Live-Site Audit · revision 3 · 15 Jul 2026

Marvel Commander Rankings

A live audit of the nineteen-configuration precon hub. This revision ships two real fixes with executed proof, and withdraws two findings from revision 2 that did not survive an actual attempt to implement them.

9.1 / 10

reviewed page

Reviewed build: content-hash 2bedcc68d80273c0, http 200 (Vercel copy is byte-identical to the GitHub mirror)

Since revision 2: default-on nineteen-view and a card-level score-provenance hint are shipped and verified by executing the page's real code; the canonical-record gap and the win-line gap are withdrawn, both turned out not to exist

Still open: recommendation-drift consistency across the individual guides, and exhaustive interface propagation beyond the card and detail-modal views

Scryfall verified vs Oracle / set data Code read from the page's JavaScript Editorial disclosed opinion, not measured

Coverage of this audit, stated honestly

The card-fact gate parsed zero card names from the rendered page, because names live in embedded JSON rather than gate-visible markup. That gate produced no meaningful automated coverage. In its place I verified a representative, load-bearing sample by hand: Crystal's colour identity, Captain America's Oracle, Loki's self-oracle, the set release date, the executable composite formula, and every published deck composite. I did not individually compare every rendered card name and rules statement against current Oracle data. The correct claim is representative and load-bearing verification, and a complete claim would require extracting every rendered card name, checking each against live Oracle text, and reporting a coverage count.

Crystal · WURG confirmed

Captain America · Hero engine

Loki · self-oracle matches

The nineteen-row propagation matrix

Every canonical configuration from CMDR_META, mapped across the data surfaces that feed scorecards, packages, win lines, and guide links. This is the reconciliation the first audit asserted without showing.

#	Config	Deck	Role	CI	Score source	Package	Win line	Deck-tech link
1	Doctor Doom	Doom	face	UBR	deck	✓	✓	✓
2	Loki	Doom	second	UBR	deck	✓	✓	✓
3	Mister Fantastic	FF	face	GRUW	deck	✓	✓	✓
4	The Thing	FF	face	GRUW	deck	✓	✓	—
5	Invisible Woman	FF	face	GRUW	deck	✓	✓	—
6	Human Torch	FF	face	GRUW	deck	✓	✓	—
7	Crystal	FF	alt	GRUW	derived	✓	✓	—
8	Power Pack	FF	alt	GRUW	derived	✓	✓	—
9	The Fantastic Four	FF	set-only	GRUW	unscored	—	—	—
10	Captain America	Avengers	face	RWU	deck	✓	✓	✓
11	Director Nick Fury	Avengers	second	RWU	deck	✓	✓	—
12	T'Challa	Wakanda	face	GW	deck	✓	✓	✓
13	Bast	Wakanda	alt	GW	derived	✓	✓	—
14	Okoye	Wakanda	alt	GW	derived	✓	✓	—
15	Shuri	Wakanda	second	GW	deck	✓	✓	—
16	Storm	Wakanda	alt	GW	derived	✓	✓	—
17	T'Chaka	Wakanda	alt	GW	unscored	—	—	—
18	Ms. Marvel	Wakanda	set-only	GW	unscored	—	—	—
19	She-Hulk	Wakanda	set-only	GW	unscored	—	—	—

Reading the matrix: package coverage is 15/19, score coverage is 15/19 (10 deck-inherited, 5 derived), win-line coverage is 15/19 (it tracks package/score exactly, every scored config has one), and guide links are 5. The four blanks (rows 9, 17, 18, 19) are set-only or thin-data configurations the page discloses as unscored.

Correction to my own revision 2: that version said win-line coverage was 11/19 and named Crystal, Power Pack, Bast, and Okoye as missing one. That number came from a grep capped with head -12, which silently dropped the remaining matches. Re-run without the cap: all 15 scored configs carry an explicit win-condition sentence, including all four named there (Crystal: "wins as a burst clock", Power Pack: "wins by connecting behind native evasion", Bast: "wins by going wide", Okoye: "wins by out-carding the table"). Withdrawn. The one real stylistic wrinkle: Captain America's line says "finishes through" instead of "wins through" — a word choice, not a gap.

The two guide groups, named separately

My first audit blurred these into "the four faces plus Doom," which double-counts Doom, since Doom is itself a precon face. Corrected from SNAIL_ARTICLE:

Hub deep-links (5)

Captain America, Doctor Doom, Mister Fantastic, T'Challa, Loki

Your rescored guides (5)

Captain America, Mister Fantastic, Loki, T'Challa, Kang

Reconciliation

Overlap 4 · hub-only Doom · rescored-only Kang (Kang is not in these four precons)

The linked precon-guide set and the recently rescored guide set serve different editorial purposes, so they should never be described as one list.

Findings on the page itself, revision 3

Attempting to implement the four "hub fixes" from revision 2 forced a deeper code read than the review pass had done, and it overturned two of my own four findings. Both retractions and both real fixes are logged here, verified against the live source, not asserted.

1 · Alternate-commander view now opens on. FIXED. allCmdrs defaulted to false; it now defaults to true while still respecting a returning visitor's own explicit toggle (stored in localStorage). Verified by extracting the page's real script into Node and running it against the real embedded data: a fresh visit now yields allCmdrs=true; a visit with a stored explicit opt-out yields false; a stored explicit opt-in yields true.

2 · Score provenance hint added to the compact card. FIXED, smaller in scope than first framed. The detail modal already explained "how this score was derived" in full prose for every scored commander; that part of my original finding was wrong. The real, narrower gap was the compact card view, which showed a role tag (face/second/alt) but never spelled out what it means for score provenance. Added a one-line hint, "Deck-evaluated score" for face/second, "Derived estimate" for alt, blank when unscored. Verified against all 19 real configs via the same Node harness: 6 deck-evaluated, 5 derived, 4 correctly blank, 4 face/second not double-labeled.

3 · Canonical record: WITHDRAWN, this was already built. My revision 2 said the dataset needed consolidating into one record per commander. It already is: buildCommanders() merges every source table into a single COMMANDERS array, which a comment in the source literally labels "ONE SELECTOR EVERY RENDERER CONSUMES," backed by a runtime assertCanonical() check that fails loudly on a count drift, a duplicate id, or a missing field. Running it live: 19 configs, {doom:2, ff:7, avengers:2, wakanda:8}, {grounded:15, partial:0, insufficient:1, setonly:3}, zero errors, both before and after my two edits above. I missed this the first time by reading the raw data tables and not searching for a builder function. The "stray first-avenger key" in the same finding is also withdrawn: it is not read into COMMANDERS at all (it isn't in CMDR_META), and it is not dead data either, it is the documented "bonus single-commander lists" feature, intentionally outside the nineteen-config framework.

4 · Empirical phrasing still exceeds cited evidence. Open. "Players testing repeatedly identify" and "strongest single line in the set" read as measured findings, while the footer cites only Scryfall, Wizards, Draftsim, and Gathering Games. This one was never a code question, so nothing above changes it. Convert to direct editorial statements, or name the specific tested games, and define "strongest line" as a specific category.

Closeout

Two of my four revision-2 findings survive: the empirical-phrasing wording (still open) and the two fixes above (now shipped and verified, not just proposed). Two do not survive: the canonical-record gap did not exist, and the win-line gap did not exist, both were errors in my own review process, one from a truncated grep, one from an incomplete code read. The page's real architecture, one canonical array with a runtime integrity check, full win-line coverage across every scored config, was already stronger than revision 2 gave it credit for. Reviewed page: 9.1 / 10. What is left, honestly: recommendation-drift consistency across the individual commander guides and exhaustive interface propagation (this review checked the compact card and the detail modal; it did not check every sort mode, filter, and comparison-table row for all nineteen configs) remain only partially tested.

Technical appendix (verification tooling)

Live build pinned cache-bypassed by critique_audit.py (content-hash 2bedcc68d80273c0). Composite reproduced from the page's overall() function: Doom 7.48, Avengers 7.41, FF 7.23, Wakanda 6.71, each matching the displayed value. Scryfall spot checks: set MSC released 2026-06-26; Crystal CI [G,R,U,W]; Captain America and Loki oracle text matched. Card-fact gate: 0 names parsed (JSON-embedded), so provenance rests on profiles_data.json stored Oracle text, which can age relative to the rendered recommendation. Matrix built by extracting CMDR_META, CMDR_PROFILES, EDHREC_UP, PRICES, GAME_PLANS, and SNAIL_ARTICLE keys and cross-tabulating per slug. Deep-links followed to final 200 (308 is Vercel's trailing-slash redirect). Vercel copy confirmed byte-identical to the GitHub mirror for this route.

Revision 3 fixes, verified by execution, not by reading. The page's real <script> block was extracted and run in Node against minimal DOM stubs, so the assertions below exercise the actual production code and the actual embedded data for all 19 configs, not a sample. Output: allCmdrs defaults true fresh, false after a stored explicit opt-out, true after a stored explicit opt-in; COMMANDERS.length===19, window.__canonicalErrors===[]; scoreOriginLabel() returned "Deck-evaluated score" for all 6 face/second configs, "Derived estimate" for all 5 scored alts, and correctly blank for the 4 unscored configs. The prior 11/19 win-line claim came from a grep | head -12 that silently dropped matches past the 12th; re-run uncapped, coverage is 15/19.

Open the live hub → This audit as Markdown →