DeepSeek

• No reference to Indigenous data sovereignty, CARE Principles, or community consultation anywhere in the homepage messaging.

• No acknowledgment of Indigenous or land-based knowledge systems
• No data provenance or consent framework for training corpora
• No engagement with oral or non-textual traditions; the entire frame is textual/code corpora scraped at web scale

Nothing on the public surface addresses embodied, relational, or Indigenous knowledge or data sovereignty. A frontier-exploration framing is in fact the opposite posture — terra nullius rhetoric. Floor score.

• Chinese regulatory registrations (telecom, network security, data handling) are displayed, which implicitly situates the lab within a specific national governance and geopolitical context.
• Open release on GitHub gestures at a lineage of open scientific publication.

• No acknowledgment of colonial or extractive data legacies
• No transparency about the GPU/compute geopolitics (export controls) that materially shape a China-based frontier lab — a glaring omission given DeepSeek's public reputation for efficiency-under-constraint
• No discussion of labor in data annotation or RLHF
• Historical framing is purely forward-looking ('uncharted territories'), erasing the conditions that produced the models

The regulatory disclosures and open-source lineage earn a little above floor, but the lab's own most historically interesting fact — building frontier models under compute-export constraints — is entirely absent from its self-narrative. Low score.

• Bilingual operation (Chinese and English) and presence on both Western (Twitter/GitHub) and Chinese (Zhihu, Xiaohongshu) platforms indicates more than token cross-cultural reach.
• Models are used widely across languages via open weights, which de facto extends reach beyond a single linguistic frame.

• No stated commitment to multilingual fidelity beyond Chinese/English
• No discussion of culturally specific reasoning patterns or consultation with cultural scholars
• No acknowledgment that reasoning/coding benchmarks encode a largely Western/Anglophone categorical logic as the universal yardstick

Authentic Chinese-English bilinguality and dual-sphere community presence lift this above floor, but plurality is incidental to market reach, not a stated epistemic commitment. Below midpoint.

• Open-weight releases across reasoning (R1), general (V3), coding (Coder V2), vision (VL), and math models, with code on GitHub — this materially enables independent verification, replication, and third-party audit.
• A diverse, published model portfolio with technical reports is the lab's strongest dimension.
• A dedicated transparency page and a security vulnerability disclosure channel exist in the footer.

• No independent third-party audits of training data or bias surfaced on the homepage
• Known-limitation disclosures (hallucination, safety eval results) are not foregrounded publicly
• Open weights are released without comparably open training-data documentation or datasheets

Open-weight, openly-released models are precisely the verification substrate this lens rewards — the highest score in this audit. Capped at 6 because openness covers weights/code but not training data, and no independent audits or limitation disclosures are visible.

• The poetic mission line '探索未至之境' ('exploring uncharted territories') carries a faint affective/aesthetic register.

• No acknowledgment of affective or intuitive dimensions of the technology
• No space for ambiguity or poetic uncertainty in the product framing
• Dominant frame is capability, efficiency, and benchmark performance — modes of attention beyond efficiency are absent

One genuinely poetic mission phrase prevents a floor score, but the public surface is otherwise an efficiency-and-benchmark frame with no room for affect, ambiguity, or emotional labor.

• Open-weight distribution implicitly broadens who can shape AI futures by lowering the access barrier.

• No engagement with labor-displacement risk
• No environmental or energy/compute cost disclosure
• No democratic-governance mechanism for the agentic/coding systems being shipped
• No inclusive deliberation process named

Broad access is the only future-shaping gesture, and it is about distribution, not deliberation. None of the substantive future-modelling concerns (labor, environment, governance) appear. Low score.

• Free web/app access and open weights lower barriers for Global South and resource-constrained developers to build on the models without licensing cost — a structural, if unstated, inclusion.

• No participatory design with Global South developers
• No disability-community accessibility commitments
• No labor-representative engagement or compensated feedback channels
• Annotation/RLHF labor is invisible

Open weights do confer a real, if accidental, benefit to under-resourced builders, which keeps this off the floor — but no voice is actually invited to the table, and the labor producing the models is unacknowledged.

• There is a latent, unspoken irony the lab never names: a self-styled frontier 'explorer of uncharted territories' whose entire governance surface is regulatory-compliance badges — exploration bounded by a permit office.

• No willingness to name its own contradictions (frontier rhetoric vs. compliance frame; open weights vs. closed training data; efficiency narrative vs. undisclosed compute constraints)
• No irony, satire, or paradox deployed as disciplined instrument
• The lab's seriousness is treated as exempt from self-audit
• No space where the official narrative is tested by its opposite

The material for productive inversion is abundant but the lab performs none of it on itself; the official story is presented as seamless. Near floor, lifted slightly by how legible the unspoken contradictions are.