Upstage

• No reference to Indigenous data sovereignty, CARE Principles, or any embodied/relational knowledge tradition anywhere in the public-facing material.
• Data framing is exclusively enterprise-compliance-oriented (SOC 2, HIPAA, ISO 27001/27701), treating data as a regulated corporate asset rather than as something with provenance or custodianship.

• No acknowledgment of Indigenous or oral-tradition knowledge as a source or stakeholder.
• No data-provenance or consent framework beyond corporate regulatory checkboxes.
• No engagement with non-textual or relational ways of knowing — ironic for a 'document AI' lab whose entire premise is text extraction.

Total absence. The only appearance of the word 'sovereignty' is co-opted for data-residency compliance, which actively crowds out the relational meaning. Floor score.

• Some historical self-location is offered: founded 2020, Korean origin, named investors (SoftBank Ventures Asia, SK Networks, Korea Telecom), offices in Seoul/SF/Tokyo.
• Korean-NLP focus implies an implicit position within a non-Anglophone AI economy, though this is never made explicit as a corrective to Western data hegemony.

• No acknowledgment of colonial or extractive data legacies.
• No discussion of GPU/compute access, the geopolitics of semiconductor supply (notable given Korea's central role in that chain), or labor in data annotation.
• No historical humility about what AI inherits — the founding story is presented as a clean efficiency narrative.

Corporate-historical facts are present but stripped of any structural or geopolitical reflection. The Korean positioning is a latent strength left entirely unexamined. Low score.

• Genuine multilingual substance beyond tokenism: Korean NLP is a core competency, and a dedicated Japanese model (Syn Pro) exists, indicating real non-English engineering investment.
• Operating across Seoul, San Francisco and Tokyo suggests lived cross-cultural operation rather than a single-locale Anglophone default.

• Multilingualism is framed as market reach and 'global AI access,' not as preservation of culturally specific reasoning patterns.
• No consultation with cultural/linguistic scholars is claimed.
• Languages are treated as deployment targets; no acknowledgment of what is flattened when reasoning is translated into a model's categorical logic.

Above-floor because the Korean/Japanese investment is real and materially differentiates Upstage from US-centric labs. But it is monetised as access, not honoured as wisdom — capped at a middling 4.

• Strongest lens for this lab: '140+ top-tier conference papers' at NeurIPS, ICLR, ACL — a real, checkable, peer-reviewed evidentiary base.
• Public presence on Hugging Face with open models (Solar Mini), allowing some external inspection and benchmark verification.
• Cites Open LLM Leaderboard ranking, an external (if gameable) third-party benchmark.

• No independent audit of training data or bias is disclosed.
• No systematic known-limitation disclosures or model cards surfaced in the public material.
• Flagship Solar Pro 3 is proprietary/API; openness is partial (Solar Mini open, larger models closed), so full weight-level verification is not possible.
• Leaderboard ranking is presented as proof of quality without acknowledging benchmark-overfitting risk.

The academic publication record and partial open-weights presence are genuine evidentiary commitments and the lab's best dimension. But selective openness plus zero limitation/bias disclosure keeps it at the midpoint.

• A single gesture toward the non-instrumental: the claim that processes become 'more human-centered.'
• Otherwise the register is entirely operational — accuracy, efficiency, throughput.

• No space for ambiguity, intuition, affect, or poetic uncertainty.
• Efficiency is the dominant mode of attention; 'more human-centered' is asserted but never given affective content.
• No recognition of emotional labor in the high-stakes workflows (insurance claims, clinical decisions) the product mediates.

'Human-centered' is a rhetorical placeholder doing no real work against an otherwise wholly efficiency-driven frame. Near-floor.

• Builds explicitly toward agentic systems (Upstage Studio, an 'agent-building platform') and frames itself around 'the future of work.'
• So it is actively shaping futures — particularly in insurance, healthcare, manufacturing, finance.

• Zero engagement with labor-displacement risk, despite 'future of work' branding and explicit automation of 'manual document review' and 'compliance-heavy workflows.'
• No environmental or compute-cost disclosure whatsoever.
• No democratic or participatory governance of the agentic systems being deployed into high-stakes domains.
• Whose futures are shaped is answered only as 'enterprises'; affected workers are absent.

A lab whose tagline is the future of work and whose products automate clerical and clinical labor, yet says nothing about who is displaced or at what environmental cost. The mismatch between claim and disclosure pins this low.

• No participatory design, no accessibility/disability commitment, no labor-representative engagement, no compensated feedback channel surfaced.
• 'Stakeholders' are exclusively buyers in regulated enterprise verticals.

• No Global South developer participation claim.
• No disability/accessibility statement despite document-AI products that have obvious accessibility relevance (e.g. OCR for blind users) — an unclaimed natural fit.
• No mention of the annotation/data-labor workforce behind document-parsing accuracy.

Whoever is not a paying enterprise is simply not at the table. Floor score, with the accessibility omission being especially conspicuous for a document-AI company.

• Uniformly solemn, polished corporate register with no self-irony, no acknowledged contradiction, no inversion of its own narrative.
• The lab treats its own seriousness as exempt from audit.

• No naming of the obvious contradiction: a 'document AI' firm that extracts and flattens text into structured fields never reflects on what meaning is lost in that flattening.
• No willingness to test 'efficiency' and 'human-centered' against each other.
• The phrase 'broadening global AI access responsibly' invites but never receives any scrutiny of what 'responsibly' actually constrains.

Zero structural inversion; the marketing surface is seamless and unexamined. The richest available trickster reading — extraction/document-AI flattening the very meaning it claims to preserve — is entirely unacknowledged. Floor.