How does DEPLOY verify physical AI claims?

DEPLOY exists to distinguish verified from claimed

Physical AI coverage is full of capability inflation, regulatory-state conflation, deployment-overstating, funding misattribution, and founder-pedigree errors. Most aggregators flatten the distinctions. Most publishers don't surface them. DEPLOY's framework discipline operates as the verification protocol that makes the distinction systematically.

This piece walks through the framework operationally: what verified-vs-claimed means at each tier, how the source-quality rubric weights primary vs secondary vs aggregator content, what cap-flag convention does in practice, the canonical worked examples that demonstrate the discipline across cohorts, how cohort architecture organizes verification at form-factor-cell granularity, and how multi-agent operational discipline + open transparency complete the credibility surface institutional partners can audit.

The framework isn't methodology for its own sake. It's the verification protocol that makes DEPLOY a reference.

---

The verified-vs-claimed framework

Every claim that surfaces on DEPLOY's registry, news, or consumer property resolves to one of four verification tiers:

• Verified: primary-source-confirmed. The claim resolves against SEC filings, FDA databases, company investor relations communications, peer-reviewed publications, or other primary-source evidence with traceable provenance.
• Stated: operator-asserted at source-quality posture. The claim is what the operator (maker, regulator, customer) said publicly with appropriate source quality, but independent verification at primary-source depth hasn't completed.
• Claimed: asserted without primary-source confirmation. The claim appears in aggregator coverage, secondary reporting, or marketing materials but lacks the primary-source anchor that would move it to verified or stated tier.
• Absence: honest-not-known. The fact would be load-bearing if known but isn't; DEPLOY surfaces the absence rather than papering over it with claimed-tier content.

Why the distinction matters: physical AI deployment claims compound errors structurally. A funding round attributed to the wrong lead investor surfaces in 50 aggregator pieces. An FDA clearance scope inflated from "remote monitoring" to "diagnosis" reframes capability claims across downstream coverage. A pilot framed as scaled deployment reshapes investor and customer expectations. The verified-vs-claimed framework catches these structurally before they compound.

---

Source-quality tier rubric

DEPLOY weights source quality structurally across four tiers:

Primary sources carry highest verification weight. These include SEC filings (10-K, 10-Q, 8-K, S-1, proxy statements); FDA databases (510(k) clearance letters, De Novo decisions, premarket approvals, IDE pivotal trial records, Warning Letters); company investor relations communications (audited financials, official funding-round press releases, IR-confirmed customer announcements); academic peer-reviewed publications (with named authors + journal venue + verifiable methodology); regulatory body primary documents (FAA Part 135 + Part 137, FCC filings, USPTO records, IRS Form 990 for nonprofits). These sources carry the primary-source anchor that resolves verified-tier claims.

Secondary established sources carry substantial verification weight. These include Reuters, Bloomberg, Wall Street Journal, Financial Times, Associated Press, and other established journalistic outlets with named-byline reporting, editorial standards, and substantive corrections practices. When primary sources aren't directly accessible, secondary established sources with verifiable primary-source attribution operate at substantial verification weight.

Secondary industry sources carry moderate verification weight. These include Robot Report, MedTech Dive, FierceBiotech, RoboticsTomorrow, IEEE Spectrum, and similar industry-trade publications with sector expertise. The moderate-weight framing matters: industry trade coverage adds expertise + context but operates at less editorial scrutiny than secondary established sources.

Aggregator content carries context-only weight, never primary. Wikipedia, SEO content sites, blog aggregators, and similar aggregator-generated content can provide topical context but never operate as primary-source verification anchors. When aggregator framing conflicts with primary sources, primary sources win.

Unverified content (self-report at unestablished outlets, social media claims, reseller marketing without confirmation, anonymous tips without verification) is flagged as such or excluded.

The discipline cuts uniformly. A Reuters report citing SEC filings operates at substantial verification weight; a Wikipedia entry citing the same SEC filings operates at context-only weight; the SEC filing itself operates at primary-source weight. Same underlying fact, different verification tiers based on source quality.

---

Cap-flag convention

When honest confidence limits exist, DEPLOY surfaces them transparently rather than hiding them.

A claim might be:

• Company-reported without independent verification: cap-flag explicit (the company says X; independent verification at primary-source depth pending).
• Regulatory clearance scope narrower than marketing implies: cap-flag explicit (FDA cleared for X scope; marketing language extends to Y scope; the clearance does not cover Y).
• Deployment scale ambiguous between built-capacity and daily-active: cap-flag both numbers honestly (system installed at N customers; daily-active utilization Z%).
• Contract value ceiling vs obligated: cap-flag the distinction (the contract has $X ceiling; obligated funding through current period is $Y).
• Timeline aspirational vs verified-pathway: cap-flag the verification posture (Musk-stated late-2020s framing; FDA pathway-to-clearance timing aspirational).

Cap-flag isn't hedging. It's the verification posture that makes the underlying assertion trustworthy at the depth where it operates. A verified claim with appropriate cap-flag is more useful than an over-stated claim that turns out to be wrong.

Example framing applied editorially: "Whoop BP Insights feature marketed by manufacturer as not-for-hypertension-diagnosis; FDA July 14 2025 Warning Letter disputes the framing; class action filed late 2025/early 2026; dispute unresolved as of May 2026 per re-verify sweep." The cap-flag surfaces manufacturer-position vs FDA-position vs class-action state without collapsing the three into one inaccurate framing.

---

Canonical worked examples

The framework operates at the discipline-in-action level. Seven worked examples from across the DEPLOY corpus demonstrate the framework operationally.

Within-cohort verified-vs-claimed pair: RingConn vs Happy Ring

Same form factor (smart ring), similar capability claims (sleep apnea detection), diametrically opposite verification posture.

RingConn markets AHI (apnea-hypopnea index) at 90.7% claimed accuracy; the AHI feature is NOT FDA-cleared; company PR explicitly states "actively pursuing clearance" and "not intended to diagnose." Happy Ring holds dual FDA 510(k) clearances: K240236 (October 8 2024 remote monitoring; not "screening" per primary-source correction) + K242224 (June 18 2025 home sleep test diagnosis support, ages 22+, clinician-directed; not autonomous diagnosis per primary-source correction).

The framework reads both honestly with their actual verification state. RingConn is the canonical worked example of AHI-claimed-not-cleared with explicit company-stated cap-flag; Happy Ring is the canonical worked example of dual-FDA-clearance verification depth. Together they demonstrate the verified-vs-claimed framework operating at form-factor-cell granularity (ring sub-cohort) on the same capability claim category (sleep apnea detection).

Within-cell AI-substance gradient: Dexcom Stelo vs Abbott Lingo

Same product category (glucose biometric for non-diabetics), same FDA clearance posture (both OTC-cleared), diametrically opposite AI-substance tier.

Dexcom Stelo operates generative AI substance: Vertex AI + Gemini Weekly Insights. Abbott Lingo operates adaptive-algorithmic substance: Lingo Count rule-based recommendations. Critical cap-flag: Lingo Live is human nutritionists, NOT AI coach. The aggregator-common confusion conflates Lingo Live human-nutritionist service with Lingo Count adaptive-algorithmic AI; DEPLOY surfaces both honestly.

The pair demonstrates the AI-substance gradient distinction within the glucose-cell at form-factor-cell granularity. Same product category + same FDA clearance posture + diametrically opposite AI-substance tier surfaces a verification-posture distinction that aggregator coverage flattens into uniform "AI-powered glucose monitoring."

Aggregator-drift correction: Figure 03 at BMW

Common aggregator framing: "Figure 03 deployed at BMW Plant Spartanburg + Leipzig."

Actual verified state: BMW Spartanburg deployment was Figure 02 completed pilot, NOT Figure 03. BMW Leipzig went to Hexagon AEON, NOT Figure. The aggregator framing collapses two distinct facts into one wrong claim about a third entity.

DEPLOY's audit-first discipline catches this structurally. The Figure 03 entity records verified Catalyst Brands Reno commercial customer; aggregator-drift framings on BMW deployment are rejected with corrected primary-source-anchored attribution.

Geography-error correction: Stryker

Common framing: "Stryker headquartered in Kalamazoo, Michigan."

Actual verified state: Stryker is headquartered in Portage, MI. SEC filings (10-K) confirm Portage. Small fact, but the discipline cuts uniformly. Verified outranks claimed at every granularity. The Stryker Mako entity anchor surfaces the Portage correction explicitly because the framework reads small facts the same way it reads large facts.

Valuation-inflation correction: Physical Intelligence

Aggregator framing: $11B valuation.

Verified state: $5.6B confirmed. The $10B and $38B figures often cited in aggregator coverage belong to Project Prometheus, a separate Jeff Bezos-backed lab commonly conflated with Physical Intelligence. DEPLOY surfaces verified $5.6B with explicit cap-flag against the conflation in the Physical Intelligence entity anchor.

The pattern recurs: large round valuations get cited from secondary aggregator coverage without checking whether the figure attaches to the named entity or to an adjacent entity with similar branding. Primary-source-anchored verification distinguishes the two.

Corporate-state correction: Covariant

Common aggregator framing: "Covariant was acquired by Intrinsic/Alphabet."

Actual verified state: Covariant continuing-but-diminished post-Amazon-AGI-departure. Founders Pieter Abbeel + Peter Chen + Rocky Duan + approximately 25% of staff moved to Amazon AGI team in August 2024. Ted Stinson took over as CEO. Legacy Covariant Brain has verified commercial deployment via KNAPP warehouse-automation channel. Non-exclusive license/IP retained.

Both true simultaneously: Covariant corporate-state continuing + Covariant capacity diminished but real. The framework surfaces both honestly in the Covariant RFM-1 entity anchor with three-fact-disambiguation (RFM-1 model + Covariant corporate-state + Amazon-acqui-hire as three distinct facts that aggregator coverage conflates).

Within-cohort autonomy-boundary distinction: Monogram mBos vs Mako vs da Vinci

Recent editorial work surfaced a four-way autonomy-boundary architectural taxonomy within the surgical cluster:

• Autonomous-execution: Monogram mBos. KUKA-based robotic arm executes the bone cut under AI control within a surgeon-approved CT plan and supervision. The ROBOT cuts, NOT the surgeon. Semi-autonomous version FDA-cleared March 17 2025; pre-commercial maturity.
• AI-augmented surgeon-controlled: Stryker Mako + Smith+Nephew CORI + Zimmer Biomet ROSA. Software plans + tracks + bounds. The SURGEON makes the cuts with AI augmentation; verified-commercial-shipped maturity.
• Replacement-robotics teleoperated: Intuitive da Vinci + Medtronic Hugo + J&J Ottava + CMR Versius. Surgeon teleoperates console-driven robotic arms; soft-tissue replacement-robotics spectrum.
• Assistive laparoscopy co-pilot: Moon Surgical Maestro. Surgeon directly manipulates laparoscopic instruments with co-pilot augmentation; assistive-niche commercial.

The verification posture differs fundamentally across all four. Trade-press coverage flattening orthopedic-robotic systems as "autonomous surgical robots" conflates AI-augmented Mako with autonomous-execution Monogram, which conflates further with replacement-robotics da Vinci. The four-way distinction is load-bearing architectural work; DEPLOY surfaces it structurally.

Critical cap-flag on Monogram framing: "semi-autonomous" is Zimmer Biomet + trade-press-sourced framing, NOT FDA-letter-verbatim. Monogram's own PR uses "robotic-assisted TKA." The cap-flag distinguishes the commercial-positioning descriptor (ZB/trade-sourced) from the FDA-letter language (primary-source-anchored).

---

Cohort architecture: form-factor-cell granularity

The framework operates at form-factor-cell granularity rather than collapsing to form-factor uniformity. Four canonical worked examples of sub-cohort treatment demonstrate the architecture compounding across cohorts:

• Biometric ring sub-cohort triangle: RingConn subscription-free + Ultrahuman hardware-subscription-augmented + Oura purchase+membership + Happy Ring FDA-cleared service-gated + Whoop subscription-only band exemplar. Same form factor + diametrically opposite verification postures.
• Biometric glucose-cell AI-substance gradient: Dexcom Stelo generative + Abbott Lingo adaptive-algorithmic + rule-based. Same product category + same FDA clearance posture + diametrically opposite AI-substance tier.
• Surgical orthopedic sub-cohort triangle: Stryker Mako large-footprint CT-based + Smith+Nephew CORI handheld imageless + Zimmer Biomet ROSA mid-size cross-domain. Form-factor + procedure-scope variance within orthopedic cohort.
• Humanoid consumer-vs-industrial sub-cohort umbrella: consumer-vs-industrial humanoid archetypes framework distinguishing verification postures + commercial-maturity timelines + brain-provider relationships + aggregator-drift patterns per archetype.

The sub-cohort treatment pattern recurs across cohorts. The framework reads at form-factor-cell granularity because that's where verification postures actually diverge.

---

Maturity-stage discipline

Deployment maturity reads through a structured discipline:

• Research: study, development, lab-only sales, single index cases, internal pilots. Lab-only sales are research, NOT commercial. Single index cases are research, NOT commercial. Maker-facility deployment (the maker deploying inside their own facility) is research, NOT commercial, per the registry rule that John Deere autonomous tractors in John Deere fields read differently from John Deere autonomous tractors in third-party-operator fields.
• Pilot: limited real-world trials at named customers with verified-pilot scope. The pilot tier is structurally distinct from research because real-world conditions + named customer counterparties are present, but pilot-not-scale framing reflects the limited deployment depth.
• Commercial: real economic work for paying customer at scaled-throughput maturity. Commercial-tier verification requires verified scaled-throughput evidence (Agility Digit at GXO Flowery Branch's 100,000-tote scaled-throughput is the canonical commercial-tier verified anchor in the humanoid cluster).
• Production: mass-produced + at-scale + multi-customer commercial deployment with verifiable production-rate cadence. Production-tier maturity is rare in the physical AI cohort outside established legacy primes.

The discipline cuts symmetrically across corporate-scale and emerging entrants. Pre-clearance is pre-clearance regardless of company scale. J&J Ottava (De Novo pre-market) and Vicarious Surgical V1 (SPAC-era distressed pre-market) sit at the same research tier despite distinct corporate-scale and capital histories. Both lack the verified gating event.

---

Multi-agent operational discipline

DEPLOY operates a multi-agent lane architecture with strict cross-lane discipline. Each lane has specific scope and reports through framework-laden response discipline.

The registry-ingest lane operates on canonical entity records with primary-source-anchored verification. The consumer-build lane operates on the consumer surface. The news-pub lane operates on editorial cluster framing + entity anchor depth. The acquisition lane operates on multimedia provenance + licensing posture. The rover surface operates on per-claim methodology display + transparency pages.

Cross-lane discipline operates through specific practices:

• Audit-first cohort-sync: verify state before assuming. Slug-audit before treating as net-new. Codebase-reality verification before treating session-memory framing as authoritative. Recent verification work surfaced 5 false-absents rejected + 3 genuine net-new confirmed (1X Redwood + Unitree H2 + Monogram) + 1 confirmed-doesn't-exist (Walker S3) across audit cycles.
• Per-anchor framework-laden discipline: V1.7 template (TLDR + structured stats + verification-posture-laden callouts + comparison-table + framework-laden FAQ + verification-anchored ArticleFooter) applied to every entity anchor.
• Em-dash grep + tsc clean before push: editorial cleanliness + type-safety gates on every ship. Em-dash grep catches a stylistic AI-generated tell; tsc catches structural errors before they reach production.
• Per-commit cadence: every entity anchor ships as a discrete commit + push; cluster intros update as separate commits. Cadence preserves the audit trail.
• Restraint discipline: the framework operates as "restraint IS the product." Sharper taxonomies + cleaner corrections + more honest cap-flag posture compound editorial credibility better than over-claimed coverage.

The multi-agent discipline is what makes the credibility surface auditable. Institutional partners who audit DEPLOY's framework discipline find the operational practice consistent with the public methodology statement.

---

Open transparency

Methodology surfaces operate at open transparency depth:

• /verified-vs-claimed: framework canonical reference. Tier definitions + source-quality rubric + cap-flag convention.
• /methodology/what-verified-means: foundational methodology entry; canonical reference for cluster intros + entity-anchor footers.
• /corrections journal: corrections shipped publicly with primary-source attribution; corrections operate as standard editorial practice, not exception.
• /funding + operational independence: operational independence statement + no commercial relationships disclosure + conflicts policy.
• /editorial-process: multi-agent lane discipline + audit-first cohort-sync + slug-audit verification practices documented operationally.

The transparency surface compounds with the framework discipline. Methodology canonical references + corrections journal + editorial process documentation + framework-laden editorial practice produce the verification authority that institutional partners (insurance underwriting + data sharing permissions + cross-device interoperability + standards body validation contexts) require.

---

Why this matters

The trust intermediaries the consumer physical AI economy needs require verification at this discipline level. Insurance underwriting requires verified-not-claimed deployment evidence + verified-not-marketed FDA-clearance scope + verified-not-aggregator-inflated commercial-volume claims. Data sharing permissions require verified-not-claimed privacy infrastructure + verified-not-marketed data-handling scope. Cross-device interoperability requires verified-not-claimed interoperability evidence + verified-not-marketed standards-compliance scope. Standards body validation requires the verification discipline applied at the per-claim depth that body certification practices already operate.

The framework discipline produces sharper taxonomies + cleaner corrections + more honest claim posture than aggregator-driven physical AI coverage. The four-way autonomy-boundary architectural taxonomy from the surgical cluster maps across other cohorts because the underlying distinction (who executes + at what verification posture + at what maturity-stage) operates the same way regardless of form factor. The biometric ring sub-cohort verified-vs-claimed pair (RingConn AHI-claimed-not-cleared vs Happy Ring dual-FDA-clearance) demonstrates the framework reading same product category + same capability claim at diametrically opposite verification posture, because that's how the underlying facts actually distribute.

This isn't methodology for its own sake. It's the verification protocol that makes DEPLOY a reference institutional partners can audit. The framework is the product.

For the canonical methodology entry, see what verified means. For the framework canonical reference, see verified vs claimed. For the editorial process documentation, see /editorial-process. For the corrections journal, see /corrections.