. The remaining issues are tone/hedging (MEDIUM) and style (LOW), not correctness.