Verification Report — Startups Research Library

Audit date: September 15, 2026 Scope: Spot-check of the highest-risk factual claims across 00-executive-summary.md and chapters 01–12. Method: Direct fetches of primary/authoritative sources (BLS, ycombinator.com, techstars.com, sbir.gov, Crowell & Moring, federalregister.gov, leg.colorado.gov, irs.gov, lw.com, mintz.com, metr.org, carta.com, news.crunchbase.com, blogs.nvidia.com, techcrunch.com, lemonsqueezy.com, artificialintelligenceact.eu, foley.com).

Process note: this session's WebSearch quota was exhausted before the audit began, so every check was performed by direct URL fetch of a primary or near-primary source. Where a source could not be reached (paywall/bot block), the item is marked UNVERIFIABLE rather than guessed at.

1. Defects found

D1 — ERROR — Chapter 01 BLS survival figures are wrong (two of four)

• File/location: 01-startup-ecosystem-and-taxonomy.md, lines 86–96 (survival table for the March 2005 establishment cohort).
• Claim as stated: March 2005 cohort survival of 80.1% at 1 year, 48.2% at 5 years, 31.2% at 10 years, 19.5% at 20 years, cited to BLS BED Table 7.
• What I found: BLS us_age_naics_00_table7.txt, "survival rates since birth" column for the establishments born March 2005, reads: March 2006 80.1, March 2010 (year 5) 46.8, March 2015 (year 10) 33.8, March 2025 (year 20) 19.5. The 1-year and 20-year values are correct; the 5-year and 10-year values do not appear anywhere in that cohort's row.
• Severity: ERROR (the chapter labels the table as drawn directly from the cited BLS file).
• Recommended correction: change 48.2% → 46.8% (5 years) and 31.2% → 33.8% (10 years). The surrounding prose ("roughly half of new establishments survive five years") should be softened to "just under half" for this cohort, which spans the Great Recession.

D2 — ERROR + INCONSISTENT — Executive Summary repeats the bad figures and attributes them to the wrong chapter

• File/location: 00-executive-summary.md, §2.2.
• Claim as stated: "Chapter 10 traced the figure and found no underlying study. The actual administrative data — BLS Business Employment Dynamics, tracking the March 2005 establishment cohort through twenty full years — shows 80.1% surviving one year, 48.2% five years, 31.2% ten years, and 19.5% twenty years."
• What I found: two problems. (a) The figures are Chapter 01's, not Chapter 10's, and two of them are wrong (see D1). (b) Chapter 10 (10-pitfalls-and-failure-modes.md, lines 62–71) uses a different cohort entirely — March 2015 — reporting 79.6% / 50.2% / 34.7% at 1/5/10 years, plus 1-year rates of 76.3% (2022), 78.2% (2023), 77.9% (2024). Every one of Chapter 10's numbers checks out exactly against BLS Table 7.
• Resolution of the apparent chapter-to-chapter contradiction: the two chapters are measuring different birth cohorts, so they are not strictly contradictory — but Chapter 10 is right and Chapter 01 is wrong, and the Executive Summary credits the wrong chapter with the wrong numbers.
• Severity: ERROR.
• Recommended correction: attribute to Chapter 01, correct the two values per D1, and add a clause noting Chapter 10 reports the March 2015 cohort (79.6/50.2/34.7) — otherwise a reader comparing the two chapters will conclude the library contradicts itself.

D3 — STALE — Executive Summary presents a February 2026 stock price as a September 2026 status

• File/location: 00-executive-summary.md, §2.12.
• Claim as stated: "Status changes verified as of September 2026 include ... Figma trading around $24 after an 80% drawdown."
• What I found: 09-case-studies.md line 317 dates that price precisely: "By February 2, 2026 the stock was at $24.00," sourced to Wolf Street. That is a seven-month-old quote being presented under a September 2026 verification banner.
• Severity: STALE.
• Recommended correction: either re-verify the current price or restate as "Figma down roughly 80% from its August 2025 peak (at $24 as of February 2026)."

D4 — INCONSISTENT — Executive Summary misattributes four enforcement/litigation items to Chapter 11

• File/location: 00-executive-summary.md, §2.10, final paragraph.
• Claim as stated: presented as part of what "Chapter 11 re-verified": "Healthline was fined $1.55M and Tractor Supply $1.35M under California privacy law; Anthropic settled copyright litigation for $1.5 billion; Moffatt v. Air Canada established that a company is responsible for what its chatbot tells a customer."
• What I found: none of these four appear in Chapter 11. All are in Chapter 10 (10-pitfalls-and-failure-modes.md lines 433, 434, 448, 469). The underlying facts are correct (see §2 below); only the attribution is wrong.
• Severity: INCONSISTENT (sourcing).
• Recommended correction: attribute to Chapter 10.

D5 — INCONSISTENT — Executive Summary mischaracterizes the CB Insights sample

• File/location: 00-executive-summary.md, §2.2.
• Claim as stated: "CB Insights' March 2026 edition samples 431 self-published post-mortems, 70% citing 'ran out of capital.'"
• What I found: 10-pitfalls-and-failure-modes.md line 94 says CB Insights analysed 431 VC-backed companies that shut down since 2023, of which 385 had identifiable failure reasons, drawn from "public post-mortems, founder interviews, and shutdown announcements." The 431 are companies, not post-mortems; the reason percentages rest on 385; and the sources are not all self-published.
• Severity: INCONSISTENT / MINOR.
• Recommended correction: "analyses 431 shut-down VC-backed companies, 385 with identifiable causes, 70% citing 'ran out of capital.'"

D6 — MINOR — Executive Summary presents a projection and an author's calculation as verified fact

• File/location: 00-executive-summary.md, §2.1.
• Claim as stated: "roughly 10,000 companies raise a first institutional venture round each year. That is 0.19% of applications, or 0.57% on the more generous denominator."
• What I found: 01-startup-ecosystem-and-taxonomy.md lines 110–112 derives the 10,000 by annualising 5,674 H1 2026 first-time financings and explicitly says the year is "on pace for more than 10,000 ... which would be a record," and labels the percentages "Author's calculation." The Executive Summary drops both the record-year caveat and the calculation label, converting a one-off projection into a steady-state fact. Note this cuts against the claim, not for it: in a non-record year the percentage would be lower.
• Severity: MINOR (estimate presented as fact — the library's own methodological standard).
• Recommended correction: "in 2026, a record year, roughly 10,000 companies are on pace to raise a first venture round — about 0.19% of applications (author's calculation)."

D7 — MINOR — Executive Summary omits the Techstars Boulder revival that Chapter 08 documents

• File/location: 00-executive-summary.md, §2.9.
• Claim as stated: "Techstars moved to $220K for 5% common in April 2025 after closing its Seattle, Boulder, and DC programs."
• What I found: 08-accelerators-investors-support.md line 98 correctly notes Boulder was closed in 2024 and later revived, and that it appears on the current program list. I confirmed on techstars.com (September 2026): the accelerator list carries Techstars Boulder; Seattle and Washington DC are absent. ~17 programs listed, consistent with the chapter's "roughly 16."
• Severity: MINOR.
• Recommended correction: "after closing Seattle and DC, and closing and later reviving Boulder."

D8 — MINOR — One-day inconsistency on the Anthropic settlement approval date

• File/location: 10-pitfalls-and-failure-modes.md, line 448 vs. source list line 750.
• Claim as stated: body text says "final approval came on 21 July 2026"; the chapter's own source entry is dated 20 July 2026.
• What I found: the cited TechCrunch article (published July 21, 2026) states approval was granted Monday, July 20, 2026, by Judge Araceli Martínez-Olguín. Amount ($1.5B), structure (~$3,000/work across ~500,000 works) and the "settled, not appealed, therefore not binding precedent" framing all check out.
• Severity: MINOR.
• Recommended correction: 20 July 2026.

D9 — MINOR / UNVERIFIABLE — The "$80B Azure backlog" is presented as a Microsoft disclosure but rests on one secondary analyst

• File/location: 02-sector-analysis.md lines 213, 680, 758; repeated in 00-executive-summary.md §2.4.
• Claim as stated: "Microsoft has disclosed an $80 billion backlog of Azure orders it cannot fulfill because of power constraints."
• What I found: the only citation is Futurum Group, a paid industry-analyst shop — no Microsoft filing, earnings call or press release is cited. Unlike nearly every other figure in Chapter 02, this sentence carries no epistemic label, yet the wording ("has disclosed") asserts a first-party disclosure. I could not trace it to a Microsoft source. This figure is load-bearing: it is the single specific datum behind one of the seven "genuinely promising" sector verdicts.
• Severity: UNVERIFIABLE as stated.
• Recommended correction: either cite the Microsoft disclosure directly or relabel as "[Reported] Futurum estimates an ~$80B Azure order backlog constrained by power availability."

D10 — MINOR — Unreconciled Carta-vs-Carta figures on the 2024 solo-founder share

• File/location: 05-founders-teams-and-gtm.md, lines 34 and 40.
• Claim as stated: line 34 cites Carta's Founder Ownership Report 2026 for solo founders at "36% of new startups in 2025, up from 31% in 2024"; line 40 cites Carta's Solo Founders Report for "30% of startups founded in 2024."
• What I found: both are accurately transcribed from their respective Carta reports; the chapter simply places two slightly different Carta numbers for the same year six lines apart without reconciling them.
• Severity: MINOR.
• Recommended correction: add a half-sentence noting the two reports were cut at different dates.

D11 — MINOR — SBIR lapse date phrased two ways (not an actual contradiction)

• File/location: 07-funding-and-financing.md line 129 vs. 08-accelerators-investors-support.md line 469.
• Claim as stated: Ch07: authorization "expired September 30, 2025." Ch08: "lapsed on October 1, 2025."
• What I found: these are consistent, not contradictory — the authorization ran through September 30, 2025, so the lapse began October 1. Confirmed against the Crowell & Moring alert both chapters cite: expired September 30, 2025; reauthorized April 13, 2026; extended through September 30, 2031, via the Small Business Innovation and Economic Security Act of 2026 (S. 3971).
• Severity: MINOR (wording only — flagged because a reader comparing the two chapters may read it as a discrepancy).
• Recommended correction: harmonise to "expired September 30, 2025 and lapsed from October 1."

D12 — MINOR — "19% slower" vs. "19% longer"

• File/location: 00-executive-summary.md §2.6 ("19% slower") vs. 03-idea-validation-and-product-building.md line 351 ("took 19% longer").
• What I found: METR's own reporting is "developers take 19% longer." A 19% increase in time is a ~16% reduction in rate, so "19% slower" is loose. METR itself is frequently summarised both ways.
• Severity: MINOR.
• Recommended correction: use "took 19% longer" in the summary to match the chapter and the source.

2. Checked and found correct

Y Combinator (Chapter 08) — fully verified on ycombinator.com, September 2026

• $500K standard deal: $125,000 post-money SAFE for 7% plus $375,000 uncapped SAFE with MFN — confirmed on /deal, including the ~2.5% worked example at a $15M post-money next round, pro rata rights, and no fees.
• Four batches a year (Winter, Spring, Summer, Fall); currently open for Winter 2027, running January–March in San Francisco; on-time deadline November 2 at 8pm PT; decisions by December 11 — confirmed on /apply. Every element of the chapter's cadence claim is correct, including the W27 deadline.
• The chapter's caution that batch-size figures are scraped third-party data (Ellenox) is appropriately labelled [Reported].

Techstars (Chapter 08) — verified on techstars.com

• $220,000 total, announced April 17, 2025: $20,000 post-money Convertible Equity Agreement converting to 5% common plus a $200,000 uncapped MFN SAFE, both converting on a priced round of ≥$1M; no program fee; Asia-Pacific uses a $100,000 SAFE. All confirmed verbatim.
• Program closures: Seattle and Washington DC absent from the current list; Boulder present (see D7).

SBIR/STTR (Chapters 07 and 08)

• Expired September 30, 2025; reauthorized April 13, 2026; extended through September 30, 2031 (Small Business Innovation and Economic Security Act of 2026, S. 3971) — confirmed.
• Strategic Breakthrough Awards up to $30 million, 100% private/non-SBIR matching requirement, 48-month maximum performance period, prior Phase I/II award required — confirmed.
• FY2027 agency-set proposal limits — confirmed.
• Phase I cap $323,090, Phase II cap $2,153,927 — confirmed on sbir.gov.
• The two chapters state the lapse consistently (see D11).

AI share of venture funding (Chapters 01, 02, 07) — verified and internally consistent

• Crunchbase (July 2026) confirms verbatim: $510 billion global H1 2026 (record), against $440 billion for all of 2025; more than 70% of global Q2 capital to AI; OpenAI and Anthropic $217 billion = 43% of H1; Anthropic $65 billion in Q2.
• PitchBook-NVCA figures are internally consistent across chapters: US H1 2026 deal value $412.7B, AI share $355.9B — 355.9 ÷ 412.7 = 86.2%, matching the 86% claim in all three chapters. Megadeal share 87.5% stated identically in Ch02 and Ch07.
• Chapter 02 line 22 explicitly explains the 70%/80%/86% spread as definitional rather than contradictory. This is handled well and is not a defect.

Nvidia / Hugging Face (Chapter 09) — verified against NVIDIA's own announcement

• NVIDIA's blog confirms: "NVIDIA has agreed to acquire Hugging Face for $12,930,300,000," announced September 3, 2026. The chapter's $12.93B figure and date are exact.
• The chapter correctly states the deal is announced, not closed, and flags antitrust review as plausible — appropriate hedging.
• (The CNBC corroborating link in the source list returned HTTP 403 to an automated fetch; this is a bot block, not evidence the article is absent. The NVIDIA primary source is sufficient.)

Legal and regulatory dates (Chapter 11) — 7 of 7 verified

State non-compete items also check out: Tennessee HB 1034 and Virginia SB 128/SB 170 effective July 1, 2026; Washington HB 1155 effective June 30, 2027 with retroactive reach.

Anthropic settlement and Moffatt (Chapter 10)

• $1.5 billion, approved July 20, 2026 (see D8 for the one-day slip), ~$3,000 per work across ~500,000 works, Judge Araceli Martínez-Olguín after Alsup's retirement. The chapter's key qualification — a settled district-court matter, therefore not binding precedent — is correct and important.
• Moffatt v. Air Canada (BC Civil Resolution Tribunal, February 2024): the quoted holding, the rejected "separate legal entity" defence, and the ~CAD $650 damages plus interest and fees are accurately stated.
• Healthline $1.55M (California AG, July 1, 2025) and Tractor Supply $1.35M (CPPA, September 30, 2025) are accurately stated and sourced to the regulators' own releases.

Bench and Lemon Squeezy (Chapter 04)

• Bench ceased operations December 27, 2024; Canadian bankruptcy filings of January 7, 2025 disclosed $2.8M cash against $65.4M liabilities; Employer.com announced the acquisition December 30, 2024, three days later. Internally coherent and correctly sourced; the chapter properly flags that Pilot and Acuity explainers are competitor-published.
• Stripe acquired Lemon Squeezy — announcement dated July 26, 2024 on Lemon Squeezy's own blog. The chapter's "July 2024" is correct.

Carta founder data (Chapter 05)

• 23.7% → 36.3% solo-founder share, 2019 → H1 2025 — verified verbatim on carta.com.
• 30% of 2024 startups solo-led, capturing 14.7% of priced-round capital — verified verbatim.
• 51–49 median split: correctly presented in Chapter 05 as the median for two-founder SaaS teams over the three years to 2025 (moving from 60–40 in 2019), with biotech lead founders at 58%. The Executive Summary's flat "Carta's median split is 51–49" compresses this but does not misstate it.
• The debunking of "65% of startups fail from co-founder conflict," and the 25–35% five-year breakup rate (40%+ at eight years for 2016–2018 cohorts), are handled carefully and labelled correctly.

METR and Veracode (Chapter 03)

• METR verified exactly against metr.org: 16 experienced open-source maintainers, 246 real issues, 19% longer with AI tools, having forecast a 24% speedup and still believing afterwards they were sped up ~20%. The chapter reproduces METR's own generalizability caveat, which is the right call.
• Veracode ~45% of AI-generated code contains vulnerabilities: the chapter labels this "[Verified with vendor-interest caveat]" and names the conflict. I did not independently re-fetch the Veracode report within this audit's budget — see §3.

BLS figures in Chapter 10 — correct

All six values check out exactly against BLS Table 7: March 2015 cohort 79.6 / 50.2 / 34.7 at 1/5/10 years; 1-year survival of 76.3% (2022), 78.2% (2023), 77.9% (2024).

Other Executive Summary figures cross-checked against their chapters and found faithful

36.2M small businesses and 82.3% nonemployer (Ch01); climate Series C −32% to an all-time low with Series D+ +78%, agtech deeptech 78% seed premium with zero $200M+ rounds, RevenueCat 52%/30%/20%, ~14,666 gaming layoffs, ~a quarter of $5.7T US health spending administrative (Ch02); Pew 8% vs 15% click-through and SparkToro 68% zero-click (Ch05); Aleph × Benchmarkit n=342, GRR 84% down four points, magic number 0.94 → 1.37, Model 4's $0.90 → $0.234 per resolution and −168% for heavy users after optimization, High Alpha 800-company panel ~10 points of early-stage compression (Ch06); Reg CF median $113,000 and 0.25% IPO rate, Pipe's exit from standalone RBF (Ch07); Howell & Hall, SMJ 2026, doi 10.1002/smj.70103 — a real and correctly described citation (Ch05); TinyPilot $598K exit, Ginkgo $20M Q2 2026 revenue after a 1-for-40 reverse split, 23andMe/TTAM (Ch09); ~$1,000 most-likely year-one revenue (Ch12).

No cross-chapter contradictions found on shared figures

Beyond the BLS cohort issue (D1/D2), I found no case where two chapters state materially different values for the same underlying quantity. The $510B/$440B/$217B/43% figures, the $412.7B/$355.9B/86%/87.5% figures, the SBIR dates, the co-founder breakup range and the solo-founder shares are stated consistently wherever they recur.

3. Not verified (reported as unverified rather than guessed)

• Veracode's ~45% vulnerable-code finding (Ch03) — not independently re-fetched in this audit. The chapter already labels the vendor conflict of interest. Treat as chapter-verified, auditor-unverified.
• Bumble's $417M market cap and $2.89 share price (Ch09, repeated in the Executive Summary) — sourced to StockStory/FinancialContent, August 2026; not independently re-verified.
• Ellenox YC batch-size counts (Ch08) — third-party scrape; the chapter already labels it [Reported] and notes YC does not publish batch sizes.
• The March 2026 FTC click-to-cancel ANPRM and its April 13, 2026 comment deadline (Ch11) — the vacatur itself is verified; the revival rulemaking detail was not separately confirmed.
• Futurum's $80B Azure backlog — see D9.

4. Summary

Twelve defects. Two are ERROR-level and share a root cause: two incorrect BLS survival percentages in Chapter 01 that the Executive Summary repeats while crediting them to Chapter 10, whose own (different-cohort) figures are correct. One item is STALE. The remainder are attribution, labelling and wording issues.

The high-risk claims most likely to have been wrong — the Nvidia/Hugging Face acquisition, YC's four-batch cadence and W27 deadline, Techstars' April 2025 terms, the SBIR lapse and reauthorization, seven legal effective dates, the AI funding share, Carta's founder data and the METR study — all verified exactly against primary sources. The library's verification discipline is, on this sample, materially better than its Executive Summary's transcription discipline: every ERROR-level defect found is in how a chapter's finding was carried forward, not in the chapter research itself, with the single exception of D1.