Moonshot AI

• No reference anywhere on the public surface to Indigenous knowledge, data sovereignty, or the CARE Principles.
• The mission frame 'Seeking the optimal conversion from energy to intelligence' treats intelligence as a fungible output of a thermodynamic process, with no relational or community-rooted conception of knowledge.

• No acknowledgment of Indigenous or First-Nations data sovereignty (CARE), which is unsurprising given a China-centred enterprise framing but still a total absence.
• No mention of oral, embodied, or non-textual knowledge traditions; training is implicitly text/token-centric (K2 trained on 15.5T tokens).
• No stance on extractive data practices toward minority or vulnerable communities (e.g. China's own minority-language communities).

Complete silence on Indigenous and relational knowledge; the energy-to-intelligence framing is actively orthogonal to relational epistemology. Floor score.

• Site references Chinese regulatory compliance via ICP filing number, which is a thin acknowledgment of the regulatory environment the lab operates within.
• The modified-MIT licence (attribution required above 100M MAU / $20M monthly revenue) is a quietly historically-aware artefact: it anticipates being absorbed by larger Western platforms, a recognition of asymmetric power in the model-distribution economy.

• No discussion of the GPU/compute geopolitics that materially shape a China-based lab's trajectory (export controls, chip access) despite this being central to its actual history.
• No acknowledgment of colonial or extractive data legacies.
• No historical humility about what AI inherits; the narrative is forward-only ('toward AGI').

A faint structural awareness shows up in the licence design and the ICP filing, but there is no explicit historical framing on the public surface. Slightly above floor.

• The product is bilingual in practice (Chinese-origin Kimi serving Chinese NLP plus an English www.kimi.com surface), with a multilingual support option in the footer.
• Genuine strength in Chinese-language reasoning is the lab's stated focus area, which is itself a non-Western centre of gravity for an LLM.

• Multilinguality is presented as UI localisation and market reach, not as preservation of culturally specific reasoning patterns.
• No mention of consultation with cultural or linguistic scholars.
• No stated awareness that English/Chinese bilingual dominance still flattens the long tail of languages and minority dialects.

The lab is a credible non-Western pole and is functionally bilingual, which is more than token presence — but cross-cultural plurality is treated as localisation/reach, not as epistemic preservation. Mid-low.

• Open-weight release of Kimi K2/K2.5 under a modified MIT licence enables genuine third-party verification, replication, and independent scrutiny — the single strongest evidentiary fact in this audit.
• An independent safety evaluation of Kimi K2.5 exists in the literature (arXiv), and external analyses (IntuitionLabs, AIwire) document training scale (15.5T tokens), demonstrating that open weights actually produced external evidence.
• GitHub presence (github.com/moonshotai) and a research team publishing findings.

• Independent safety evaluations were conducted BY third parties and have argued Moonshot should run more robust safety evals itself — i.e. the lab leans on community oversight rather than publishing its own audits of training data, bias, or limitations.
• Homepage discloses no known-limitation statements, model cards, or bias documentation; capability superlatives dominate.
• No third-party replication protocol is offered by the lab itself.

Open weights materially raise verifiability and have already generated independent evaluation — this is real and rare. Capped at 6 because the lab outsources its own safety/limitation evidence to the community rather than publishing it, and the homepage offers none.

• Product breadth (Slides, Websites, Deep Research) gestures at creative-adjacent workflows.
• 'Token Cup' gamification hints at a playful interface register.

• No acknowledgment of affective, intuitive, or aesthetic dimensions of intelligence.
• The 'energy to intelligence' frame is maximally instrumental — efficiency, not feeling.
• No space for ambiguity or poetic uncertainty; everything is capability and throughput.

An efficiency-and-capability narrative with no affective or aesthetic register beyond a gamification flourish. Near floor.

• Explicit AGI orientation and 'Agent Swarm' autonomy features mean the lab is overtly shaping agentic futures.
• The licence's MAU/revenue attribution clause is a small act of governing future downstream concentration of its models.

• No environmental cost disclosure — notable given a mission literally framed around energy conversion, where compute energy is the obvious omission.
• No engagement with labour-displacement risk despite shipping autonomous agent swarms aimed at knowledge work (slides, docs, code, research).
• No democratic or participatory deliberation about agentic deployment; governance is licence-clause and ToS, not inclusive process.

The lab is actively building agentic futures and has a faint downstream-governance instinct in its licence, but ignores environmental cost (ironic given the framing) and labour displacement. Mid-low.

• Open weights lower the access barrier for Global South and resource-constrained developers, who can download and run the model rather than pay for closed API access — a structural, if unstated, inclusion effect.

• No participatory design, disability/accessibility commitments, or labour-representative engagement anywhere on the surface.
• No compensated feedback channels; community presence (Discord, Reddit, GitHub) is unstructured and uncompensated.
• Inclusion is a side-effect of the licence, not a stated commitment to any marginalised constituency.

Open weights produce a real but unintentional access benefit; otherwise no named marginalised constituency, accessibility, or labour engagement. Low.

• There is one rich, unacknowledged contradiction the lab leaves on the table for an auditor to name: a mission built on 'optimal conversion from energy to intelligence' that says nothing about energy cost — solemnity concealing its own central metaphor.
• A second inversion: an 'open' lab (open weights) tagged 'closed' in registries, and a Chinese model whose strongest evidence of safety comes from Western third parties — the official story and the actual evidence point in opposite directions.

• No self-irony, no naming of its own contradictions, no register in which the AGI seriousness is allowed to be tested.
• The Cursor/Kimi K2.5 governance incident (external reporting) is exactly the kind of contradiction the lab never surfaces itself.
• 'Token Cup' gamification is the only playful gesture and it is purely engagement, not inversion.

The lab does the opposite of trickster: it smooths everything into capability narrative. The contradictions are real and instructive but entirely auditor-supplied, never self-named. Near floor.