Citation Audit: The LULU Q4 FY25 Supplier Read-Through

Page references are to the PDF as distributed by Wall Street Prompt (the printed footer number is the PDF page minus 2). The report is quoted minimally, for review purposes. Nothing in this document is investment advice; it evaluates research citations, not securities.

The question

This audit answers one question: when the report states a checkable fact, does the record contain that fact?

The report is the flagship public demonstration of the supply-chain read-through prompt our forward test runs live. It presents itself as a blind ex-post reconstruction of what the workflow would have produced before LULU's 17 March 2026 print, and it discloses that reconstruction status prominently. We credit that disclosure in full, and the reconstruction question is outside this audit's scope. So is the quality of the report's predictions, which its own scorecard already grades and which the forward test addresses prospectively.

Summary of results

The report makes 126 externally checkable factual claims. We verified each against primary sources in the language of origin: SEC and HKEX filings, Taiwan MOPS revenue tables, Korean DART disclosures, earnings-call transcripts, archived aggregator captures, and the Taiwanese, Korean, and Chinese press record. Every adverse verdict was then attacked in a second pass run as defense counsel for the report. That pass overturned two of our verdicts, downgraded two more, and corrected three errors in our own working notes. The figures below are the post-review result.

*Three supplier-list claims pair a checkable list element, which verified, with an exposure percentage the report itself discloses as its own triangulation; no public support exists for those percentages, consistent with that disclosure.

By claim category:

Read as rates: verbatim quotes verify in full 7 times in 16; filing numerics 20 in 23. The sampling rule for a reader follows directly: check the quotes and attributions, spare the arithmetic.

Two findings frame everything else.

The mechanical data layer is excellent. All ten Taiwan monthly-revenue claims are exact to the printed decimal against the TWSE MOPS tables, including the year-to-date cumulatives. Every LULU actual-print figure reconciles to SEC filings, including a China segment figure that appears in no single document and reproduces only by subtracting the Q3 10-Q from the 10-K; it survives that arithmetic to the thousand of dollars. The management-guidance quotes are verbatim against call transcripts, including the seven-component Q4 gross-margin bridge. Three Chinese quotes that the PDF itself corrupts with unrendered glyphs recover and verify word for word in the cited article. Where the workflow transcribed a number from a filing, an aggregator table, or a transcript, it transcribed it correctly.

The errors concentrate where judgment about sources was required. Fourteen claims mismatch the record and three could not be found in any public source. The failures are patterned rather than random: quotes attributed to more official or more in-window sources than the ones that carried them; translations that add words the original does not contain, in each case in the direction of the report's thesis; supplier-list details that misread the list the report says it parsed, including one supplier ruled out that is on the list; a filing date inverted from pre-print to post-print; an inventory direction inverted; an analyst-action tally overstated. Sections below give each finding with its evidence.

The errors that have a direction all lean the same way. Thirteen findings adverse in whole or in part carry an identifiable direction. All thirteen run toward the report's thesis, or toward its evidence looking more official, more in-window, more contrarian, or cleaner-cut than the record supports (C036, C045, C046, C047, C053, C054, C063, C070, C072, C083, C109, C110, C125). None runs the other way. The remainder are direction-neutral identity and precision slips, plus one mechanical extraction error with no narrative content (C086).

Not all of it matters equally. Seven adverse findings are signal-relevant, meaning a reader who relied on them took away a different read than the record supports: the supplier wrongly ruled out (C030), the pre-print filing dated post-print and excluded as unusable (C036), the client-status upgrade (C053), the contradicted absence claim behind the pivot read (C054), the interpolated tariff-sharing cohort (C070), the five-times-repeated quote with no locatable source behind the soft FY26 read (C072), and the inverted inventory direction (C086). The rest are precision and identity errors, real but not read-changing.

A third finding concerns the reader's position. The PDF carries 148 link annotations: 64 are internal navigation, and all 84 external hyperlinks resolve to wallstreetprompt.com (43) or davewang.ai (41). None points to a filing, an article, or a transcript. To be precise about what this is and is not: the 84 resolve to just two unique URLs, the firms' two homepages, repeated across pages in the manner of masthead links, and there is nothing improper in a document linking its publisher. The finding is the absence those links leave behind, a 41-page evidence document whose only clickable destinations are its publisher's homepages. (Re-verified 2026-08-04: both targets are bare homepages returning HTTP 200, with no onward redirect toward any source.) The document embeds no files, and its text layer prints no source URL anywhere; the only third-party domain string in the whole text sits inside the report's own note on sources that returned HTTP 403 errors. On sourcing, the report states (p9) that the full URL bibliography for the supplier source set "sits in the supporting workflow output files", a directory of per-supplier extracts that is not part of the PDF and is not linked from it. Ten of its sixteen verbatim quotes sit in sources a reader can reach free and in full; of the five sourced to sell-side notes, one is publicly readable end to end. A reader without a Korean research terminal and a Chinese research-aggregator subscription cannot replicate the quote layer, and the report presents all quotes with uniform confidence regardless of stratum. Every error documented in this audit is invisible from inside the document; each one surfaced only on primary-source re-verification.

One number this audit cannot supply is a base rate. No equivalent claim-level audit of human-authored sell-side or independent research exists, so 66.7 percent verified grades against no curve, and we decline to guess whether a human analyst's 41-page document would score better or worse under the same protocol. The forward test publishes a provenance verification rate with every sealed run, which starts building that base rate in public.

Why this matters

The audited report is the strongest public artifact of its genre, produced by the firm that leads the genre, and that is what makes its failure profile useful: it is the best available evidence of where AI-assisted research actually breaks, and the break has a shape a reader can act on.

If you produce research with these tools. The six failure patterns documented below are what an unverified generation pipeline outputs, whoever runs it. The controls follow from the patterns: pull every load-bearing quote back to the document that supposedly carries it; print the original language beside any translation; record the edition and capture date of every foundation document you parse; enumerate before writing "all", "none", or "every"; log the searches behind any absence claim before treating silence as signal; and run one pass whose only instruction is to break the conclusion. The companion checklist (github.com/SPEC-research/forward-test/blob/main/checklist.md) states the standard in eight questions.

If you consume it. Spot-checking the numbers now samples the wrong layer: this report survives any figure check a reviewer would run, with all ten Taiwan monthlies exact to the decimal, while the quote its FY26 read leans on five times has no locatable source. The five-minute check that works: pick two or three load-bearing quotes, follow each to its claimed source, and confirm the source carries it. Ask whether verification ran as a step separate from generation, and ask to see its output. The red flags visible in any finished document: every link resolving to the producer's own domain, uniform confidence across quotes from filings and from paywalled foreign-language notes, "all" and "none" tallies, absence-of-coverage presented as signal, and translations with no original printed.

If you allocate. Figure accuracy in a manager's AI research exhibit is evidence of very little, because transcription is the layer these systems stopped getting wrong. Grade the attribution layer instead: who re-verified the original-language quotes and against what, where the bibliography is and whether it resolves outside the manager's own materials, and what edition control exists on foundation data. A manager who answers with model quality has answered.

If you own the controls. Review-by-reading is not a control for this failure class. At most two of the seventeen adverse findings here were catchable from the document alone: the broker romanization contradicted by the report's own printed Chinese, and the Bonding figure attributed to a regulatory filing on one page and to a broker note on another. Everything else required primary re-verification across three languages and five disclosure systems. The legally loaded variant is misattribution to regulatory filings, which appears twice (C070, C083). The operational conclusion: verification is a separately run function whose output ships with the research.

Reading the pattern

This audit grades claims; it alleges no intent. What the findings support is narrower and, for the category, more important: the errors follow the characteristic failure modes of AI-assisted research pipelines. Real language migrating to a more authoritative or better-timed source. Translations drifting toward the thesis. A stale edition of a foundation document persisting into conclusions its current edition contradicts. A number lifted from an adjacent row. A tally rounded up to "all". Each error is small, most have innocent mechanisms, several sit one caveat away from defensible, and every one of them survived into a finished, confident, 41-page document because generation and verification are different functions and only one of them was run. The direction of the surviving errors follows from the same fact: a generation pipeline optimizes for narrative coherence, so the claims that outlive drafting are disproportionately the ones that fit the thesis. Nothing about that requires intent. It requires only the absence of an adversarial pass.

That is the case for the category's next layer, and it is why this audit publishes alongside a live, pre-registered forward test of the same method. The method's data layer is strong enough to be worth testing live. The sourcing layer is exactly where live testing needs an independent referee.

Method

Claim extraction. Every page of the 41-page PDF was read and every externally checkable factual claim ledgered with its page, attributed source, and verification route: 126 claims across six categories (verbatim quotes 16, filing numerics 23, supplier relationships 30, calendar facts 17, consensus and actuals 22, other facts 18). Excluded from the denominator: the workflow's own predictions and confidence tags, its self-described architecture, its supplier-weighting constructs, and the five supplier exposure percentages the report's own p13 describes as triangulated "directional inputs, not audited revenue facts" (we checked those for public support anyway and report the result, but a blank there is consistent with the report's own caveat, and we grade it accordingly).

Verification. Primary sources only, in the original language, with no verdict resting on a source that repeats the report's own text. Consensus figures were checked against the aggregators the report attributes them to, using Wayback captures where the live pages have moved on; those verdicts are against aggregator data, not filings, and are labeled as such. Claims whose attributed source is paywalled or unreachable are graded UNVERIFIABLE, and UNVERIFIABLE is never counted as an error.

Disconfirmation. Every MISMATCH, NOT FOUND, and PARTIAL verdict (36 items) was then attacked in a separate pass instructed to act as defense counsel for the report: re-read the claim in context, hunt for any edition, provider, translation, fiscal convention, or rounding under which the report is right, and re-fetch the underlying evidence. Two of our verdicts were overturned outright, two were downgraded, and the pass surfaced three errors in our own notes, which we corrected. Fifteen surviving adverse verdicts carry mandatory caveats, and those caveats are printed with the findings below.

Fairness rules applied throughout. The report's disclosed caveats are credited, and a claim the report itself flags as weak is not treated as a confident assertion. Absence claims ("no coverage exists") are graded as weak-verifiable. Date and figure conventions were resolved in the report's favor wherever a legitimate convention supports its number. The supplier-list findings were tested against all three published editions of LULU's list (April 2025, October 2025, May 2026) to exclude edition drift as an innocent explanation before any mismatch was recorded.

What holds up

The strongest work in the report is exactly where its genre usually fails, and that deserves plain statement.

• Taiwan monthlies, 10 of 10 exact. Eclat January and February 2026, Quang Viet November and December 2025 and January and February 2026 plus the FY2025 cumulative, FENC December 2025 through February 2026: every figure and growth rate matches MOPS to the printed decimal, and the 17-month-high characterization for Eclat's January holds on the same inclusive count Taiwanese press used.
• The actual-print record, exact throughout. Revenue, EPS, margins, segment figures, inventory, the FY26 and Q1 FY26 guides, and the full gross-margin bridge all reconcile to the 8-K, 10-K, 10-Q, and call transcripts (C112 to C120, C122). The China Mainland Q4 figure of $528.4M is not printed in any LULU release; it reconciles exactly from the 10-K minus the Q3 10-Q, which means the workflow did the two-filing arithmetic and did it right.
• Exchange-filing and wire-service quotes verify. Regina's "brand partners' order placement decisions have stabilised" is verbatim in the HKEX interim. The Haitong International title clause is verbatim on the full-text carrier. The three FENC quotes whose characters the PDF fails to render all recover (瓶, 為, 產) and verify in the cited UDN article, byline and date included.
• The consensus layer is real. Segment-level estimates match FXStreet to the figure and to the analyst count; the -580bp margin guide and the inventory guide are verbatim from the Q3 call; the Reuters/LSEG FY26 expectations check out through syndications and CNBC independently.
• The report polices its own window at least once. It correctly identifies the widely-cited FENC Middle East commentary as post-cutoff (24 March 2026 Yuanta conference) and excludes it (C096). Its hedge on Shenzhou is exactly right: the list on its face contains no Shenzhou entity, the report caveats that tier-2 fabric exposure may not appear on a final-assembly disclosure, and that is precisely what the record shows (C014; see context note below).
• Its self-flagged weak spots are genuinely weak, honestly flagged. The FY26 China consensus range the report marks WEAK is confirmed unverifiable (C107), and the exposure percentages its p13 calls triangulations have, as that caveat predicts, no public source (C013, C018, C019, C021).

What does not hold up

Fourteen MISMATCH and three NOT FOUND verdicts survived the disconfirmation pass, along with adverse elements inside several PARTIALs. They cluster into six patterns. Where the disconfirmation pass attached a mandatory caveat to a surviving verdict, the caveat is stated with the finding here and verbatim in the appendix.

1. The supplier list the workflow says it parsed

The report's foundation document is LULU's October 2025 Active Supplier List, which it cites by name. Its facility-level readings of that list are wrong in five places, and the errors were tested against all three published editions before being recorded.

• A supplier ruled out that is on the list (C030, MISMATCH). The report excludes Stella International (1836.HK) as a LULU supplier. The October 2025 list carries Golden Star Company Limited, Hai Phong, Tier 1 Footwear, parent "Stella International Trading (Macao Commercial Offshore) Limited"; the May 2026 edition carries the same row under "Stella International Hong Kong Limited". Stella is absent only from the April 2025 edition, which suggests a stale earlier capture never re-checked against the edition the report claims.
• Crystal International: six facilities claimed, four on the list (C013, MISMATCH). "Crystal SL Global" parents exactly four rows in every edition. The only other "Crystal" row is an unrelated Colombian company.
• Hojeon: one facility in Vietnam claimed; the list shows two, both in Indonesia (C016, MISMATCH). No edition of the list has ever carried a Hojeon facility in Vietnam. Caveat: the exposure element of the claim (roughly 20 to 22 percent of FY24 apparel revenue per Korean trade press) is genuinely supported by a published analyst projection, so the mismatch is confined to the facility count and country.
• Eclat: six facilities plus one subcontractor claimed; the list shows five plus two (C011, PARTIAL). The total of seven Eclat-linked rows matches, so this reads as a classification slip rather than an invented count.
• The list-level totals (C008, PARTIAL). "Approximately 115 facilities" matches only an unstated Tier-1-plus-subcontractors reading of a 142-row list, and "roughly twenty" Asian-listed parents overcounts the 14 to 16 identifiable on any defensible inclusion rule.

Context note credited to the report (C014, VERIFIED): its claim that Shenzhou is absent from the October 2025 list is true as printed, since the relevant row's printed parent is "Master Limited (Vietnam)". The analytically important context is that Shenzhou's exposure is on the list through that row (Gain Lucky (Vietnam), Tier 2, Fabric-Vertical, Tay Ninh), and a Shenzhou-named Tier 1 facility appeared on the April 2025 edition. The report's own tier-2 hedge anticipated exactly this.

2. Quotes and figures attributed to sources that do not carry them

Four claims move real material onto wrong sources. In each case the move makes the evidence look more official or more in-window than it is.

• The Hojeon tariff quote (C070, PARTIAL). The Korean sentence naming tariff cost-sharing as a Q4 drag is genuine company language, verbatim in three independent 27 February 2026 articles. The report attributes it to the 26 February DART filing. The filing's entire reason line reads 환율변동 및 관세 영향으로 인한 이익 감소 (profit decrease due to FX movements and tariff impact); the quoted sentence evidently comes from Hojeon's IR communication to media. See also pattern 3 on this quote's translation.
• The December Shenzhou quote that appears to be an August quote (C063, NOT FOUND). "Actively cultivating segment leaders such as Lululemon" is attributed to a Hexun sector strategy of 7 December 2025. That article exists and contains no Lululemon mention. Near-identical language sits in a Huaxin Securities Shenzhou initiation dated 27 August 2025, verified from the note PDF, and 华鑫 (Huaxin) is phonetically adjacent to 和讯 (Hexun). Caveat: the quote is probably real but misattributed rather than invented, and a China-reachable browser check of the Hexun page remains open as a final control. If the attribution is wrong as it appears, a pre-window August quote was moved inside the December observation window.
• Hojeon net income from the wrong document (C067, MISMATCH). The claimed KRW 11.3 billion (-51.2%) does not match the cited preliminary disclosure (KRW 11.12 billion, -52.0%); it matches the later audited annual report. Caveat: the figure is the later audited actual, not an invented value, and a May 2026 author reading FnGuide would see exactly these numbers. Revenue and operating income in the same claim are exact.
• The Regina Bonding figure (C083, MISMATCH). The scorecard attributes "Bonding +40% with 'Lululemon Air Support' cited" to Regina's interim results. The interim contains no +40% figure, no Lululemon, and no Air Support (zero occurrences in both the results announcement and the full interim report); the segment figures in the same claim are exact. The +40% is genuine content of the CMBC International initiation the report cites adjacently. Caveat: this is misattribution of a real figure plus an unverifiable product claim, not fabrication of the number.

The same pattern appears in miniature inside a verified claim: the Eclat "flat to low" customer-inventory quote verifies against CNA's 13 November 2025 coverage, and the report's p25 restatement re-dates it to the 5 March 2026 board statement, whose article contains no customer-inventory sentence (C046, VERIFIED with that caveat).

3. Translations that add words

Two quoted translations insert phrases the original does not contain, and both insertions run toward the report's thesis.

• "With US customers" (C070). The Korean names 미국발 관세 이슈, tariff issues originating from the US, and contains no word for customers. The report's quoted translation inserts "with US customers" and then builds its tariff cost-sharing argument on the inserted cohort. The direction of the inference is plausible; the counterparty is unstated in the Korean. In fairness, the report repeatedly caveats that this is a cohort phrase rather than a LULU-specific one.
• "Of four major clients" (C053, MISMATCH). The Taiwanese record supports "fastest-growing" for LULU at Quang Viet and simultaneously frames LULU as a newly acquired, low-base customer outside the named major-client set (the majors are listed and LULU is explicitly set apart: 除了去年加入的Lululemon外). The insertion upgrades LULU into the major-client cohort, which changes the signal: a top-four client decelerating is a different fact from a low-base new client growing fast.

4. Quotes with no locatable source

• The five-times-repeated management quote (C072, NOT FOUND). "High-base growth slowdown at large-share customers", attributed to Hojeon management and load-bearing for the report's soft FY26 read, appears in no locatable Korean source: the DART filing lacks it, the three same-day press articles lack it, indexed Korean coverage lacks it, and the one findable broker document reads opposite in tone (Daishin, September 2025: major customers including Lululemon keep growing). If it exists it sits inside a paywalled note or an unindexed interview; as printed it is unsupported, and no Korean original is given.
• "Margin normalisation expected" (C047, PARTIAL). Attributed to Eclat's CFO on 5 March 2026. Caveat: the 5 March article does carry backward-looking margin-recovery language (Q4 gross margin 30.19 percent, back above 30 percent), so the quote has a real kernel in the claimed vehicle; forward-looking normalisation language appears in the record only around the 17 March investor conference.
• "Tale of two markets" (C124, PARTIAL). Absent from both accessible full transcripts of the call. It appears in third-party earnings summaries that attribute it to management. Caveat: the report most likely echoes secondary summaries rather than coining the phrase; it remains unplaceable as call language on the primary record. The claim's other two phrases are verbatim in the transcript.

5. Window and absence claims

• A filing date inverted (C036, MISMATCH). The report dates Tristate's FY25 annual results to 27 March 2026, ten days post-print, and excludes them as unusable. The results were filed 17 March 2026 at 21:38 HKT, roughly seven hours before LULU's release. Nothing was filed on 27 March. The error inverts the report's own usability framing: the filing was pre-print.
• The absence claim behind the Quang Viet read (C054, MISMATCH). The report states LULU is not mentioned in any in-window article on Quang Viet and reads the silence as signal. Commercial Times, 17 December 2025 print (online 16 December), inside the window, ties Quang Viet to LULU down jackets and outerwear with orders up 20 percent this year. The secondary element (a 30 percent Jordan expansion earmarked for Alo Yoga) conflates two separate 2025 disclosures.
• A conference that did not happen on that date (C042, MISMATCH). The report has Eclat presenting at a Yuanta conference on 12 February 2026. The MOPS record dated 12 February announces attendance at a Yuanta conference on 17 March, Eclat's Q4 call. Caveat: Eclat did attend J.P. Morgan's Taiwan CEO-CFO Conference on 17 February 2026, so an in-window mid-February broker appearance existed; the report has the wrong host and date for it, or converted an announcement date into an event date.
• The Eclat order-book trajectory (C045, PARTIAL). Three of its four touchpoints verify or are fairly inferred, including full-capacity wording we initially missed in the 5 March article and corrected on review. The August anchor is misdated: six-month visibility was restored by 7 August 2025, and the three-month reading dates to May 2025. The misdating moves the order-book recovery inside the observation window when it predates it.

6. The market-record layer

• "All six firms cut" (C110, MISMATCH). The report has Goldman, BTIG, Telsey, Evercore, UBS, and BofA all cutting LULU price targets in the seven days before the print. Four did (UBS 9 March, Goldman 11 March, BTIG 12 March, Evercore 16 March). Telsey published a Market Perform reiteration on 10 March with an unchanged $215 target, and BofA took no action between its December 2025 raise and June 2026. Caveat: the report's "uniformly negative direction of travel" gloss is directionally fair; "all six" is not.
• The rating distribution (C109, MISMATCH). The claimed 1 Strong Buy / 2 Buy / 28 Hold / 2 Sell across 33 analysts and a $190 to $200 median target match no archivable aggregator: MarketBeat's 12 March capture shows 2/2/30/1 of 35 with a $225.22 mean, Benzinga's 17 March capture shows 25 analysts at $210.40. Caveat: the claim's shape is right (85 to 90 percent Hold, real final-week cuts, a low-side cluster at $175 to $192), no checkable source supports the printed numbers, and a terminal source such as LSEG or Bloomberg could still vindicate them.
• The Q1 FY26 EPS consensus (C108, MISMATCH on that element). The only locatable contemporaneous print is $2.07 (LSEG via CNBC) against the claimed $2.09. Caveat: a different aggregator could have printed $2.09 and none was found; $2.09 is not disproven. The revenue element corroborates.
• The day-after move (C125, PARTIAL). The 17 March close is exact, and the claimed after-hours snapshot ($161.98, +1.3 percent) is defensible as an initial reaction measured against the 16 March close, the standard extended-hours convention, in a session Reuters documents as two-sided and ending lower. The surviving error: 18 March closed up 3.84 percent, and the report's "~+5%" matches only an intraday print.
• One extraction inversion (C086, MISMATCH). The report has Regina's inventory building 18.5 percent half-on-half. The interim shows inventories declining 9.6 percent against 31 March 2025 and 15.0 percent year on year. One arithmetic identity is worth recording: inventories at 30 September 2025 divided by the adjacent trade-receivables comparative equals exactly +18.5 percent, consistent with a row-misalignment extraction error. We offer that as a plausible mechanism, not a finding. The capex elements of the same claim are exact.

One naming error rounds out the set: the broker behind the December Shenzhou notes is Guosheng Securities (国盛证券), legible in the report's own partially rendered Chinese, romanized throughout as "Guosen", which is a different, real firm (C062, MISMATCH). Caveat: all four underlying broker documents are genuine and correctly dated; in a report sold on original-language sourcing, pointing readers at the wrong firm is a substantive error rather than a typo.

What a reader can and cannot check

The report's premise is that its edge comes from primary sources in three languages. The audit confirms the sources are overwhelmingly real. It also establishes that the report, as distributed, gives its reader no path to them.

• The PDF's annotation table, embedded-file table, and text layer were each inspected in full. Of 148 link annotations, 64 are internal navigation and all 84 external hyperlinks resolve to wallstreetprompt.com (43) or davewang.ai (41); the document embeds no files; and no text-layer string points to any source. Re-extracted and redirect-tested 2026-08-04: the 84 external annotations carry two unique homepage URLs, both returning HTTP 200 with no onward redirection. The bibliography statement on p9 places the full URL bibliography in "the supporting workflow output files", per-supplier extracts in a directory the PDF neither contains nor links.
• Of the 16 verbatim quotes, 10 sit in sources a reader can reach free and in full. The sell-side stratum is the weak layer: of five quotes sourced to broker notes, one (Haitong) is publicly readable end to end; the DB Securities quote is reachable only through Korean press excerpts; the CMBC International and Wedbush notes are paywalled or unlocatable.
• The report presents all quotes with uniform confidence regardless of stratum. Nothing in its text distinguishes the quote a reader can click through to HKEX from the quote that exists, if at all, behind a Korean research terminal.

This matters because of where the errors sit. The transcription layer, which a reader could in principle spot-check, is nearly flawless. The errors live in the attribution and translation layer, which a reader cannot check at all without redoing the sourcing from scratch. We had to: the re-verification ran across the primary record in Chinese, Korean, and English, five disclosure systems (SEC EDGAR, HKEX, TWSE MOPS, DART, and archived aggregator captures), and the region-locked press stratum. That is the cost profile of the control this genre is missing.

Unverifiable items

Two claims are graded UNVERIFIABLE. Neither is counted as an error.

• C082. The CMBC International initiation on Regina Miracle exists, is correctly dated 1 December 2025, and its accessible abstract confirms Bonding growth figures and Lululemon among customers. Whether it "explicitly names Lululemon Air Support [leggings]" cannot be confirmed; the note body is paywalled, and no accessible summary links Lululemon to Air Support. One independent observation: Air Support is a Lululemon sports-bra line, which fits Regina's specialty; the report's own bracketed "[leggings]" is likely a product-category error even if the note names the product.
• C111. The Wedbush "FY26 Street is conservative" view could not be located through any public channel, and Wedbush is absent from the findable pre-print commentary set. A paywalled note may exist. The claim's embedded proxy-fight fact verifies (CNBC, 17 March 2026); the 15-quarter beat-streak fact was not cleanly checkable in public archives and remains open.

Our own errors, and what could change

The disconfirmation pass exists because auditors make the same class of mistakes as the audited. Ours, for the record: we initially misdated UBS's March cut, missed the full-capacity wording in the 5 March CNA article, wrote that the Eclat-Nulu tie had no public support when a secondary source exists, over-penalized a hedged Shenzhou claim that is true as printed, and applied segment-level pedantry to an accurate description of FENC. All five are corrected in our working files, and the two affected verdicts were re-graded in the report's favor before this draft was assembled.

Open items that could move specific verdicts, listed so that anyone with the access can check us: the early-March 2026 LULU rating distribution and median target from a professional terminal (C109); any second aggregator print of the Q1 FY26 EPS consensus (C108); the Daishin, NH, DB, CMBC, and Wedbush note PDFs (C078, C075, C076, C082, C111); the Hexun article from a China-reachable browser (C063); the customs entry-effective date of the 19 percent Bangladesh rate (C057). If any verdict here is wrong, we will correct it in place, dated, and say so at the top of this document.

Publication note, 18 August 2026

Two disclosures that sit outside the audit as delivered on 10 August. They are recorded here rather than in the text so that the delivered document and its published hash stay identical.

How this audit was run. The claim ledger, the primary-source retrieval and the first-pass grading were produced with AI assistance: language-model agents fetching and reading the underlying filings, transcripts, revenue tables and press articles in their original languages, each verdict recorded against the specific document it rests on. The disconfirmation pass that attacked our own adverse verdicts ran the same way under a separate instruction, and it overturned two of our verdicts, downgraded two more and corrected three errors in our notes. SPEC Research reviewed the output and is responsible for every verdict as published. We state this because the audit asks readers of AI-assisted research to ask who re-verified the sources and against what. That question applies to us. The answer that matters is in the appendix: every verdict names the document it rests on, so a reader can check the record rather than trust the process.

On the link annotations. The audit reports 84 external link annotations resolving to two homepages. The geometry, for precision: 80 of the 84 sit in two identical rectangles repeated in the page footer across 40 pages, which is template chrome and carries no implication. Four are placed in the body of the document, and those four resolve to the same two homepages. The finding is the one the audit states, unchanged: the document's only clickable destinations are its publisher's.

Right of reply

This draft was sent to Wall Street Prompt on Monday 10 August 2026 with five business days, through 5:00 PM New York time on Monday 17 August 2026, to respond before publication. Any response is printed verbatim and unedited alongside this audit, at whatever length. Corrections they substantiate will be made in the text, attributed.

Response from the report's authors

No response was received at either address or through LinkedIn by the deadline of 5:00 PM New York time on Monday 17 August 2026. Delivery and receipt were not independently confirmed. The offer does not expire: a response provided for publication at any point prints here verbatim and unedited, at whatever length, and any finding shown to be wrong is corrected in the text and credited.

Appendix: the full claims ledger with verdicts

Verdict key: V = verified in full against the primary or attributed source. V-list = the supplier-list element verifies; the exposure element is the report's own disclosed triangulation, for which no public support was found. CORR = corroborated via secondary sources only. P = partial. MM = mismatch. NF = not found in any public source. UNV = attributed source inaccessible; graded neither right nor wrong.

Consensus rows marked V against FXStreet, Zacks, or Reuters/LSEG are verified against the report's attributed aggregator, which is itself aggregator data rather than a filing.

A. Framing and calendar (pp. 5-10)

B. Supplier discovery, exposure, non-suppliers, HK calendar (pp. 11-14)

C. Per-supplier extracts (pp. 15-21, 24)

D. Pre-print consensus (pp. 30-31)

E. Actual print and reaction (pp. 32-34, 38)