Phase Space AI

Financial Model Notes

Boston Scientific [BSX]

Boston Scientific [BSX] — Financial Model Notes

Data provenance

TTM to 2026-06-30

Defects found and how each was handled

1. AV BALANCE_SHEET returns NULL for BSX cash and every debt tag → net cash computed as $0 against actual net debt of $9,576m. A $9,576m error, 12.3% of enterprise value, and it moves the required CAGR by ~1.9pp. This is exactly known-silent-failures.md §9.6 ("a completeness flag that tests only one side" / "fail closed"). Fixed by hand from EDGAR. The same defect hit EW in the comparator set — AV returned zero cash and zero debt, so EW's $49,549m EV is market cap only. 2 of 21 medtech names checked (9.5%) carry this.

2. EDGAR and AV disagree on BSX GAAP operating income. EDGAR us-gaap:OperatingIncomeLoss for the quarter ending 2026-03-31 is $1,101m; AV reports $1,072m. A $29m gap (2.7% of the line, 0.56pp of margin). The company's own CFO stated GAAP operating margin of 21.2% on the Q1 2026 call, which ties to EDGAR's $1,101m/$5,203m = 21.16%, not to AV's 20.60%. EDGAR/company adopted. This is the only check with power — a comparison against a source AV did not produce — exactly as the brief specifies.

3. AV depreciationAndAmortization is signed NEGATIVE for BSX (−$351m in the quarter ending 2026-06-30). Confirmed by av_vs_edgar.py (DNA_SIGN flag). AV's ebitda field is therefore unusable and was never read. EBITDA computed as operating income + |D&A| from the cash-flow statement.

4. AV OVERVIEW is one quarter stale, and two of its fields are mislabelled. - RevenueTTM = $20,614m = Q2'25+Q3'25+Q4'25+Q1'26 — it excludes the Q2 2026 quarter that INCOME_STATEMENT already carries. True TTM is $20,997m; a −1.8% understatement. - QuarterlyRevenueGrowthYOY 0.116 = Q1 2026, not Q2 2026. - OperatingMarginTTM = 0.206, which is exactly Q1 2026's single-quarter GAAP margin, not a TTM figure. True TTM is 21.60%. Verified as a systematic mislabel: DXCM's OperatingMarginTTM 0.214 = its Q1 2026 quarter (21.42%), and ATRC's 0.0629 = its Q2 2026 quarter (6.29%). 3 of 3 names. Do not use this field.

5. AV transcripts lose speaker identity on some utterances. Four Q4 2025 attributions return speaker: "Unknown Executive", title: "Executive", including the load-bearing PFA share-loss concession. The utterance content is intact and the prepared-vs-Q&A classification still holds (it keys off the first Analyst-titled utterance), but the specific executive cannot be named. Quoted as "BSX executive."

6. AV has BSX's Q2 2026 financials but not its Q2 2026 transcript (transcript_2026Q2.json is empty while INCOME_STATEMENT carries the 2026-06-30 quarter). The transcript endpoint lags the statement endpoint. All organic-growth disclosure in this memo therefore stops at Q1 2026; Q2 2026's reported $5,443m (+7.55% YoY) has no organic decomposition available and none is asserted.

7. A prepared/Q&A splitter keyed on the operator's "question-and-answer" phrase is wrong. The operator's opening remarks contain that phrase, so the split fired at index 0 and reported n_prepared = 0 on 6 of 13 BSX quarters and 7 of 13 DXCM quarters — silently attributing all prepared remarks to Q&A. Fixed by splitting at the first utterance whose title contains "Analyst". Worth recording: the first version produced a confident, complete-looking mention-frequency table that was wrong for half the periods.

Internal-consistency checks run

• grossProfit + costOfRevenue == totalRevenue: passes. This proves nothing (brief §"Two known AV defects", item 2) and was not relied on.
• m_EBIT,T 22.0% ≤ m_gross,T 70.0%: passes, 48pp headroom.
• Terminal margin 22.0% vs trailing actual 21.60%: above, +0.4pp. Not the §9.8 defect.
• Opex bridge reconciles: 70.0 − 10.5 − 33.5 − 3.0 − 1.0 = 22.0. Exact.
• No workbook was built this run, so §1–§3 of known-silent-failures.md (Excel named ranges, shared Excel instance, unevaluated openpyxl formulas) do not apply.