Vocabulary used across the Tractatus framework, the sovereign-record papers, and the agenticgovernance.digital blog — with cross-references between concepts and pointers into the texts that develop each one. The intended way in for readers approaching this work for the first time. English and te reo Māori.
231 entries ·
v1.0 ·
Last built against sources: 2026-05-06 ·
CC BY 4.0
Audience: Written for readers without a technical background — parish treasurers, rūnanga secretaries, family-history curators, and anyone reading the research papers or blog posts who would like a quick definition without leaving the page. Each entry covers a term a non-specialist might not recognise from context. There is no assumed level of prior knowledge; the explanations are for everyone.
Translation: English and te reo Māori are first-class. Standard technical terms are translated by DeepL. Te ao Māori vocabulary is presented in te reo with the English explanation. Entries marked [DeepL] are machine translations awaiting human review.
Suggest a term
Tell us a term you'd like explained. Anonymous. We'll review and add it if it fits the glossary's scope.
231 of 231 entries shown
Architecture & Sovereignty PrimitivesHanganga me ngā Mātāpono Mana Motuhake(21)
BoundaryEnforcer — Kaitiaki Rohe
A Tractatus framework service that blocks the AI from making certain classes of decision on its own — those that involve values, ethics, or cultural context — and instead routes them to human deliberation.
[DeepL] He ratonga anga Tractatus e aukati ana i te AI kia kaua e whakatau i ētahi momo whakataunga māna anō — arā, ko ēnei e pā ana ki ngā uara, ki ngā tikanga, ki te horopaki ahurea rānei — ā, ka tukuna ēnei ki te whiriwhiringa a te tangata.
Federation that is bilateral and strictly bounded — each tenant chooses its federation partners and what specific records may travel in each direction, rather than joining an open network where any participant can address any other.
[DeepL] He federation ā-rua, ā, he tino herea — ka kōwhiri ia kaiwhakamahi i āna hoa federation me ngā rekoata motuhake ka taea te tuku i ia ahunga, kaua e hono ki tētahi whatunga tuwhera e taea ai e ngā kaiuru katoa te whakapā atu ki tētahi atu.
Cognition sovereignty — rangatiratanga o te māramatanga
The principle that a community using an AI is not just choosing a tool but importing a cognitive substrate — its training corpus, its assumptions, its registers. Cognition sovereignty asks who shaped that substrate and on what authority. Paper B's cognition layer is the architectural answer for the project's communities.
[DeepL] Ko te mātāpono, kāore te hapori e whakamahi ana i te AI e kōwhiri noa ana i tētahi taputapu, engari kei te kawemai i tētahi paparanga mātauranga — arā, tōna kohinga whakangungu, ōna whakapae, ōna reo. Ka pātai te rangatiratanga mātauranga, nō wai i hanga taua paparanga, ā, i runga i te mana āhea. Ko te paparanga mātauranga o Pepa B te whakautu hanganga mō ngā hapori o te kaupapa.
A failure pattern in which two systems intended to check each other share enough underlying assumptions that they fail in the same way at the same time. Guardian Agents use mathematical similarity (not another generative model) precisely to avoid this — the watcher must operate in a different epistemic domain from the watched.
[DeepL] He tauira hapa i roto i tēnei ka hapa ngātahi ngā pūnaha e rua e whāia ana kia tirotiro tētahi i tētahi, nā te nui o ngā whakaaro taketake e tiritiri ana, ā, ka hapa rātou i te ara kotahi i te wā kotahi. Ka whakamahi ngā Agent Guardian i te ōritetanga pāngarau (ehara i tētahi atu tauira whakaputa) kia karo ai i tēnei — me mahi te kaitirotiro i roto i tētahi rohe mātauranga rerekē i te mea e tirotirohia ana.
The use of cryptographic signatures and proof chains so that every record carries verifiable evidence of who created it, when, and that it has not been altered since. Underpins the claim that sovereignty is a property of the record rather than a permission the operator could revoke.
[DeepL] Ko te whakamahi i ngā haina whakamunatanga me ngā mekameka taunakitanga, kia mau ai ki ia rēhita he taunakitanga ka taea te whakamana mō te tangata i hanga, te wā i hanga ai, me te kore i whakarerekēhia mai i taua wā. E tautoko ana i te kī e mea ana he āhuatanga o te rēhita te rangatiratanga, ehara i te whakaaetanga ka taea e te kaiwhakahaere te whakakore.
The principle that the community whose life produces the data also governs what is done with it — including where it lives, who may read it, who may copy it, and what an AI may learn from it. Distinct from individual privacy: a collective right held by the community.
[DeepL] Ko te mātāpono, ko te hapori i whakaputa i ngā raraunga e whakahaere ana i ngā mahi e mahia ana ki aua raraunga — tae atu ki te wāhi e noho ai, ko wai ka taea te pānui, ko wai ka taea te kape, me ngā mea ka ako ai tētahi AI. He rerekē i te tūmataitinga takitahi: he tika ā-hapori e puritia ana e te hapori.
The bounded message format used when one tenant communicates with another across a federation link. The envelope carries only what the sending tenant has explicitly chosen to share, with provenance attached, so that federation does not become a back-door to cross-tenant content access.
[DeepL] Ko te hōputu karere herea e whakamahia ana ina whakawhiti kōrero tētahi kaipā ki tētahi atu mā tētahi hononga federation. Ka kawea e te kopaki anake ngā mea kua tohua motuhake e te kaipā tuku kia tohatoha, ā, kua tāpirihia ngā pārongo takenga, kia kore ai e noho te federation hei kuaha huna mō te uru ki ngā ihirangi a ētahi atu kaipā.
The structural shape of a Tractatus deployment: codebase + framework services + Guardian Agents + community-owned records + community-aligned AI. Discussed in the Five-Component Tractatus Architecture blog post.
[DeepL] Te āhua hanganga o tētahi whakaurunga Tractatus: pūtake waehere + ratonga anga + Kaihautū Guardian + ngā rekoata e puritia ana e te hapori + AI e hāngai ana ki te hapori. I matapakihia i te pou rangitaki 'Five-Component Tractatus Architecture'.
A four-phase verification layer that checks AI responses against the community's own corpus and policies, using mathematical similarity rather than another generative model — the watcher operating in a different epistemic domain from the system it watches, to avoid common-mode failure.
[DeepL] He paparanga whakamana e whā ngā wāhanga e tirotiro ana i ngā whakautu AI ki te kohinga me ngā kaupapa here a te hapori, mā te whakamahi i te ōritetanga pāngarau, kaua mā tētahi atu tauira whakaputa — ko te kaitirotiro e mahi ana i tētahi rohe mātauranga rerekē i te pūnaha e tirohia ana, kia kore ai e puta he hapa aronga kotahi.
Limits of the sayable — Ngā here o ngā mea e taea te kī
Wittgenstein's observation that some matters can be specified, measured, and verified, while others — involving values, ethics, cultural context — fundamentally cannot. What cannot be systematised must not be automated; the framework's foundational constraint.
[DeepL] Ko te kitenga a Wittgenstein, ka taea te tautuhi, te ine, me te whakamana i ētahi take, engari kāore e taea tērā mō ētahi atu — arā, ngā uara, te matatika, me te horopaki ahurea. Me kaua e aunoa ngā mea kāore e taea te whakarite pūnaha; koinei te here taketake o te anga.
Member-driven sovereign portability — Kawe rangatiratanga e ārahi ana e ngā mema
The right and the technical means for any member to export their own contributions in a portable form, without the operator's permission and without losing the cryptographic provenance that makes those records verifiable elsewhere.
[DeepL] Te tika me ngā tikanga hangarau kia taea e ia mema te kaweake i āna ake koha i roto i tētahi āhua kawe, me te kore whakaaetanga a te kaiwhakahaere, me te kore ngaro o te takenga whakamunatanga e whakamana ana i aua rekoata ki ētahi atu wāhi.
Pluralistic Deliberation Orchestrator — Kaiwhakahaere Whiriwhiringa Maha
A Tractatus framework service that, when the AI hits a values conflict, halts and coordinates deliberation among the affected stakeholders rather than choosing on its own. Implements Berlin's value pluralism as a runtime architectural primitive.
[DeepL] He ratonga anga Tractatus e, ina pāngia te AI e tētahi taupatupatu uara, ka whakamutu, ka whakarite i te whakawhitiwhiti whakaaro i waenga i ngā hunga whai pānga, kaua e whiriwhiri māna anake. Ka whakatinana i te maha-uara o Berlin hei pūtake hanganga whakahaere-ā-wā.
Read further:Tractatus working paper; Blog: Guardian Agents
Was this useful?Thanks — recorded.
Proof chain — Meaka whakamana
A sequence of cryptographic signatures that links a record to its origin, its prior states, and the policy decisions that governed it. The platform periodically re-verifies older proof chains in batches so that tampering anywhere in the chain becomes detectable.
[DeepL] He raupapa hainatanga whakamuna e hono ana i tētahi rēkoata ki tōna takenga, ki ōna āhua o mua, me ngā whakataunga kaupapa here i whakahaere i a ia. Ka whakamātautau anō te papanga i ngā mekameka taunakitanga tawhito ia wā i roto i ngā kohinga, kia kitea ai mēnā ka whakarerekēhia tētahi wāhanga o te mekameka.
A platform design in which the sovereignty a community needs is a property of the records themselves rather than a concession the operator may revoke. Cryptographic provenance, tenant-bounded policy enforcement, federation envelopes, and member-driven export each contribute to making the records irrevocably the community's.
[DeepL] He hoahoa tūāpapa e whakarite ana kia noho rangatiratanga o te hapori hei āhuatanga o ngā rēkōta anō, kaua hei mana whakaae ka taea e te kaiwhakahaere te whakakore. Ko te takenga whakamunatanga, te whakatinanatanga kaupapa here e herea ana ki ngā kaipā, ngā anga kotahitanga, me te kaweake e ārahi ana ngā mema, ka āwhina katoa kia kore rawa e taea te whakakore i ngā rēkōta, ā, kia noho rātou hei rawa o te hapori.
One of the foundational architectural building blocks that, composed together, make sovereignty a property of the records rather than a permission the operator may revoke. Paper A names cryptographic provenance, tenant isolation, the federation envelope, member-driven sovereign portability, the tenant key store, the proof chain, and the policy-inheritance engine as primitives. "Primitive" in the computer-science sense: irreducible building blocks that compose to produce the architectural property.
[DeepL] Ko tētahi o ngā pūtake hanganga matua e huihui ana hei whakarite kia riro te rangatiratanga hei āhuatanga o ngā rekoata, kaua hei whakaaetanga ka taea e te kaiwhakahaere te whakakore. Ko te Pepa A e kī ana ko te takenga whakamunatanga, te wehewehe kiritaki, te anga kotahitanga, te kawe rangatiratanga e ārahi ana e ngā mema, te whare kī kiritaki, te mekameka taunakitanga, me te pūnaha tuku tuku kaupapa here hei pūtake. Ko te kupu "pūtake" i te horopaki pūtaiao rorohiko: he pūtake hanga kāore e taea te whakaiti, ā, mā te whakakotahi ka puta he whakamana.
Read further:Paper A §5; glossary first-category name
Was this useful?Thanks — recorded.
Surface — Whārangi
A generic term in this project's vocabulary meaning the bounded set of points at which a particular kind of interaction can take place. Always paired with a qualifier — attack surface, API surface, evaluation surface, review surface, write surface, audit surface, request surface, governance surface — that names what kind of interaction. Asking "what is the surface for X?" is the project's default way of bounding a concern: name the surface, then design defences or affordances for it. The framing pairs naturally with sovereignty primitives (which compose to bound a surface) and with threat models (which enumerate what an adversary could do at a given surface).
[DeepL] He kupu whānui i roto i te rārangi kupu o tēnei kaupapa e tohu ana i te huinga herea o ngā tohu e taea ai te tūhonohono motuhake. Ka hono tonu ki tētahi kupu whakamārama — mata whakaeke, mata API, mata aromatawai, mata arotake, mata tuhi, mata arotake pūtea, mata tono, mata whakahaere — e tohu ana i te momo tūhonohono. Ko te pātai "He aha te mata mō X?" te huarahi taunoa a te kaupapa hei here i tētahi āwangawanga: ingoaia te mata, kātahi ka hoahoa ārai, ka whakarato āheinga rānei mōna. Ka hono tēnei anga ki ngā pūtake rangatiratanga (e huihui ana hei here i tētahi mata) me ngā tauira mōrearea (e whakarārangi ana i ngā mahi ka taea e tētahi hoariri i tētahi mata).
The architectural rule, enforced at the data layer, that one tenant's content is not exposed to or modifiable by another tenant — including by a platform administrator. Tenant isolation is treated as foundational; if code allows a platform admin to read tenant content, that is a security defect, not a feature.
[DeepL] Ko te whakarite hanganga e whakarite ana kia kore rawa e kitea, e whakarerekē rānei e tētahi atu kaiwhakamahi te ihirangi o tētahi kaiwhakamahi — tae atu ki te kaiwhakahaere papanga. Ka whakaarohia te wehewehe kaiwhakamahi hei tūāpapa; mēnā ka āhei te waehere i te kaiwhakahaere papanga ki te pānui i te ihirangi o tētahi kaiwhakamahi, he hapa haumaru tērā, ehara i te āhuatanga.
A query, operation, or service that always filters by the requesting tenant's identifier. The standing rule is: every database query filters by tenantId; never bypass.
[DeepL] He pātai, he whakahaere, he ratonga rānei e tātari tonu ana mā te tuakiri o te kaipātai e tono ana. Ko te ture tū tonu: ia pātai pātengi raraunga ka tātari mā te tenantId; kaua rawa e karo.
The database identifier that binds every record to a single tenant. Stored on every domain model. Indexed for fast filtering. The mechanism by which tenant isolation is enforced at the data layer.
[DeepL] Ko te tohu raraunga e here ana i ia rekoata ki tētahi kaipā kotahi. Kei ia tauira rohe. Kua tāurutia kia tere ai te tātari. Ko te pūnaha e whakatinana ana i te wehewehe kaipā i te papa raraunga.
A development-time governance framework for AI assistance — six services that constrain what the AI building the platform may decide on its own (BoundaryEnforcer, ContextPressureMonitor, CrossReferenceValidator, InstructionPersistenceClassifier, MetacognitiveVerifier, PluralisticDeliberationOrchestrator) — together with Guardian Agents that verify outputs after the fact. Open source under Apache 2.0.
[DeepL] He anga whakahaere i te wā whakawhanake mō te āwhina AI — e ono ngā ratonga e here ana i ngā whakatau a te AI e hanga ana i te papanga (BoundaryEnforcer, ContextPressureMonitor, CrossReferenceValidator, InstructionPersistenceClassifier, MetacognitiveVerifier, PluralisticDeliberationOrchestrator) — me ngā Agent Kaitiaki e whakamana ana i ngā hua i muri i te mahi. Puna tuwhera i raro i te raihana Apache 2.0.
Isaiah Berlin's argument that some values are genuinely incommensurable — privacy and safety cannot be ranked on a single scale without imposing one community's priorities on everyone else. The architecture must accommodate plural moral frameworks rather than flatten them.
[DeepL] Ko te whakapae a Isaiah Berlin, arā, he tino kore e taea te whakataurite i ētahi uara — kāore e taea te whakarārangi i te tūmataitinga me te haumaru i runga i tētahi ine kotahi, me te kore e whakanoho i ngā manakohanga o tētahi hapori ki runga i ētahi atu katoa. Me whakaae te hoahoanga ki ngā anga matatika maha, kaua e whakaiti i a rātou.
He Tangata, He Karetao, He Ātārangi — He Tangata, He Karetao, He Ātārangi
"A person, a puppet, a shadow" — the title of one of Dr Taiuru's framings, distinguishing what an AI is not (he tangata, a person) from what it might be (he karetao, a puppet operated by another, or he ātārangi, a shadow that follows the form of a person without the substance).
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
A gathering or meeting, often on a marae and conducted under tikanga — much more than a Western "meeting". A hui carries whakatau (welcome), karakia, formal speeches, shared kai, and the relational obligations that frame whatever business is on the agenda.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
Read further:Village glossary (platform); Paper A §10
Was this useful?Thanks — recorded.
Iwi — Iwi
A tribe — the largest contemporary Māori political and kinship grouping, descended from a common ancestor. Comprises multiple hapū, which in turn comprise multiple whānau.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
Guardianship, stewardship — exercised in obligation to ancestors, descendants, and the environment. Distinct from property ownership: a kaitiaki holds something on behalf of others rather than owning it outright. The platform's data-sovereignty framing is closer to kaitiakitanga than to property rights.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
A ritual incantation, prayer, or blessing — spoken or chanted to open gatherings, consecrate food, mark transitions, or farewell the dead. Karakia frame what follows; they are not a religious appendix but part of how the action proceeds.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
A purpose, principle, or guiding philosophy that shapes how something is done. Dr Taiuru's Kaupapa Māori AI Framework names a kaupapa for AI grounded in te ao Māori.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
The specific protocols of a particular marae — how visitors are welcomed (pōwhiri), where people sit, the order of speaking, the karanga (call), the songs, the customs around shared food. Each marae has its own kawa; respecting it is part of being a guest.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
A gift offered in reciprocal relationship — not a payment for service but a way of strengthening the bond between giver and receiver. The platform's funding model uses koha rather than subscription pricing.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
Speech, talk, discussion, conversation — and also narrative, account, story, teaching. Kōrero holds the spoken word, the genealogical account, and the negotiated meaning together as one continuous practice rather than three separate things.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
Authority, prestige, dignity, intrinsic worth. Held by people, places, and taonga; can be enhanced or diminished by how others act in relation to it. The platform's privacy posture rests on the recognition that members carry mana that the system must not erode.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
Reciprocal care, hospitality, generosity in relationship. The koha system through which the platform sustains its work draws on manaakitanga as much as on whanaungatanga.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
The communal meeting ground at the heart of a Māori community — both a physical place (with its wharenui, wharekai, urupā) and a social institution where tikanga is practised.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
Mātauranga is knowledge — accumulated learning. Mātauranga Māori is the body of Māori knowledge specifically: intergenerationally-transmitted understanding that includes but is not limited to language, environment, ancestry, ritual, and craft. A taonga that the platform's data-sovereignty design is structured to protect rather than appropriate.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
Life force, vital essence. Held by living beings, places, and taonga; can be diminished by mistreatment. One of the four dimensions of Mead's Tikanga Test.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
Hirini Moko Mead's evaluative surface for measuring whether a practice is tika — through tapu (sacredness), mauri (life force), take-utu-ea (the obligation cycle of issue, response, and resolution), and whanaungatanga (relationship). Paper B's full empirical companion will engage this test directly.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
Self-determination, chiefly authority. Article 2 of Te Tiriti affirms tino rangatiratanga (full self-determination) over taonga. In the data context: the right of a community to govern its own data and the systems that act upon it.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
A council, assembly, or governance body — often the modern administrative form through which an iwi or hapū exercises its rangatiratanga. Several rūnanga deploy or trial the platform.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
The cycle of issue (take), response (utu), and resolution (ea). A te ao Māori frame for thinking about how harms are addressed and balance restored. One of the four dimensions of Mead's Tikanga Test.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
The extended Māori funeral rite — three or more days of mourning, formal speeches, song, and shared food on the marae, culminating in burial. Tangihanga is one of the deepest expressions of whanaungatanga and one of the most consequential commitments a person can make to their wider whānau.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
Treasure — anything valued, including tangible objects, places, language, knowledge, and relationships. Te Tiriti Article 2 protects taonga; current reading extends this to community data and the AI systems that mediate community life.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
Sacredness — a state requiring respect and care; certain matters are tapu and must be approached with the appropriate protocol. One of the four dimensions of Mead's Tikanga Test.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
The Māori world — the worldview, values, knowledge, and ways of being that have shaped Māori life. The platform's te ao Māori engagement is grounded in the recognition that this worldview is a body of substantive content, not an aesthetic overlay on a Western system.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
The Māori language. A taonga in its own right; one of New Zealand's three official languages. The platform's translation workflow uses DeepL for te reo (DeepL supports MI as a target language) for initial drafts; cultural-loaded terms are flagged for human review.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
Customary Māori practice — the right way of doing things in a given context, encompassing protocol, ethics, and relational obligations. Tikanga is context-specific; what is correct depends on the people, place, and purpose.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
Reciprocity — the response that completes a relationship. Often mistranslated as "revenge"; in tikanga it carries the broader sense of restoring balance, including through gift, gratitude, or repair.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
Genealogy — the network of ancestral and descendant relationships that places a person within whānau, hapū, iwi, and waka. Whakapapa is not just a family tree; it situates a person in the world by naming their connections, and it carries obligations as well as identity.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
Extended family — broader than the nuclear family, encompassing multiple generations and a wider circle of kinship. The whānau Tier-1 cohort is trained on contexts particular to this scale of community.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
Relationship-building, kinship, the cultivation and maintenance of connection. One of the four dimensions of Mead's Tikanga Test. The platform's koha framing draws on whanaungatanga.
He whakamārama ā-tangata e tatari ana — Te kupu i te reo Māori, te whakamārama i te reo Pākehā mō nāianei.
Governance, Law & PolicyWhakahaere, Ture, me te Kaupapa Here(27)
AI Act (EU) — Ture AI (EU)
The European Union's regulatory framework for AI systems, finalised 2024. The platform's governance services align with several Act provisions, including Article 22 on automated decision-making.
[DeepL] Te anga whakahaere a te Uniana o Europi mō ngā pūnaha atamai, i whakaoti i te tau 2024. E hāngai ana ngā ratonga whakahaere o te tūāpapa ki ētahi tikanga o te Ture, tae atu ki te Tuhinga 22 mō te whakatau aunoa.
A permissive open-source licence allowing use, modification, and redistribution with limited conditions (notice preservation, patent grant, no trademark licence). The Tractatus framework is licensed under Apache 2.0.
[DeepL] He raihana puna tuwhera whakaae e whakaae ana ki te whakamahi, ki te whakarerekē, me te tohatoha anō me ngā here iti (me tiaki te pānui, me tuku raihana pātengi, kāore he raihana tohu hokohoko). Kei raro i te raihana Apache 2.0 te anga Tractatus.
CARE Principles for Indigenous Data Governance — Ngā Mātāpono o CARE mō te Whakahaere Raraunga Taketake
Collective benefit, Authority to control, Responsibility, Ethics. An international framework for indigenous data governance that complements the FAIR data principles by asserting collective and authority dimensions FAIR omits.
[DeepL] Painga ā-rōpū, Mana whakahaere, Haepapa, Tikanga. He anga ā-ao mō te whakahaere raraunga taketake e whakakī ana i ngā mātāpono raraunga FAIR mā te whakapuaki i ngā āhuatanga ā-rōpū me te mana whakahaere kua mahue e FAIR.
CC BY 4.0 (Creative Commons Attribution 4.0) — Tohu Whakaaturanga 4.0 a Creative Commons
A permissive content licence allowing redistribution and adaptation provided the original author is credited. Used for Paper A, Paper B, and the project's published essays.
[DeepL] He raihana ihirangi whakaae e whakaae ana kia tohatoha anō, kia whakarerekē rānei, me te whakamōhio ki te kaituhi taketake. Ka whakamahia mō te Pepa A, te Pepa B, me ngā tuhinga kua whakaputaina o te kaupapa.
Shared governance between two or more parties with distinct sources of authority — most often, in the New Zealand context, between iwi or hapū and the Crown (or a local council). Distinct from token consultation: each party retains its own authority and the structure is designed to keep that authority intact.
[DeepL] Ko te whakahaere ngātahi i waenga i ngā rōpū e rua, e maha rānei, me ngā puna mana motuhake — i te nuinga o te wā, i te horopaki o Aotearoa, i waenga i ngā iwi, ngā hapū rānei me te Karauna (te kāwanatanga rānei, te kaunihera ā-rohe rānei). He rerekē ki te uiui tohu noa: ka pupuri tonu ia rōpū i tōna ake mana, ā, kua hoahoatia te hanganga kia mau tonu ai taua mana.
A structured assessment, required under GDPR for high-risk processing, of how a processing activity affects data subjects and how risks are mitigated. The platform produces DPIAs as living documents alongside major architectural changes.
[DeepL] He aromatawai hanganga, e hiahiatia ana i raro i te GDPR mō ngā tukanga tūraru-nui, hei aromātai i te pānga o tētahi mahi tukatuka ki ngā tangata raraunga, me te āhua e whakaitihia ai ngā tūraru. Ka whakaputa te papanga i ngā DPIAs hei tuhinga ora i te wā e puta ana ngā panonitanga hanganga matua.
Data subject request (DSR) — Tono a te tangata e pā ana ki ngā raraunga (DSR)
A formal request under GDPR or similar legislation in which a person asks the controller to access, correct, port, or delete their data. The platform processes DSRs through a Bull-backed queue with automatic retry and audit logging.
[DeepL] He tono ā-ture i raro i te GDPR, i ētahi atu ture ōrite rānei, e tono ana te tangata ki te kaiwhakahaere kia uru, kia whakatika, kia whakawhiti, kia muku rānei i āna raraunga. Ka tukatuka te papanga i ngā tono DSR mā tētahi rārangi tatari tautokohia e Bull, me te whakamātau anō aunoa me te rēhita arotake.
Māori AI governance researcher whose Kaupapa Māori AI Framework, and more recent inquiry into AI personhood constituted by Māori knowledge, frame much of the project's te ao Māori engagement. Reviewed the Tiriti compliance statement v0.1 in 2026. His own work and writing are at taiuru.co.nz.
[DeepL] He kairangahau whakahaere AI Māori nāna i waihanga te Kaupapa Māori AI Framework, ā, ko tana rangahau hou mō te tangata AI i hangaia e te mātauranga Māori e whakarite ana i te nuinga o ngā mahi whakawhanaungatanga ki te ao Māori o te kaupapa. I a ia i arotake i te tau 2026 i te tauākī hāngai Tiriti v0.1.
Updated 2026-05-06 in response to reader feedback — added link to Dr Taiuru's own site
Was this useful?Thanks — recorded.
Elinor Ostrom — Elinor Ostrom
Nobel-winning political economist whose work on governing the commons frames the platform's view of community-scale digital infrastructure as a commons rather than a market or a service.
[DeepL] He ōhanga-tōrangapū i toa i te Tohu Nobel, ko āna mahi mō te whakahaere i ngā rawa tūmatanui e whakarite ana i te tirohanga o te tūāpapa ki ngā hanganga matihiko ā-hapori hei rawa tūmatanui, ehara i te mākete, i te ratonga rānei.
An information-sharing arrangement involving the United States and partner states. Cited in the project's threat model as one of the jurisdictional pressures that motivate the platform's sovereign infrastructure posture.
[DeepL] He whakaritenga tiritiri pārongo e whai wāhi ana te United States me ōna hoa rangatira. I whakahuatia i roto i te tauira tūraru o te kaupapa hei tētahi o ngā pēhanga ā-ture e akiaki ana i te tūnga hanganga rangatiratanga o te papa.
EUPL-1.2 (European Union Public Licence) — EUPL-1.2 (Rāhina Tūmatanui o te Uniana o Europi)
A copyleft software licence approved by the European Commission, compatible with several other open-source licences. The community platform code is licensed under EUPL-1.2.
[DeepL] He raihana pūmanawa copyleft i whakaaetia e te Komihana o Europi, e hāngai ana ki ētahi atu raihana puna tuwhera maha. Kei raro i te raihana EUPL-1.2 te waehere tūāpapa hapori.
GDPR — Ture Whakahaere Whaiaro ā-Iwi o te Uniana o Europi
The General Data Protection Regulation — the European Union's data protection law. Imposes obligations on any controller of EU residents' personal data; shapes the platform's consent model, data-subject-request handling, and data-export architecture.
[DeepL] Te Ture Whānui mō te Tiaki Raraunga — te ture tiaki raraunga a te Uniana o Europi. Ka whakanoho i ngā herenga ki ia kaiwhakahaere o ngā raraunga whaiaro o ngā kainoho o te Uniana o Europi; ka ārahi i te tauira whakaaetanga o te papaanga, te whakahaere tono a te tangata e pā ana ki āna raraunga, me te hanganga kaweake raraunga.
GDPR's restriction on solely-automated decision-making. Requires that significant decisions about a person not be made by an algorithm without human meaningful intervention. Per-request inference routing is logged so the platform can demonstrate Article-22 compliance.
[DeepL] Te here a te GDPR mō ngā whakatau āunoa anake. E hiahiatia ana kia kaua e mahia e tētahi tātaitanga ngā whakatau nui mō tētahi tangata, mehemea kāore he urupare whai tikanga a te tangata. Ka tuhia ngā huarahi whakapae i ia tono, kia taea e te papanga te whakaatu i tōna ū ki te Tuhinga 22.
Dr Karaitiana Taiuru's prescriptive framework of Te Tiriti-grounded principles for AI consent, Māori data sovereignty, and full-chain accountability. The platform's governance design discharges the duties this framework asserts. Dr Taiuru's own writing is at taiuru.co.nz.
[DeepL] Ko te anga whakaritenga a Dr Karaitiana Taiuru o ngā mātāpono i runga i Te Tiriti mō te whakaaetanga AI, te rangatiratanga raraunga Māori, me te haepapa puta noa i te mekameka katoa. Ka whakatutuki e te hoahoa whakahaere o te tūāpapa ngā haepapa e whakapuaki ana tēnei anga.
Updated 2026-05-06 in response to reader feedback — added link to Dr Taiuru's own site
Was this useful?Thanks — recorded.
Minutes — Meneti
The official record of what happened at a meeting: who attended, what was discussed, what motions were moved, how members voted, and what was resolved. Approved at the following meeting; signed by the chair; kept indefinitely as part of the governance record.
[DeepL] Ko te pūrongo ā-ture mō ngā mea i tupu i te hui: ko wai i tae mai, he aha ngā kaupapa i matapaki, he aha ngā tono i whakatakotoria, me pēhea ngā pōti a ngā mema, me ngā whakataunga i puta. I whakamanahia i te hui i muri mai; i hainatia e te heamana; ka puritia mō ake tonu atu hei wāhanga o ngā pūrongo whakahaere.
A formal proposal put before a board or committee for decision. Convention requires a motion to be moved by one member and seconded by another before discussion; once carried by vote, it becomes a resolution and is recorded in the minutes.
[DeepL] He tono ā-ture i tukuna ki te poari, ki te komiti rānei mō te whakatau. E ai ki ngā tikanga, me whakatakotohia te motini e tētahi mema, ā, me tautokona e tētahi atu i mua i te matapaki; ka whakamanahia mā te pōti, ka noho hei whakataunga, ā, ka tuhia ki ngā meneti.
Polycentric governance — Kaitiakitanga ā-pokapū maha
Elinor Ostrom's framing in which multiple overlapping authorities co-govern a commons rather than a single central body. Cited in the platform's design as a structural alternative to single-vendor or single-state AI governance.
[DeepL] Ko te anga a Elinor Ostrom e whakaatu ana i ngā mana maha e whakawhānui ana, e whakahaere tahi ana i tētahi rawa tūmatanui, kaua ko tētahi mana matua kotahi. I whakahuatia i roto i te hoahoa o te papa hei kōwhiringa hanganga mō te whakahaere AI a tētahi kaiwhakarato kotahi, a tētahi kāwanatanga kotahi rānei.
The minimum number of board, committee, or assembly members required to be present before formal business can proceed. Set by the body's own rules. The platform tracks attendance against quorum and surfaces the result before formal motions are recorded.
[DeepL] Ko te iti rawa o ngā mema o te poari, o te komiti, o te huihuinga rānei e hiahiatia ana kia tae mai i mua i te haere tonu o ngā mahi ā-ture. Kua whakaritea e ngā ture ake o te rōpū. Ka aroturuki te papanga i te tae mai ki te kōrum, ā, ka whakaatu i te hua i mua i te tuhi i ngā tono ā-ture.
A formal decision recorded by a board or committee after a motion has been moved, seconded, discussed, and carried by vote. Resolutions form the persistent governance trail of the organisation; later decisions reference them by date and number.
[DeepL] He whakataunga ā-ture i tuhia e tētahi poari, e tētahi komiti rānei i muri i te whakatakoto, te tautoko, te matapaki, me te whakaaetanga mā te pōti. Ko ngā whakataunga te ara whakahaere mau tonu o te whakahaere; ka tohu ngā whakataunga ā muri ake ki a rātou mā te rā me te nama.
The Māori Data Sovereignty Network. "Mana" here carries authority and standing; "raraunga" is data. The network's principles articulate Māori data sovereignty as a collective right requiring both consent and ongoing kaitiakitanga.
[DeepL] Te Whatunga Rangatiratanga Raraunga Māori. Ko te kupu "mana" i konei e tohu ana i te mana me te tū; ko te kupu "raraunga" te raraunga. E whakamārama ana ngā mātāpono o te whatunga i te rangatiratanga raraunga Māori hei tika ā-hapori e hiahiatia ana te whakaaetanga me te kaitiakitanga tonutanga.
The 1840 treaty between the British Crown and many Māori rangatira (chiefs) that founds the constitutional relationship between Māori and the New Zealand state. Article 2 protects taonga — treasures including, in current reading, the data and cognition that mediate community life. The architecture's Te Tiriti framing engages this Article-2 duty.
[DeepL] Ko te Tiriti o Waitangi o te tau 1840 i waenga i te Karauna Piritene me ngā rangatira Māori maha, nāna i whakatū te hononga ture matua i waenga i ngā Māori me te kāwanatanga o Aotearoa. E tiaki ana te Wāhanga 2 i ngā taonga — arā, ngā rawa whakahirahira, tae atu ki ngā raraunga me te mātauranga e whakawhiti ana i te oranga hapori i te panui o nāianei. Ka whai wāhi te anga Te Tiriti o te hanganga ki tēnei haepapa o te Wāhanga 2.
A statement of how a system, organisation, or platform discharges its obligations under Te Tiriti o Waitangi. The platform publishes a Tiriti Compliance Statement; Dr Karaitiana Taiuru reviewed v0.1.
[DeepL] He tauākī mō te āhua e whakatutuki ai tētahi pūnaha, whakahaere, rānei tūāpapa i āna herenga i raro i Te Tiriti o Waitangi. Ka whakaputa te tūāpapa i tētahi Tauākī Whakatutuki Tiriti; i arotakehia e Dr Karaitiana Taiuru te putanga v0.1.
United States legislation that gives US authorities jurisdiction over data held by US-owned cloud and SaaS providers, regardless of where in the world that data physically resides. The principal legal reason the platform's vendor-prohibition rule excludes all US-owned infrastructure from the production request path.
[DeepL] He ture a te United States e tuku ana ki ngā mana o te United States te mana whakahaere ki ngā raraunga e puritia ana e ngā kaiwhakarato kapua me ngā ratonga pūmanawa hei ratonga (SaaS) nō te United States, ahakoa kei hea i te ao e noho ana aua raraunga ā-tinana. Ko te take ture matua i aukatia ai e te ture aukati kaiwhakarato o te papanga ngā hanganga katoa nō te United States i te ara tono whakaputa.
Vendor prohibition (US) — Te aukati i ngā kaihoko (Amerika)
The standing rule that no US-owned cloud, SaaS, or managed service may sit in the production request path — neither at the data layer nor at the inference layer. Driven by the CLOUD Act's reach over US-owned infrastructure.
[DeepL] Ko te ture tū tonu, kāore he kapua, he ratonga pūmanawa hei ratonga (SaaS), he ratonga whakahaere rānei e puritia ana e ngā Amerikana e taea te noho i roto i te ara tono whakaputa — kāore i te papa raraunga, kāore hoki i te papa whakamātau. Nā te whānuitanga o te Ture Kapua ki ngā hanganga e puritia ana e ngā Amerikana.
Read further:Codebase rule (CLAUDE.md); Paper A §1
Was this useful?Thanks — recorded.
Visibility — Te kitea
Who can see a given piece of content. The platform's visibility levels are All members, Specific subgroups, or Only me — never "public". Members-only by default; the framing is itself a privacy commitment.
[DeepL] Ko wai ka taea te kite i tētahi wāhanga ihirangi kua whakaritea. Ko ngā taumata kitea o te papanga ko Ngā mema katoa, Ngā rōpū iti motuhake, me Au anake — kāore rawa e wātea ki te marea. Mema anake te taunoa; ko tēnei anō te whakapuakitanga o te ū ki te tūmataitinga.
A formal expression of preference on a motion, recorded with the voter's identity, timestamp, and choice (in favour, opposed, abstain). The platform records individual votes for the audit trail; presented or aggregated according to the body's rules.
[DeepL] He whakapuakitanga ā-ture o te manakohanga mō tētahi tono, kua tuhia me te tuakiri o te pōti, te tohu wā, me tōna kōwhiringa (e tautoko ana, e whakahē ana, e kore e pōti). Ka tuhia e te papanga ngā pōti takitahi mō te ara arotake; ka whakaaturia, ka whakakotahitia rānei i runga i ngā ture o te rōpū.
A Waitangi Tribunal claim concerning Māori data sovereignty and the relationship between te ao Māori and emerging technology. Part of the contemporary New Zealand context the platform's governance design engages with.
[DeepL] He kerēme a te Kōti Waitangi e pā ana ki te rangatiratanga raraunga Māori me te hononga i waenga i te ao Māori me ngā hangarau e puta ake ana. He wāhanga tēnei o te horopaki o Niu Ioka o nāianei e pā ana ki te hoahoa whakahaere o te tūāpapa.
Identity & AccessTuakiri me te Whakaaetanga Uru(10)
Accreditation gate — Kaute whakamanatanga
An onboarding step that confirms a prospective tenant owner is who they claim to be and intends the platform for legitimate community use. Existing tenants are grandfathered; the gate applies only to new tenant creation.
[DeepL] He hipanga whakauru e whakamana ana he tangata tūturu te kaipupuri rēti e hiahiatia ana, ā, e whakamahere ana ia ki te whakamahi i te papanga mō te whakamahinga ā-hapori tika. Ka whakamanahia ngā kaipupuri rēti o nāianei; ka pā anake te kuaha ki te waihanga kaipupuri rēti hou.
A bounded read-only mode used by the demo Villages to let prospective members get a feel for the platform without joining. Tightly scoped; never used to expose live tenant content.
[DeepL] He aratau pānui-anake ā-rohe e whakamahia ana e te whakaaturanga o ngā Kāinga hei āwhina i ngā mema pea kia rongo i te papanga, kia kore ai e hono atu. He tino herea; kāore rawa e whakamahia hei whakaatu i ngā ihirangi o ngā kaipāpā noho ora.
The user-record flag that marks an account as a platform administrator. Used to grant access to system operations; never used to grant cross-tenant content access.
[DeepL] Ko te tohu rekoata-kaiwhakamahi e tohu ana i tētahi pūkete hei kaiwhakahaere papanga. Ka whakamahia hei tuku uru ki ngā whakahaere pūnaha; kāore rawa e whakamahia hei tuku uru ki ngā ihirangi o ētahi atu teneti.
The platform's privacy posture. Content is shared with named groups ("all members", "selected members", "only me", "everyone in community") never with abstract categories like "public" or "visitors". The terminology choice is itself a privacy commitment.
[DeepL] Te tūnga tūmataitinga o te papa. Ka tohatohahia ngā ihirangi ki ngā rōpū kua ingoa ("ngā mema katoa", "ngā mema kua tohua", "au anake", "ngā mema katoa o te hapori"), kāore rawa e tohatohahia ki ngā kāwai whānui pērā i te "tūmatanui" me te "manuhiri". Ko te kōwhiringa kupu nei he ū ki te tūmataitinga.
The senior administrative role within a tenant — the community member who provisioned it and holds the final say on tenant-scope decisions.
[DeepL] Te tūranga whakahaere matua i roto i tētahi kaiwhakamahi — te mema o te hapori i whakarite ai, ā, ko ia te tangata whai kupu whakamutunga mō ngā whakataunga e pā ana ki te whānuitanga o te kaiwhakamahi.
A role that authorises infrastructure-level operations (creating tenants, billing, health monitoring) — never tenant content access. The isPlatformAdmin flag authorises tenant-scoped system operations only; if code allows a platform admin to read tenant content, that is a security defect.
[DeepL] He tūranga e whakaaetia ana ki ngā mahi taumata hanganga (te waihanga kiritaki, te nama, te aroturuki hauora) — kāore rawa e whakaaetia kia uru ki ngā ihirangi a te kiritaki. Ko te haki isPlatformAdmin e whakaaetia ana anake ngā mahi pūnaha e pā ana ki te kiritaki; mēnā ka whakaaetia e te waehere kia pānui te kaiwhakahaere papanga i ngā ihirangi a te kiritaki, he hapa haumaru tērā.
The per-tenant address prefix — e.g. whanau.mysovereignty.digital. The platform routes by subdomain to identify which tenant a request is bound to.
[DeepL] Te tapanga whakauru wāhitau mō ia kaiwhakamahi — hei tauira whanau.mysovereignty.digital. Ka whakatere te papa mā ngā subdomain kia tautuhi ai ko tēhea kaiwhakamahi e hono ana te tono.
One community on the platform — a Village, a rūnanga, a parish, an extended family, a small business. Each tenant has its own subdomain, its own corpus, its own members, its own governance, and its own situated language layer cohort assignment.
[DeepL] He hapori kotahi i runga i te papanga — he kāinga, he rūnanga, he pārihi, he whānau whānui, he pakihi iti. He subdomain motuhake tō ia kaipā, he kohinga kupu motuhake, he mema motuhake, he whakahaere motuhake, me tana tohinga rōpū paparanga reo ā-tūnga motuhake.
A person legally empowered to act on behalf of an organisation, marae, or charitable trust. Trustees carry governance obligations beyond ordinary membership — fiduciary duty, accountability for decisions, signatory authority — and the platform reflects these in role configuration where governance product types are in use.
[DeepL] He tangata kua whakamanahia ā-ture kia mahi mō tētahi whakahaere, tētahi marae, tētahi rāngai pūtea āwhina rānei. He kawenga whakahaere ā ngā kaitiaki e nui ake ana i ngā kawenga o te mema noa — te haepapa pono, te haepapa mō ngā whakataunga, te mana haina — ā, ka whakaata tēnei i te papanga i roto i te whirihoranga tūranga, i ngā wā e whakamahia ana ngā momo hua whakahaere.
The actor a threat model is concerned with — anyone or anything whose interests could lead them to attack the system. Includes opportunistic attackers, motivated attackers, insiders, and (for sovereignty-grade systems) jurisdictional pressure on infrastructure providers.
[DeepL] Ko te tauira whakatumatuma e pā ana ki ngā kaiwhakaari — ko te hunga, ko ngā mea rānei e āhei ana ā rātou hiahia ki te whakaeke i te pūnaha. Kei roto i tēnei ngā kaiwhakaeke whai wā, ngā kaiwhakaeke whai take, ngā tāngata o roto, ā, mō ngā pūnaha taumata rangatiratanga, ko te pēhanga ā-ture ki ngā kaiwhakarato hanganga.
An encryption algorithm and mode (Advanced Encryption Standard, 256-bit key, Cipher Block Chaining). Used for at-rest encryption of certain platform data.
[DeepL] He tātai whakamunatanga me tētahi aratau (Paerewa Whakamunatanga Matatau, kī 256-bit, here here poraka whakamuna). Ka whakamahia hei whakamuna i ngā raraunga papanga e noho ana i te wā rokiroki.
The total set of points at which an outside actor could try to enter a system, extract data, or interfere with operation. A larger attack surface is harder to defend; the platform's vendor-prohibition rule and minimal-dependency posture exist partly to keep the surface small.
[DeepL] Ko te huinga katoa o ngā wāhi e taea ai e tētahi kaiwhakahaere o waho te ngana ki te uru ki roto i te pūnaha, ki te tango raraunga, ki te whakararuraru rānei i ngā mahi. He uaua ake te tiaki i tētahi mata whakaeke nui; ko te ture aukati kaiwhakarato o te papanga me tōna tūnga whakawhirinaki iti e wāwāhi ana kia iti tonu te mata whakaeke.
A tamper-evident record of significant operations — who did what, when, against which record — kept for forensic and compliance purposes. The platform's audit log includes export operations (GDPR DSR fulfilment), administrative actions, and per-request inference-routing decisions.
[DeepL] He rēhita e kitea ana mēnā kua whakarerekēhia, o ngā mahi nui — ko wai i mahi i te aha, i te wāhea, ki tēhea rēhita — e puritia ana mō ngā take tātari ā-ture me te ū ki ngā ture. Kei roto i te rangitaki arotake o te papanga ngā mahi kaweake (hei whakatutuki i te DSR o te GDPR), ngā mahi whakahaere, me ngā whakataunga whakatere whakamātau mō ia tono.
Proving who you are. The platform authenticates members by email + password backed by salted hash, with optional second factors. Distinct from authorisation, which proves what you may do.
[DeepL] Te whakaū ko wai koe. Ka whakamanahia e te papanga ngā mema mā te īmēra me te kupuhipa, ā, ka tautokona e te hash tote, me ētahi atu āhuatanga tuarua hei kōwhiringa. He rerekē tēnei i te whakaaetanga, e whakaū ana i ngā mahi ka taea e koe.
Proving what you may do once authenticated. The platform's authorisation enforces tenant isolation, role boundaries (member, owner, platform admin), and feature-gate policies. Mistakes here are a more frequent source of breach than authentication failures.
[DeepL] E whakaatu ana i ngā mahi ka taea e koe i muri i tō whakamanatanga. Ka whakatinana te whakamanatanga o te papaanga i te wehewehe o ngā kaiwhakamahi, ngā rohe tūranga (mema, rangatira, kaiwhakahaere papaanga), me ngā kaupapa here ā-kuaha āhuatanga. He puna pakaru nō ngā hē i konei i te nuinga ake i ngā hē whakamanatanga.
Automated traffic from non-human clients with hostile intent — credential stuffing, scraping, comment spam, denial-of-service. Defended against by rate limiting, challenge questions on signup forms, and behavioural detection. The platform's invitation-based onboarding also limits bot exposure on the member side.
[DeepL] Ngā rerenga aunoa mai i ngā kiritaki kāore he tangata, he whāinga kino — te whakakī ingoa me ngā kupuhipa, te kohikohi raraunga, te parahanga kōrero, me te aukatinga ratonga. Ka ārai i ēnei mā te here tere, ngā pātai wero i ngā puka rēhita, me te kitenga whanonga. Ka herea hoki te tukanga whakauru mā te tono a te papanga, kia iti ake te whakaaturanga o ngā pōtae ki te taha mema.
An attack in which the attacker exhaustively tries every plausible secret for a single account until one works. Defended against by complexity requirements on credentials, rate limiting on authentication endpoints, and account-lockout after repeated failures.
[DeepL] He whakaeke e whakamātau whānuitia ana e te kaiwhakaeke ngā kupuhipa (rānei ngā kī) katoa e taea ana mō tētahi pūkete kotahi kia kitea tētahi e whai hua ana. Ka ārai i tēnei mā ngā kaupapa here kupuhipa pakari, te here tere i ngā tauranga whakamanatanga, me te aukati pūkete i muri i ngā hapa.
An HTTP header that tells the browser which sources of script and style it may execute. The platform's CSP forbids inline scripts; all JavaScript must live in external files. This prevents large classes of injection attack.
[DeepL] He upoko HTTP e whakamōhio ana ki te kaitirotiro ko ēhea puna o ngā tuhinga me ngā āhua ka taea e ia te whakahaere. Ka aukati te CSP o te papanga i ngā tuhinga ā-roto; me noho katoa ngā JavaScript ki ngā kōnae o waho. Mā konei ka ārai i ngā momo whakaeke werohanga nui.
Read further:Codebase rule (CLAUDE.md)
Was this useful?Thanks — recorded.
Credential stuffing — Whakakī ingoa me ngā kōrero urunga
An automated attack in which the attacker tries username-and-password pairs leaked from other breaches against the platform's login form, betting that members reused credentials. Defended against by rate limiting, MFA, and detecting login attempts from anomalous locations.
[DeepL] He whakaeke aunoa e whakamātau ana te kaiwhakaeke i ngā huinga ingoa kaiwhakamahi me ngā kupuhipa i pakaru i ētahi atu pakaru, ki te whakamātau ki te takiuru ki te pae, e whakapono ana ka whakamahia anō e ngā mema ā rātou mōhiohio takiuru. Ka ārai tēnei mā te here tere, te whakamanatanga tohu-rua (MFA), me te kitea o ngā whakamātau takiuru mai i ngā wāhi rerekē.
An attack class in which a malicious site causes a user's browser to issue an authenticated request to a different site. Mitigated by requiring a per-session token on every state-changing request; the platform's middleware enforces this app-wide.
[DeepL] He momo whakaeke e akiaki ana i te pūtirotiro a te kaiwhakamahi ki te tuku tono kua whakamanahia ki tētahi atu pae mā te pae kino. Ka ārai tēnei mā te tono i tētahi tohu mō ia wāhanga i ia tono e whakarerekē ana i te tūnga; ka whakatinana tēnei e te papanga waenga o te papaanga puta noa i te taupānga.
An attack class in which an attacker arranges for malicious script to be served from a trusted origin, executing in another user's browser with that origin's privileges. Defended against by escaping output, using CSP to forbid inline scripts, and sanitising rich content before render.
[DeepL] He momo whakaeke e whakaritea ana e te kaiwhakaeke kia tukuna he tuhinga kino mai i tētahi puna e whakawhirinaki ana, kia whakahaere ai i roto i te kaitirotiro o tētahi atu kaiwhakamahi me ngā mana o taua puna. Ka ārai i tēnei mā te karo i ngā putanga, mā te whakamahi i te CSP hei aukati i ngā tuhinga ā-roto, me te whakamā i ngā ihirangi whai rawa i mua i te whakaatu.
CVE (Common Vulnerabilities and Exposures) — CVE (Ngā ngoikoretanga me ngā whakaaturanga noa)
A globally-coordinated catalogue of publicly-known security vulnerabilities, each with a unique identifier of the form CVE-YYYY-NNNNN. Vendors and security tools track CVEs to know which versions of which software need patching.
[DeepL] He rārangi ā-ao kua whakaritea o ngā ngoikoretanga haumaru e mōhiotia ana e te marea, ā, ia ngoikoretanga he tuakiri ahurei i te āhua CVE-YYYY-NNNNN. Ka whai i ngā CVE ngā kaiwhakarato me ngā taputapu haumaru kia mōhio ai rātou ko ēhea putanga o ngā pūmanawa e hiahiatia ana kia whakatikahia.
Denial of service (DoS / DDoS) — Te aukatinga ratonga (DoS / DDoS)
An attack that aims to make a service unavailable to legitimate users by overwhelming it with traffic. The distributed variant (DDoS) uses many sources at once. Defended against by upstream filtering at the cloud-provider edge, rate limits, and circuit-breakers in the application.
[DeepL] He whakaeke e whāia ana kia kore ai e wātea te ratonga ki ngā kaiwhakamahi mana, mā te whakawhē i a ia ki te nui rawa o ngā rerenga raraunga. Ko te momo tohatoha (DDoS) e whakamahi ana i ngā puna maha i te wā kotahi. Ka ārai i tēnei mā te tātari i ngā rerenga raraunga e rere mai ana i ngā hononga matua o te kaiwhakarato kapua, ngā here tere, me ngā pūwaha aukati huringa i roto i te taupānga.
A CVE in a library the platform depends on rather than in the platform's own code. The April 2026 sweep remediated 19 dependency vulnerabilities to zero; a 48-hour patch cycle was adopted to keep new disclosures from accumulating.
[DeepL] He ngoikoretanga haumaru (CVE) kei roto i tētahi whare pukapuka e whakawhirinaki ana te papanga, kaua ki roto i te waehere o te papanga anō. I whakatika e te tirotiro o Aperira 2026 ngā ngoikoretanga haumaru e 19 o ngā whakawhirinaki kia kore rawa; i whakamahia he huringa whakatika 48-haora kia kore ai e kohi ngā whakamōhiotanga hou.
Encryption at rest — Whakamunatanga i te wā e tū ana
Data encrypted while stored on disk, so that physical access to the disk does not yield readable content. The platform uses AES-256-CBC for tenant data at rest on both servers, completed in the April 2026 security posture sweep.
[DeepL] Kua whakamunatia ngā raraunga i te wā e penapenahia ana ki te kōpae, kia kore ai e taea te pānui i ngā ihirangi ahakoa te uru ā-tinana ki te kōpae. Ka whakamahi te papanga i te AES-256-CBC mō ngā raraunga rēniti i te wā e tau ana ki runga i ngā tūmau e rua, ā, i oti i te tirohanga tūnga haumaru o Aperira 2026.
Encryption in transit — Whakamunatanga i te haerenga
Data encrypted while moving between systems — for instance, between a member's browser and the platform — so that anyone observing the network cannot read the contents. SSL/TLS provides this on the public web; the platform also encrypts data flowing between its own services.
[DeepL] Ka whakamunatia ngā raraunga i te wā e neke ana i waenga i ngā pūnaha — hei tauira, i waenga i te pūtirotiro a te mema me te papaanga — kia kore ai e taea e te hunga e mātakitaki ana i te whatunga te pānui i ngā ihirangi. Ka whakaratohia tēnei e SSL/TLS i runga i te ipurangi tūmatanui; ā, ka whakamunatia hoki e te papaanga ngā raraunga e rere ana i waenga i āna ake ratonga.
A pattern of putting structural barriers in place so that a future mistake — accidentally committing a confidential file, accidentally pushing a draft to a public remote — fails closed rather than succeeding. The platform's restricted-class .gitignore patterns and pre-push file-list audit are forward-protection measures.
[DeepL] He tauira tēnei o te whakatū i ngā ārai hanganga kia aukati i tētahi hē ā muri ake — pērā i te tuku hē i tētahi kōnae tūmataiti, i te tuku hē i tētahi tuhinga tuatahi ki tētahi pūnaha mamao tūmatanui — kia kati, kaua kia angitu. Ko ngā tauira .gitignore ā-kāwai herea o te papanga me te arotake rārangi kōnae i mua i te tuku he tikanga tiaki ā mua.
Authentication that requires two or more independent factors — something you know (a password), something you have (a phone or hardware key), or something you are (a biometric). Defends against credential leaks because a leaked password alone is insufficient.
[DeepL] Te whakamana e hiahiatia ana kia rua, kia neke atu rānei ngā āhuatanga motuhake — he mea e mōhio ana koe (he kupuhipa), he mea kei a koe (he waea pūkoro, he kī taputapu rānei), he mea ko koe (he tohu ā-tinana, he tohu ā-whatu rānei). Ka ārai i te rerenga o ngā tohu uru, nā te mea kāore e ranea te kupuhipa anake kua rerengia.
A regularly-updated list of the ten most consequential web-application security risks, maintained by the Open Web Application Security Project. The platform's HARD RULES on input handling, CSP, CSRF, and authorisation map directly onto OWASP categories.
[DeepL] He rārangi e whakahou ana ia wā o ngā tūraru haumaru taupānga tukutuku tekau tino whai pānga, e tiakina ana e te Open Web Application Security Project. Ko ngā ture mārō o te papanga mō te whakahaere urunga, CSP, CSRF, me te whakaaetanga e hāngai tika ana ki ngā kāwai OWASP.
Social-engineering attacks in which the attacker impersonates a trusted party (a bank, a colleague, the platform itself) to trick a member into revealing credentials or approving a transaction. Defended against by member education, MFA, and clear sender-domain conventions.
[DeepL] Ngā whakaeke hangarau ā-pāpori e whakakāhua ana te kaiwhakaeke hei tangata whakawhirinaki (pēke, hoa mahi, te tūāpapa anō rānei) kia whakapohehe i tētahi mema kia whakaatu i ōna tohu uru, kia whakaae rānei ki tētahi whakawhitinga moni. Ka ārai ēnei mā te mātauranga o ngā mema, te whakamana-ā-maha (MFA), me ngā tikanga mārama mō ngā rohe ingoa kaituku īmēra.
Principle of least privilege — Mātāpono o te mana iti rawa
The standing rule that every account, service, and process is given only the minimum access needed to do its job — and no more. Limits the blast radius when a credential is compromised. Underpins the platform-admin / member separation and the per-tenant database scoping.
[DeepL] Ko te ture tū tonu e mea ana kia whakawhiwhia ia pūkete, ia ratonga, ia tukanga ki te uru iti rawa e hiahiatia ana kia tutuki ai tana mahi — ā, kāore e nui ake. Ka here i te awhe pānga ina pakaru tētahi tohu whakau. E tautoko ana i te wehenga i waenga i te kaiwhakahaere papanga me ngā mema, me te whakarite i te rahinga o ngā papa raraunga mō ia tangohanga.
A constraint on how often a given client may issue requests within a window. The platform applies rate limits to all API surfaces. Defends against brute-force attacks, scraping, and bot traffic; also catches accidental client-side bugs that loop on requests.
[DeepL] He here mō te auau ka tukuna e tētahi kiritaki ngā tono i roto i tētahi matapihi. Ka whakamahi te papa i ngā here tere ki ngā atanga API katoa. Ka tiaki i a ia i ngā whakaeke kaha-whānui, i te kohikohi raraunga, me ngā rerenga bot; ā, ka hopu hoki i ngā pepeke ohorere i te taha kiritaki e huri ana i ngā tono.
The practice of periodically replacing API keys, encryption keys, and similar secrets so that a leaked secret has bounded usefulness. The platform's mdsl Ops Hub schedules a randomly-selected secret each month from a configurable pool, with an inverse-recency weight so freshly-rotated secrets are less likely to be picked next.
[DeepL] Ko te tikanga o te whakakapi i ngā kī API, ngā kī whakamunatanga, me ērā atu mea huna i ngā wā katoa, kia whai here ai te whaihua o tētahi mea huna kua pakaru. Ka whakaritea e te mdsl Ops Hub o te papaanga tētahi mea huna kua tohua ā-tūturu ia marama mai i tētahi puna ka taea te whakarite, me te taumaha huri-whakamuri kia iti ake ai te tūpono ka tohua anō ngā mea huna kua huri hou.
An attack class in which an attacker supplies database-language fragments through an input field, hoping the application splices them into a database query rather than treating them as data. Defended against by parameterised queries; the platform's MongoDB driver parameterises by default.
[DeepL] He momo whakaeke e tuku ana te kaiwhakaeke i ngā wāhanga reo pātengi raraunga mā tētahi pae urunga, e tūmanako ana ka whakauru te taupānga i ēnei ki tētahi pātai pātengi raraunga, kaua e whakaarohia hei raraunga. Ka ārai tēnei mā ngā pātai kua whakaritea ki ngā tawhā; ko te taraiwa MongoDB o te papanga e whakarite ana i ngā tawhā i te taunoa.
Secure Sockets Layer / Transport Layer Security — the protocol that encrypts data in transit between browsers and servers. The padlock in the browser address bar.
[DeepL] Secure Sockets Layer / Transport Layer Security — te kawa e whakamuna ana i ngā raraunga e haere ana i waenga i ngā kaitirotiro me ngā tūmau. Ko te pāhī i te pae wāhitau o te kaitirotiro.
Read further:Standard web infrastructure
Was this useful?Thanks — recorded.
Threat model — Mōdela mōrearea
A structured statement of who might attack a system, what they want, what they could do, and what defences answer each move. Paper A §4 sets out the platform's threat model: adversaries, sovereignty invariants, and testable predicates.
[DeepL] He tauākī hanganga e whakaatu ana ko wai pea ka whakaeke i tētahi pūnaha, he aha tā rātou e hiahia ana, he aha ngā mahi ka taea e rātou, me ngā ārai e whakautu ana ki ia nekehanga. Te Puka A §4 e whakarārangi ana i te tauira mōrearea o te tūāpapa: ngā hoariri, ngā āhuatanga tū rangatiratanga, me ngā whakapae ka taea te whakamātautau.
Vulnerability disclosure — Whakaputanga o ngā ngoikoretanga haumaru
The process by which a security flaw is reported to the affected party, fixed, and (typically after some delay) disclosed publicly. The platform's posture is to disclose promptly with credit; the codebase tracks disclosed vulnerabilities and their patch dates.
[DeepL] Ko te tukanga e pūrongo ai tētahi hapa haumaru ki te rōpū kua pāngia, ka whakatikahia, ā, (i te nuinga o te wā, i muri i tētahi whakaroa) ka whakapuakina ki te marea. Ko te tū o te papanga he whakaputa wawe me te whakamōhio i te kaituhi; e aro turuki ana te pūtake waehere i ngā ngoikoretanga kua whakapuakina me ngā rā whakatika.
What a community AI has learned and remembers about a member personally across conversations — preferences, recurring questions, prior context. Members have full visibility and control: anything can be inspected and removed. The pattern itself is significant: memory is a member-controlled asset, not a vendor-controlled extraction.
[DeepL] He aha ngā mea kua akohia e te AI hapori, ā, e maumahara ana mō tētahi mema i ngā kōrero – ngā manakohanga, ngā pātai e hoki mai ana, me ngā horopaki o mua. He tino kitea, he mana whakahaere hoki ā ngā mema: ka taea te tirotiro, te tango rānei i ngā mea katoa. He mea nui te tauira nei: he rawa e whakahaerehia ana e ngā mema te mahara, ehara i te tangohanga e whakahaerehia ana e te kaihoko.
Training a model for a community type that does not yet have a deployment tenant. The project does not do this; the corpus is then necessarily speculative rather than the community's own content.
[DeepL] Te whakangungu i tētahi tauira mō tētahi momo hapori kāore anō kia whai kaipāpā tuku. Kāore te kaupapa e mahi ana i tēnei; nā reira, he whakapae noa te kohinga raraunga, ehara i te ihirangi ake o te hapori.
A general-purpose language model used as the starting point before any community-specific layering. The project's current base is the 14-billion-parameter Qwen2; FAQ layering and steering vectors are then applied to make it answer in a specific community's register.
[DeepL] He tauira reo whānui e whakamahia ana hei tīmatanga i mua i ngā paparanga mō ia hapori. Ko te turanga o tēnei kaupapa ināianei ko te Qwen2 me ōna tawhā e 14 piriona; ka tāpirihia ngā paparanga FAQ me ngā arataki ā-vector kia whakautu ai ia i runga i te reo o tētahi hapori motuhake.
Whether the model's grounded responses cite the community-corpus passages they draw from, and whether those citations are precise. Tracked alongside refusal discipline as a measure of model trustworthiness.
[DeepL] Kei te whakahua e ngā whakautu tūhonohono o te tauira ngā wāhanga o te kohinga hapori i tangohia ai e rātou, ā, he tika rānei aua tohutoro. Ka whai i te whakahaere whakahē hei ine i te pono o te tauira.
A measure of how much the model's working memory is being stretched — token usage relative to its context window. The Tractatus ContextPressureMonitor service raises verification intensity when pressure rises, on the principle that mistakes correlate with crowded context.
[DeepL] He ine i te nui o te whānuitanga o te mahara mahi o te tauira — te whakamahinga tohu i runga i tōna matapihi horopaki. Ka whakanui te ratonga Tractatus ContextPressureMonitor i te kaha whakamana ina piki te pēhanga, i runga i te mātāpono e hono ana ngā hē ki te horopaki kapi.
Synthetic training examples in which a model's wrong answer is paired with a corrected good answer. The project does not use them: in testing they introduced a bias surface that the model overfitted, degrading performance overall. Behavioural correction is done with steering vectors instead.
[DeepL] Ngā tauira whakangungu hangarua e honoa ana te whakautu hē o te tauira ki tētahi whakautu tika kua whakatikahia. Kāore te kaupapa e whakamahi ana i ēnei: i te whakamātautau, i whakaurua e rātou he mata whakatōpū hē i whakapāpāhia e te tauira, ā, i heke ai te mahi whānui. Ka mahia te whakatikatika whanonga mā ngā pūwāhi ārahi.
A Tractatus framework service that validates a proposed AI action against stored explicit instructions. When the AI's proposed action contradicts a stored instruction, the instruction wins.
[DeepL] He ratonga anga Tractatus e whakamana ana i tētahi mahi AI kua tūtohutia ki ngā tohutohu mārama kua penapenahia. Ki te taupatupatu te mahi kua tūtohutia e te AI ki tētahi tohutohu kua penapenahia, ka toa te tohutohu.
The principle that the model's own behaviour (its grounding, its refusal habits) and the post-hoc safety filter are distinct layers; both must operate; both must be reported in evaluation. The filter alone is never treated as the whole answer to grounding.
[DeepL] Ko te mātāpono ko te whanonga ake o te tauira (tōna tūāpapa, ōna tikanga whakahē) me te tātari haumaru i muri i te tauira he paparanga motuhake; me mahi ngātahi rāua; me rīpoata rāua i roto i te arotake. Kāore te tātari anake e whakaarohia hei whakautu katoa mō te tūāpapa.
The held-aside questions a candidate model is scored against. The project's standing evaluation set draws from the active tenant corpora plus a corpus of community-relevant out-of-corpus questions, so that refusal behaviour is measured alongside grounded-answer behaviour.
[DeepL] Ko ngā pātai kua waiho hei paearu mō te aromatawai o tētahi tauira kaitono. Ka whiwhi te huinga aromatawai tūnga o te kaupapa i ngā koropore kiritaki e whakamahia ana, me tētahi koropore pātai e pā ana ki te hapori, ā, kāore i roto i ngā koropore, kia taea ai te ine i ngā whanonga whakahē i te taha o ngā whanonga whakautu whai turanga.
A technique for extending what a community AI can answer by adding curated question-and-answer pairs at runtime, rather than retraining the model itself. The pairs live in the platform database; they are consulted before the model generates its answer. Every FAQ entry is anchored to a verifiable source — no aspirational answers.
[DeepL] He tikanga hei whakawhānui i ngā whakautu ka taea e tētahi AI hapori te tuku, mā te tāpiri i ngā tokorua pātai me ngā whakautu kua whiriwhiria i te wā e rere ana te pūnaha, kaua e whakahou i te tauira anō. Kei roto i te pātengi raraunga o te tūāpapa ēnei tokorua; ka tirohia i mua i te waihanga a te tauira i tana whakautu. Ia tomokanga FAQ e hono ana ki tētahi puna ka taea te whakamana — kāore he whakautu wawata.
Instruction persistence — Te mau tonu o ngā tohutohu
The architectural pattern of classifying every instruction by how strongly it should override later contradictory model behaviour, then storing it externally so the model's training priors cannot quietly overwrite it.
[DeepL] Ko te tauira hoahoanga e whakarōpū ana i ia tohutohu i runga i te kaha e tika ana kia whakakorehia ai ngā whanonga tauira e whakahē ana ā muri ake, ā, ka penapenahia ki waho kia kore ai e taea e ngā tūtohu whakangungu o te tauira te whakakapi i taua tohutohu.
A pattern in a model's outputs that comes from the structure of the training data and architecture rather than from any conscious choice — for instance, a frontier model defaulting to property-rights language when asked about Māori data guardianship, because that is what the global corpus reflects.
[DeepL] He tauira i ngā putanga o tētahi tauira rorohiko e ahu mai ana i te hanganga o ngā raraunga whakangungu me te hoahoa, ehara i te mea nā tētahi kōwhiringa mōhio — hei tauira, ka huri aunoa tētahi tauira matatau ki te reo mō ngā mana rawa whenua ina pātaihia mō te tiakitanga raraunga Māori, nā te mea koirā te mea e whakaata ana i te kohinga raraunga ā-ao.
A Tractatus framework service that triggers an AI self-check on alignment, coherence, and safety before executing a complex operation. Used selectively (not on every task) to avoid overhead on routine work.
[DeepL] He ratonga anga Tractatus e whakaoho ana i tētahi arotakenga ā-AI mō te taurite, te hono, me te haumaru i mua i te whakahaere i tētahi mahi matatini. Ka whakamahia ā-kōwhiringa (ehara i ia mahi katoa) kia karo ai i te taumaha o ngā mahi auau.
A question that has no answer grounded in the community's own content. The intended response is a refusal, not a guess. Measuring how often the model correctly refuses out-of-corpus questions is part of evaluation.
[DeepL] He pātai kāore he whakautu e ū ana ki ngā ihirangi a te hapori. Ko te whakautu e hiahiatia ana he whakahē, ehara i te matapae. Ko te ine i te auau e tika ana ka whakahē te tauira i ngā pātai kāore i roto i te kohinga kuputuhi he wāhanga o te aromatawai.
Refusal discipline — Te whakatikatika mō te whakahē
The rate at which a model correctly declines to answer questions outside its corpus. A measurable property; the project tracks refusal-by-default as one of the ways modified models drift away from the FAQ-layered baseline.
[DeepL] Ko te auau e whakakore tika ai tētahi tauira i te whakautu i ngā pātai kei waho o tōna kohinga raraunga. He āhuatanga ka taea te ine; e aroturuki ana te kaupapa i te whakakore-a-taunoa hei kotahi o ngā huarahi e rereke haere ana ngā tauira kua whakarerekētia i te paerewa FAQ-ā-papa.
Situated Language Layer (SLL) — Papanga Reo Whakaritea
A community-scoped small language model trained on the tenant's own content, governed by the tenant's own authority, and operated on infrastructure inside the tenant's own jurisdictional reach. Replaces a frontier vendor model as the runtime cognition layer of the sovereign-record platform.
[DeepL] He tauira reo iti kua tautuhia e te hapori, kua whakangungua ki ngā ihirangi ake o te kaiwhakamahi, e whakahaerehia ana e tōna ake mana, ā, e whakahaerehia ana i runga i ngā hanganga kei roto i tōna ake rohe mana. Ka whakakapi i tētahi tauira kaiwhakarato ā-mua hei paparanga māramatanga i te wā whakahaere o te papanga rēkōta rangatira.
Small adjustments applied at inference time that nudge a model's behaviour in a particular direction — for example, refusing more reliably on a class of out-of-corpus questions, or down-weighting a known mechanical bias. Steering vectors operate on the model's activations rather than its training data, so the underlying corpus is left as the community provided it. Discussed extensively across the project's research because they are the principal mechanism by which the platform corrects model behaviour without modifying weights.
[DeepL] He whakarerekētanga iti e whakamahia ana i te wā whakatau hei akiaki i te whanonga o te tauira ki tētahi ahunga motuhake — hei tauira, kia whakakore pono ake i ngā pātai o tētahi kāwai kāore i roto i te kohinga. Ka mahi ki ngā whakahohe, ehara i te kohinga whakangungu, nō reira ka waiho te kohinga i te āhua i whakaratoa e te hapori.
The illustrative case where Claude Code was told to use port 27027 but used 27017 instead — because MongoDB's default port is 27017 and the model's statistical priors auto-corrected the explicit instruction. The incident motivated the architectural rule that human instructions must be classified, persisted, and validated against — not absorbed into the model's prior.
[DeepL] Ko te tauira whakaahua i mea atu ai ki a Claude Code kia whakamahi i te tauranga 27027, engari i whakamahia e ia te 27017 — nā te mea ko te tauranga taunoa o MongoDB ko te 27017, ā, nā ngā priors tauanga o te tauira i whakatika aunoa i te tohutohu mārama. Nā tēnei aitua i whakaoho i te ture hanganga e mea ana me whakarōpū, me rokiroki, me whakamana ngā tohutohu a te tangata — kaua e whakaurua ki ngā priors o te tauira.
Read further:Tractatus working paper; multiple blog posts
Was this useful?Thanks — recorded.
Tier-1 cohort — Rōpū tuatahi
A community-type model that is trained and deployed in production today. The five Tier-1 cohorts are whānau (Māori extended-family contexts), episcopal (Anglican parish), generic-community fallback, family-history, and small-business.
[DeepL] He tauira momo hapori kua whakangungua, ā, e whakamahia ana i te whakaputanga i tēnei rā. Ko ngā rōpū e rima o te Tier-1 ko whānau (ngā horopaki whānau whānui Māori), epikopā (parīhi Anglikana), te kōwhiringa hapori whānui, te hītori whānau, me te pakihi iti.
A community-type model that has been designated but not yet commissioned because no first tenant of that type is in deployment. The current four Tier-2 cohorts are conservation, diaspora, clubs, and alumni.
[DeepL] He tauira ā-hapori kua tohua, engari kāore anō kia whakamanahia, nā te mea kāore he kaipāmu tuatahi o taua momo e whakamahia ana. Ko ngā rōpū Tier-2 e whā o nāianei ko te tiaki taiao, te diaspora, ngā karapu, me ngā alumni.
The standing rule that no Tier-2 cohort is commissioned for training until at least one tenant of that type is in deployment. Prevents corpora grounded in speculation rather than the community's own content, and conserves scarce GPU time for production cohorts.
[DeepL] Ko te ture tū tonu, kāore he rōpū Tier-2 e whakaritea mō te whakangungu kia tae rā anō kia kotahi te kaiwhakamahi o taua momo e whakamahia ana. Ka ārai i ngā koropora e hangai ana ki ngā whakapae, kaua ki ngā ihirangi a te hapori, ā, ka tiaki i te wā GPU iti mō ngā rōpū whakaputa.
One model build for one community type — a Tier-1 cohort that has been trained on its tenant-type's corpus and deployed. "Cohort" rather than "model" because each carries the standing operating discipline rather than a one-off configuration.
[DeepL] He tauira kotahi mō tētahi momo hapori — he rōpū Tier-1 kua whakangungua ki te kohinga raraunga mō tōna momo kaipāmu, ā, kua whakatinanahia. "Rōpū" kē i te "tauira", nā te mea ka mau ia i te tikanga whakahaere mau tonu, ehara i te whirihoranga kotahi anake.
Training data deduplication — Te tango tārua o ngā raraunga whakangungu
The standard text-cleaning step of removing repeated content from a training corpus. The project does not do this for community corpora: apparent duplicates (canonical lines, repeated tikanga formulae, recurrent governance phrasings) are reinforcement that shapes the corpus's emphasis structure, not noise.
[DeepL] Ko te hipanga paerewa o te horoi tuhinga he tango i ngā ihirangi tāruarua i roto i tētahi koropore whakangungu. Kāore te kaupapa e mahi ana i tēnei mō ngā koropore hapori: ko ngā tāruarua e kitea ana (ngā rārangi mana, ngā tauira tikanga tāruarua, ngā rerenga whakahaere e puta anō ana) he whakakaha e āhua ana i te hanganga aro o te koropore, ehara i te haruru.
The in-platform AI assistant that answers questions about how to use the system, the community's own customs, and locally-curated knowledge. Trained on the community's own corpus rather than on a vendor's global one; the day-to-day expression of the situated language layer.
[DeepL] Ko te kaiāwhina AI kei roto i te papanga e whakautu ana i ngā pātai mō te whakamahi i te pūnaha, ngā tikanga motuhake o te hapori, me ngā mōhiotanga ā-rohe kua whakaritea. I whakangungua ki tōna ake kohinga rauemi a te hapori, kaua ki tētahi kohinga ā-ao a te kaiwhakarato; ko te whakaaturanga o ia rā o te paparanga reo ā-horopaki.
Any technique that alters the parameters of the base model itself — fine-tuning, distillation, preference-tuning, and so on. The project's standing position: every weight-modification protocol tested to date has underperformed FAQ layering on the project's evaluation set. Until one is identified that demonstrably improves, none is adopted.
[DeepL] He tikanga katoa e whakarerekē ana i ngā tawhā o te tauira taketake anō — te whakangāwari āta, te whakawāwā, te whakangāwari i runga i ngā manakohanga, me ērā atu. Te tūnga o te kaupapa: kua whakamātauria ngā kawa whakarerekē taumaha katoa tae noa mai ki tēnei rā, ā, kāore i eke ki te whai hua o te whakakotahitanga FAQ i runga i te huinga aromatawai a te kaupapa. Kāore he mea ka whakaaetia kia whakamahia, kia kitea tētahi e whakaatu ana i te whakapai ake.
Process & Operating DisciplineTukanga me te Whakatakotoranga Mahi(10)
Cache bust — Whakakore i te pātaka raraunga
A query-string version stamp appended to every JavaScript and CSS file load, updated on each deploy so that browsers fetch the new file rather than serve the old one from cache. Without cache busting, deploys can briefly serve the new code that then gets overridden by an old cached worker.
[DeepL] He tohu putanga rārangi pātai e tāpirihia ana ki ia utaina o ngā kōnae JavaScript me CSS, ā, ka whakahouhia ia wā ka tukuna he whakahou, kia tiki ai ngā kaitirotiro i te kōnae hou, kaua e tuku i te kōnae tawhito mai i te pūkaupapa. Mēnā kāore he tikanga pakaru pūkaupapa, ka taea e ngā whakahou te tuku wawe i te waehere hou, ā, ka whakakapiia taua waehere e tētahi kaimahi tawhito kei roto i te pūkaupapa.
A neural machine translation service. The platform's standing translation tool — used for all content translation including te reo Māori, with culturally-loaded terms flagged for human review.
[DeepL] He ratonga whakamāori mīhini whatunga neorana. Ko te taputapu whakamāori tū o te tūāpapa — e whakamahia ana mō te whakamāori i ngā ihirangi katoa, tae atu ki te reo Māori, ā, ka tohu i ngā kupu whai tikanga ahurea kia arotakehia e te tangata.
The community deploy script. Runs ESLint, validates Content Security Policy, performs the cache bust, updates the service worker version, and pushes to both application servers. Bypassing it is a standing rule violation.
[DeepL] Ko te tuhi whakaurunga a te hapori. Ka whakahaere i a ESLint, ka whakamana i te Kaupapa Here Haumaru Ihirangi, ka pakaru i te pūkairangi, ka whakahou i te putanga o te kaimahi ratonga, ā, ka tuku ki ngā tūmau taupānga e rua. He hē ki te ture tūmau te karo i tēnei.
A working mode for the AI assistant: bounded to an approved plan; only the planned files are touched; anything unexpected stops execution and returns to PLAN. Commit diffs match plan diffs line-for-line.
[DeepL] He aratau mahi mō te kaiāwhina AI: herea ki tētahi mahere kua whakaaetia; ko ngā kōnae kua whakamaheretia anake ka paingia; ka mutu te whakahaere i ngā mea ohorere, ā, ka hoki ki te MAHERE. Ka ōrite ngā rerekētanga komiti ki ngā rerekētanga o te mahere, rārangi ki rārangi.
A documented decision in which the AI assistant consults the Tractatus framework before making an architectural change. Each consultation is recorded both as a markdown document and as a database record on each app server, so the consultation history is verifiable.
[DeepL] He whakataunga kua tuhia e pā ana ki te kaiawhina AI e tirotiro ana i te anga Tractatus i mua i te whakarerekē hanganga. Ka tuhia ia tirotiro hei tuhinga markdown, hei rēhita raraunga hoki i ia tūmau taupānga, kia taea ai te whakamana i te hītori o ngā tirotiro.
The standing rule that the session handoff document is created at the first natural breakpoint of a session and updated continuously through the session — never reconstructed at the end. Each meaningful event becomes a row in the change log.
[DeepL] Ko te ture tū tonu, ka waihangatia te tuhinga whakawhiti wāhanga i te wā wehenga ā-taiao tuatahi o te wāhanga, ā, ka whakahouhia tonu i te roanga o te wāhanga — kāore e hangaia anō i te mutunga. Ka noho ia kaupapa whai tikanga hei rarangi i te rārangi panoni.
A short period during which active members are warned of an imminent service interruption and then locked out while a deploy runs. Mandatory before every push that triggers a service restart on the application servers.
[DeepL] He wā poto e whakatūpatohia ana ngā mema kaha mō tētahi aukatinga ratonga e tata mai ana, ā, ka aukatia rātou i te uru i te wā e rere ana te deploy. Me mahia i mua i ia push e whakaoho ana i te tīmatanga anō o te ratonga i runga i ngā tūmau taupānga.
A working mode for the AI assistant that builds the platform: outputs only, no state-changing actions. The plan must restate the task, list files to modify with diff summary, list scope exclusions, list reversibility per step, list verification steps, and list unknowns or blockers. Persisted to docs/plans/PLAN_*.md.
[DeepL] He aratau mahi mō te kaiāwhina AI e hanga ana i te tūāpapa: whakaputa anake, kāore he mahi e whakarerekē ana i te tūnga. Me whakahoki anō te mahere i te mahi, me rārangi i ngā kōnae hei whakarerekē me te whakarāpopototanga diff, me rārangi i ngā aukatinga ā-rohe, me rārangi i te āhei huri whakamuri mō ia hikoinga, me rārangi i ngā hikoinga whakamana, me rārangi i ngā mea kāore i te mōhiotia, i ngā ārai rānei. Ka penapenahia ki docs/plans/PLAN_*.md.
A three-line falsifiability declaration the AI assistant must output before any state-changing action: ACTION (exact command), EXPECT (one-line success condition), STOP-IF (one-line invalidation condition). After the action: a one-line OBSERVED note. Defeats the failure mode in which confident multi-paragraph reasoning hides an action's actual shape from operator audit.
[DeepL] He whakapuaki whakamātautau e toru ngā rārangi me whakaputa e te kaiāwhina AI i mua i tētahi mahi e whakarerekē ana i te tūnga: ACTION (whakahau tika), EXPECT (tikanga angitu kotahi te rārangi), STOP-IF (tikanga whakakore kotahi te rārangi). I muri i te mahi: he tuhipoka kotahi te rārangi mō OBSERVED. Ka aukati i te aratau hapa e huna ana e te whakaaroaro mārama, whai wāhanga maha, te āhua tūturu o tētahi mahi i te aroturuki a te kaiwhakahaere.
A working mode for the AI assistant: runs automatically after EXECUTE; observed results, not assumed; one line per plan claim with the actual command, stdout/stderr, and exit code. On failure, surface and wait — do not patch on the fly.
[DeepL] He aratau mahi mō te kaiāwhina AI: ka rere aunoa i muri i te EXECUTE; ngā hua i kitea, ehara i te mea i whakapaetia; kotahi rārangi mō ia kerēme mahere me te whakahau tūturu, stdout/stderr, me te waehere putanga. Ki te hē, whakaatuhia ngā hapa ka tatari — kaua e whakatika i te wā tonu.
Infrastructure & DeploymentHanganga Tūāpapa me te Tukunga(16)
Airwallex — Airwallex
A New Zealand-licenced payment provider. The platform's sole payment processor. Stripe was removed in 2026 because of US ownership; Paddle was rejected; Airwallex remains the single hardcoded provider.
[DeepL] He kaiwhakarato utu kua whiwhi raihana i Aotearoa. Ko ia anake te kaiwhakahaere utu o te papanga. I tangohia a Stripe i te tau 2026 nā te mea he rangatiratanga nō Amerika; i whakakorehia a Paddle; kei te noho tonu a Airwallex hei kaiwhakarato kotahi kua tuhia ki roto i te waehere.
The total set of endpoints, methods, parameters, and response shapes that an API exposes to its callers — everything an external program can call. Larger surface = more to test, more to document, more to maintain, more to break when something changes. Related to but broader than "attack surface" (which is the security-framed subset that an attacker could reach).
[DeepL] Ko te huinga katoa o ngā tauranga mutunga, ngā tikanga, ngā tawhā, me ngā āhua whakautu e whakapuaki ana te API ki ōna kaiwhakawhiti karanga — ko ngā mea katoa ka taea e tētahi hōtaka o waho te karanga. He mata nui ake = he nui ake hei whakamātautau, hei tuhi tuhinga, hei tiaki, hei pakaru ina huri tētahi mea. He hononga ki te "mata whakaeke", engari he whānui ake (ko te wāhanga i raro i te anga haumaru e taea ana e te kaiwhakaeke te tae atu).
A New Zealand-owned cloud provider used by the platform for both compute and object storage. Hosts the GPU on which the situated language layer runs during business hours, and the application server that runs the Catalyst-tenant subdomains.
[DeepL] He kaiwhakarato kapua nō Aotearoa e whakamahia ana e te papa hei tukatuka me te rokiroki o ngā tūemi. E manaaki ana i te GPU e rere ana te paparanga reo i ngā haora mahi, me te tūmau taupānga e whakahaere ana i ngā subdomains o Catalyst-tenant.
A model-inference path that uses ordinary processors rather than a GPU when no GPU is available. Slower, but keeps the system serviceable. The platform's CPU-fallback path is part of how the inference layer stays inside the sovereign envelope at all times.
[DeepL] He ara whakatau tauira e whakamahi ana i ngā pūtukatuka noa, kaua ko te GPU, ina kāore he GPU e wātea ana. He pōturi ake, engari ka mau tonu te pūnaha kia taea tonu te whakamahi. Ko te ara hokinga ki te CPU o te papaanga he wāhanga o te āhua e noho ai te paparanga whakatau i roto tonu i te anga rangatiratanga i ngā wā katoa.
A graphics card connected to a workstation via a high-speed external interface. The project uses a home eGPU as the off-business-hours inference path, providing failover when the Catalyst A6000 is shelved.
[DeepL] He kāri whakairoiro kua honoa ki tētahi papamahi mā tētahi atanga ā-waho tere-nui. Ka whakamahi te kaupapa i tētahi eGPU kāinga hei huarahi whakatau i ngā wā kāore i te mahi pakihi, ā, ka whakarato whakawhiti aunoa ina whakatāhia te Catalyst A6000.
The federation of independent servers running open social protocols (most prominently ActivityPub) — Mastodon, PeerTube, Pixelfed and so on. The platform interacts with the fediverse via Mastodon for non-confidential public communications.
[DeepL] Ko te kotahitanga o ngā tūmau motuhake e whakahaere ana i ngā kawa pāpori tuwhera (ko ActivityPub te mea tino rongonui) — arā ko Mastodon, PeerTube, Pixelfed me ērā atu. Ka hono te papanga ki te fediverse mā Mastodon mō ngā whakawhitiwhiti kōrero tūmatanui kāore i te muna.
The runtime path a request follows from a member's question to the model's answer — including which model is selected, where it runs, what fallback applies if the primary path is unavailable, and what is logged at each step.
[DeepL] Te ara whakahaere e whai ana te tono mai i te pātai a te mema ki te whakautu a te tauira — tae atu ki te ingoa o te tauira kua tohua, te wāhi e whakahaerehia ai, te tikanga whakakapi ka whakamahia mēnā kāore te ara matua e wātea ana, me ngā tuhinga e tuhia ana i ia hikoinga.
A decentralised microblogging platform on the fediverse. The project publishes selected announcements via Mastodon; every post is drafted in markdown, reviewed, and published only on operator instruction.
[DeepL] He papanga rangitaki-iti kore-pokapū i runga i te fediverse. Ka whakaputa te kaupapa i ngā pānui kua tohua mā Mastodon; ia pou ka tuhia i roto i te markdown, ka arotakehia, ā, ka whakaputaina anake i runga i ngā tohutohu a te kaiwhakahaere.
A document-oriented database. The platform's primary data store; tenant-aware via tenantId field on every collection.
[DeepL] He pātengi raraunga e aro ana ki ngā tuhinga. Ko te putunga raraunga matua o te papa; e mōhio ana ki ngā kaipāmu mā te mara tenantId i ia kohinga.
A French-headquartered cloud provider. Hosts the EU-sovereign application server that runs the family-history tenant.
[DeepL] He kaiwhakarato kapua kei Parīhi tōna tari matua. E manaaki ana i te tūmau taupānga rangatiratanga o te Uniana o Europi e whakahaere ana i te teneti hītori whānau.
A point in the inference path at which a policy decision is made — for instance, whether the model may answer a question of a given class. Per-request policy-gate verdicts are logged for audit.
[DeepL] He wāhanga i te ara whakatau e whakatauhia ai tētahi whakatau kaupapa here — hei tauira, mēnā ka taea e te tauira te whakautu i tētahi pātai nō tētahi kāwai kua whakaritea. Ka tuhia ngā whakataunga kuaha kaupapa here mō ia tono hei arotake.
Progressive Web App (PWA) — Taupānga Tukutuku Whanake (PWA)
A web application installable to a device's home screen, with offline capability and push notifications. The Village apps are PWAs.
[DeepL] He taupānga tukutuku ka taea te tāuta ki te mata kāinga o te pūrere, me te āheinga mahi ahakoa kāore he hononga ipurangi, me ngā pānui pana. He PWA ngā taupānga o The Village.
A local agent provided by Proton that exposes a Proton mailbox to standard SMTP and IMAP. The platform sends transactional email through Bridge running on the Catalyst host.
[DeepL] He kaituku ā-rohe nā Proton e whakaatu ana i tētahi pouaka īmēra Proton ki ngā paerewa SMTP me IMAP. Ka tukuna e te papanga ngā īmēra whakawhitinga mā Bridge e rere ana i runga i te tūmau Catalyst.
Read further:Codebase
Was this useful?Thanks — recorded.
REST API — Atanga Hōtaka Taupānga Whakatā
An HTTP-based API style organised around resources (nouns) and standard verbs (GET to read, POST to create, PUT to replace, PATCH to partially update, DELETE to remove). The platform's APIs follow REST conventions; clients authenticate with cookies, send CSRF tokens on writes, and receive JSON responses.
[DeepL] He momo API e hāngai ana ki te HTTP, e whakarōpūhia ana ki ngā rauemi (ngā ingoa) me ngā tūmahi paerewa (GET hei pānui, POST hei waihanga, PUT hei whakakapi, PATCH hei whakahou wāhanga, DELETE hei tango). E ū ana ngā API o te papanga ki ngā tikanga REST; ka whakamana ngā kiritaki mā ngā pihikete, ka tuku tohu CSRF i ngā tuhi, ā, ka whiwhi urupare JSON.
A browser-side script that the platform installs once per device, enabling features like push notifications. The platform's service worker is push-only, with zero caching, to avoid stale-content failure modes.
[DeepL] He tuhinga hōtaka kei te taha o te kaitirotiro e tāutahia kotahi mō ia pūrere e te papanga, e whakaahei ana i ngā āhuatanga pērā i ngā pānui pana. He kaimahi ratonga a te papanga e pana anake ana, kāore he penapena raraunga, kia kore ai e puta ngā hapa o ngā ihirangi tawhito.
A graphics processing unit hosted on infrastructure that is outside US legal reach — in this project's case, an A6000 at Catalyst Cloud during business hours and a non-US home eGPU at other times. Used for inference of the situated language layer.
[DeepL] He pūtukatuka whakairoiro e tū ana i runga i te hanganga kei waho o te mana ture o Amerika — i tēnei kaupapa, he A6000 i Catalyst Cloud i ngā haora mahi, ā, he eGPU kāinga kei waho o Amerika i ētahi atu wā. Ka whakamahia hei whakatau i te paparanga reo e hāngai ana ki te horopaki.
AI Industry & StackTe Ahumahi me te Pae Atamai(52)
Agentic workflow — Rauemi mahi ā-mahi
A multi-step process orchestrated by an AI agent — for instance, research-then-summarise-then-publish, or detect-anomaly-then-investigate-then-alert. The defining shape: an outer loop where the agent plans, acts, observes, and adjusts. The Tractatus framework constrains what an agentic workflow may decide on its own.
[DeepL] He tukanga maha-hipanga e whakaritea ana e tētahi kaiārahi AI — hei tauira, rangahau, whakarāpopoto, whakaputa; rānei kitea he hē, rangahau, whakamōhio. Ko te āhua matua: he porowhita o waho e whakamahere ana, e mahi ana, e aroturuki ana, e whakatikatika ana te kaiārahi. Ka herea e te anga Tractatus ngā whakatau ka taea e tētahi rerenga mahi ā-kaiārahi te whakatau māna anō.
A system in which a language model is given goals, a set of tools, and the autonomy to choose which tool to call next based on intermediate results — so that a single high-level instruction can drive a sequence of actions. Distinguished from a single-turn chatbot by the agent's loop over actions and observations.
[DeepL] He pūnaha e whakawhiwhia ana te tauira reo ki ngā whāinga, ki tētahi kohinga taputapu, me te mana motuhake ki te whiriwhiri ko tēhea taputapu ka karangahia ā muri ake i runga i ngā hua waenga — kia taea ai e tētahi tohutohu taumata-nui kotahi te ārahi i te raupapa mahi. Ka wehe i tētahi chatbot hurihanga kotahi nā te hurihanga a te māngai i waenga i ngā mahi me ngā kitenga.
The general problem of making AI systems behave in accordance with human intent and values. Encompasses RLHF, constitutional AI, red teaming, the design of guardrails, and the architectural choice of which decisions to delegate to AI in the first place. The field has no agreed solution; the architectural alternative is to bound what AI may decide.
[DeepL] Ko te raru whānui o te whakarite kia mahi ngā pūnaha AI i runga i ngā hiahia me ngā uara a te tangata. Ka kapi i roto i tēnei te RLHF, te AI ture-matua, te red teaming, te hoahoa o ngā here ārai, me te kōwhiringa hanganga mō ngā whakataunga hei tuku ki te AI i te tuatahi. Kāore he rongoā kua whakaaetia e te rāngai; ko tētahi atu huarahi hanganga, ko te here i ngā whakataunga ka taea e te AI.
Artificial General Intelligence (AGI) — Atamai Whānui Hangarua (AGI)
An AI system that matches or exceeds human cognitive capability across a broad range of intellectual tasks rather than excelling at one. The term is contested: it has no agreed technical definition, no agreed evaluation, and vendors use it inconsistently — OpenAI, Anthropic, DeepMind, and others each work to their own informal definitions. Frontier vendors increasingly describe their roadmaps as oriented toward AGI; sceptics note that capability benchmarks routinely declared "AGI-adjacent" later prove easy to game and easy to ceiling. Whether current frontier models constitute, approach, or merely simulate the property is a live and largely philosophical debate. Distinct from ASI (Artificial Superintelligence), the further hypothetical of an AI surpassing human capability on substantially every dimension at once.
[DeepL] He pūnaha atamai (AI) e taurite ana, e neke atu rānei i te āheinga hinengaro a te tangata puta noa i ngā momo mahi hinengaro whānui, kaua e tino kaha ki tētahi anake. He kupu tēnei e taupatupatuhia ana: kāore he whakamārama hangarau kua whakaaetia, kāore he aromatawai kua whakaaetia, ā, ka whakamahia e ngā kaihoko i runga i te huarahi kāore i te ōrite — ko OpenAI, Anthropic, DeepMind me ētahi atu e mahi ana i runga i ā rātou ake whakamārama ā-roto. Kei te kaha ake ngā kaihanga hangarau matatau ki te whakaahua i ā rātou mahere ara hei aronga ki te AGI; e kī ana ngā tārewa kāore i te mārama, ā, ka kitea he māmā te tākaro i ngā paearu pūkenga e kīia ana he "AGI-adjacent", ā, he māmā hoki te aukati i tō rātou tipu. He wero ora, he tautohetohe whānui hoki te pātai mēnā e hanga ana, e tata ana, e whakaahua noa rānei ngā tauira hangarau matatau o nāianei i tēnei āhuatanga. He rerekē tēnei i te ASI (Artificial Superintelligence), arā, te ariā whakamua mō tētahi AI e whakawhiti ana i ngā pūkenga a te tangata i ngā āhuatanga katoa i te wā kotahi.
A hypothetical AI system substantially exceeding human capability across substantially every intellectual dimension simultaneously. ASI sits on the speculative end of the AI capability spectrum — beyond AGI in scope and uniformly aspirational; no consensus on whether it is reachable, when, or how it would be governed if it were. The term shapes the safety + alignment discourse but has no operational referent today.
[DeepL] He pūnaha AI āhua whakaaro e tino nui ake ana i ngā pūkenga a te tangata puta noa i ngā āhuatanga hinengaro katoa i te wā kotahi. Kei te pito whakaaroaro o te whānuitanga āheinga AI te ASI — he nui ake i te AGI i tōna whānuitanga, ā, he wawata tonu; kāore he whakaaetanga mō te āhei rānei, mō te wā rānei, mō te pēhea rānei e whakahaere ai mēnā ka taea. Ka ārahi te kupu nei i te kōrero mō te haumaru me te whakatikatika, engari kāore he tauira whakahaere i tēnei rā.
Read further:AI industry standard; speculative term
Was this useful?Thanks — recorded.
Chain of thought (CoT) — Meitua whakaaro (CoT)
A prompting technique (and now a training target) in which the model is asked to write out its intermediate reasoning before its final answer — typically improving accuracy on multi-step problems. The basis on which the recent reasoning models were built.
[DeepL] He tikanga akiaki (ā, ināianei he whāinga whakangungu) e tonoa ana te tauira kia tuhi i āna whakaaro waenga i mua i tana whakautu whakamutunga — he mea whakapai ake i te tika o ngā raruraru maha-hipanga. Ko te tūāpapa i hangā ai ngā tauira whakaaro o nā tata nei.
The pre-processing step in a RAG pipeline that splits source documents into passages of a chosen size — small enough to embed cleanly, large enough to carry meaningful context. Chunk-size choice and chunk-overlap policy are among the most consequential design decisions in a RAG system.
[DeepL] Ko te hipanga whakarite-mua i roto i te paipa RAG e wehe ana i ngā tuhinga pūtake ki ngā wahanga i te rahi kua tohua — iti rawa kia taea te whakauru maeneene, nui rawa kia mau ai te horopaki whai tikanga. Ko te kōwhiri rahi wāhanga me te kaupapa here whakawhāiti wāhanga ētahi o ngā whakatau hoahoa tino hua nui rawa atu i roto i tētahi pūnaha RAG.
Closed-weight / proprietary model — Mōdeli taumaha kati / mōdeli rangatōpū
A model whose parameters are not released — it is only accessible behind the vendor's API. GPT-4 / GPT-5, Claude 4, and Gemini are closed-weight. Closed-weight models typically lead capability benchmarks but cannot be self-hosted.
[DeepL] He tauira kāore ōna tawhā i tukuna — ka taea anake te uru mā te API a te kaiwhakarato. He tauira taumaha kati a GPT-4 / GPT-5, Claude 4, me Gemini. Ko ngā tauira taumaha kati e ārahi ana i ngā paearu āheinga, engari kāore e taea te manaaki māu anō.
The platform's own published constitution — the rules and principles the platform is built to uphold, separate from any AI's training. Members can read it; the AI assistant building the platform is held to it. Distinct from constitutional-AI training, though conceptually related.
[DeepL] Ko te ture matua kua whakaputaina e te papaanga — ko ngā ture me ngā mātāpono e hangaia ana te papaanga kia ū ki a rātou, motuhake i ngā whakangungu a te AI. Ka taea e ngā mema te pānui; ā, me ū hoki te kaiāwhina AI e hanga ana i te papaanga ki ēnei ture. He rerekē tēnei i ngā whakangungu AI ā-ture matua, ahakoa he hononga ā-ariā.
A training method (introduced by Anthropic in 2022) in which the model is given a written constitution — a list of values it should uphold — and then trained to critique and revise its own outputs against that constitution. A direct conceptual ancestor of the platform's Constitution-of-MDSL framing.
[DeepL] He tikanga whakangungu (i whakaurua e Anthropic i te tau 2022) e whakawhiwhia ana te tauira ki tētahi ture ā-tuhi — he rārangi uara hei tautoko — ā, ka whakangungua kia arotake, kia whakarerekē hoki i āna ake hua i raro i taua ture. He tūpuna ariā tūturu o te anga 'Constitution-of-MDSL' o te papanga.
The maximum number of tokens a model can attend to in one request — including the system prompt, the conversation so far, retrieved documents, and the generated response. Frontier models in 2025-2026 commonly support 200K to 2M token windows; larger windows enable longer documents but with rising cost and (often) degrading quality past a point.
[DeepL] Ko te nui rawa o ngā tohu ka taea e tētahi tauira te aro ki roto i tētahi tono kotahi — tae atu ki te tono pūnaha, te kōrero tae noa ki tēnei wā, ngā tuhinga kua tiki, me te whakautu kua whakaputaina. I te tau 2025–2026, ka tautoko whānuitia e ngā tauira Frontier ngā matapihi tohu 200,000 ki te 2,000,000; mā ngā matapihi nui ake e āhei ai ngā tuhinga roa ake, engari ka piki haere te utu, ā, he maha ngā wā ka heke te kounga i muri i tētahi taumata.
A Chinese AI lab that released open-weight reasoning and coding models from 2023 onward. DeepSeek R1 (early 2025) demonstrated frontier-class reasoning capability at materially lower training cost than US-vendor frontier models, drawing wide industry attention. Open-weight; downloadable and self-hostable.
[DeepL] He taiwhanga rangahau AI nō Haina i tukuna ngā tauira whakaaro me ngā tauira waehere taumaha-tuwhera mai i te tau 2023. I whakaatu a DeepSeek R1 (i te tīmatanga o te tau 2025) i te āheinga whakaaro taumata-mātanga, ā, he iti ake ngā utu whakangungu i ngā tauira taumata-mātanga a ngā kaiwhakarato o Amerika, nā reira i kukume ai i te aro nui o te umanga. Taumaha-tuwhera; ka taea te tikiake, ka taea hoki te manaaki māu anō.
Direct Preference Optimization (DPO) — Te Whakapai Tūtohu Tūturu (DPO)
An alternative to RLHF for post-training: directly optimises the model on preference pairs without fitting a separate reward model first. Lower compute cost than RLHF in many setups; in active industry use from 2023 onward. Like RLHF, one of the weight-modification protocols the platform has tested and not adopted.
[DeepL] He huarahi kē atu ki te RLHF mō muri i te whakangungu: ka whakapai tika i te tauira ki runga i ngā tokorua manakohanga, kāore e hiahiatia kia whakaurua tuatahi he tauira utu motuhake. He māmā, he iti ake te utu i te RLHF; kua whānuitia te whakamahinga mai i te tau 2023. Pērā i te RLHF, ko tētahi o ngā kawa whakarerekē taumaha i whakamātauria e te papaanga, engari kāore i whakaaetia.
A smaller model trained to imitate a larger 'teacher' model — preserving much of the teacher's behaviour at lower inference cost. DeepSeek's R1 distillates and Llama's 'instruct' variants are distillations. Distillation is one of the weight-modification protocols the platform has tested and not adopted.
[DeepL] He tauira iti ake i whakangungua kia tārite i tētahi tauira 'kaiako' nui ake — e tiaki ana i te nuinga o ngā whanonga a te kaiako, ā, he iti ake te utu whakatau. Ko te R1 a DeepSeek he whakawhāngai, ā, ko ngā momo 'instruct' a Llama he whakawhāngai anō. Ko te whakawhāngai tētahi o ngā kawa whakarerekē taumaha i whakamātauria e te papanga, engari kāore i whakaaetia.
A numeric representation of a piece of content — text, image, audio — such that semantically similar items are nearer to each other in the vector space. The mathematical basis of retrieval-augmented generation, semantic search, and the Guardian Agents' verification.
[DeepL] He whakaaturanga tau mō tētahi wāhanga ihirangi — kuputuhi, whakaahua, oro — kia tata ake ai ngā mea ōrite ā-aronga ki a rātou anō i roto i te wāhi ā-vector. Ko te tūāpapa pāngarau mō te whakaputa kua whakakahaia mā te tiki ihirangi, te rapu ā-aronga, me te whakamana a ngā Guardian Agents.
A specialised model that takes content and produces an embedding vector. Often distinct from the generative model used to write answers. OpenAI's text-embedding-3, Cohere Embed, and open models like nomic-embed and bge-m3 are common choices.
[DeepL] He tauira motuhake e tango ana i te ihirangi, ā, ka whakaputa i tētahi huinga ā-roto. He maha ngā wā ka rerekē i te tauira whakaputa e whakamahia ana ki te tuhi i ngā whakautu. Ko text-embedding-3 a OpenAI, ko Cohere Embed, me ērā atu tauira tuwhera pērā i a nomic-embed me bge-m3 ētahi kōwhiringa noa.
Extended thinking / reasoning tokens — Tohu whānui mō te whakaaro me te arorau
A capability in 2025-2026 frontier models (Claude's extended thinking, OpenAI's o-series reasoning) where the model spends additional inference time and tokens on hidden internal reasoning before producing the visible answer. Trades inference cost for better outcomes on hard problems.
[DeepL] He āheinga i roto i ngā tauira ā-mua o te tau 2025–2026 (te whakaaro whānui a Claude, te whakaaro o-series a OpenAI), e whakapau ana te tauira i te wā tātaitanga tāpiri me ngā tohu tātaitanga ki te whakaaro ā-roto huna i mua i te whakaputa i te whakautu e kitea ana. Ka whakawhiti i te utu tātaitanga kia pai ake ngā hua mō ngā raru uaua.
A weight-modification technique that continues training a base model on new data, typically smaller and more specialised than the original training corpus. Part of the family of approaches the project has tested and not adopted.
[DeepL] He tikanga whakarerekē taumaha e haere tonu ana te whakangungu i tētahi tauira pūtake ki ngā raraunga hou, he iti ake, he motuhake ake i te kohinga whakangungu taketake. He wāhanga nō te whānau o ngā huarahi kua whakamātauhia e te kaupapa, engari kāore i whakaaetia.
A large, general-purpose model — usually a language model — trained on a broad corpus and used as the starting point for many downstream applications. The term covers models in the GPT, Claude, Gemini, and Llama families. Distinguished from task-specific models which are trained narrowly for one job.
[DeepL] He tauira nui, whānui te whakamahinga — he tauira reo i te nuinga o te wā — i whakangungua ki tētahi kohinga raraunga whānui, ā, ka whakamahia hei tīmatanga mō ngā taupānga maha e whai ake nei. Ka kapi tēnei kupu i ngā tauira o ngā whānau GPT, Claude, Gemini, me Llama. He rerekē ēnei i ngā tauira motuhake mō tētahi mahi kotahi, i whakangungua kia arotahi ki taua mahi anake.
A general-purpose vendor language model at or near the leading edge of capability — typically trained by a US-based provider on a global corpus. The project's situated language layers exist as community-controlled alternatives so that mediation of community queries does not import a frontier model's training assumptions.
[DeepL] He tauira reo kaiwhakarato whānui kei te taumata tiketike rānei, tata rānei ki te taumata tiketike rawa o ngā āheinga — i te nuinga o te wā ka whakangungua e tētahi kaiwhakarato kei Amerika ki runga i tētahi kohinga raraunga ā-ao. Ko ngā paparanga reo ā-horopaki o te kaupapa e tū ana hei kōwhiringa e whakahaerehia ana e te hapori, kia kore ai e kawemai ngā whakapae whakangungu o tētahi tauira rohe ki te whakahaere pātai a te hapori.
The property of a model's answer being anchored in a retrievable, citable source rather than confabulated from training memory. RAG is the architectural answer to the grounding problem; the platform's FAQ-layering plus citation-discipline check are its grounding layers.
[DeepL] Ko te āhuatanga o te whakautu a tētahi tauira e ū ana ki tētahi puna e taea te tiki, e taea te whakahua, kaua e hangaia noa iho i ngā mahara whakangungu. Ko RAG te whakautu hanganga ki te raru o te whakapūtanga; ko te paparanga FAQ o te tūāpapa me te whakamātautau i te tika o ngā tohutoro ōna ngā paparanga whakapūtanga.
A separate layer (often a smaller model or a rules engine) that screens model inputs and outputs for policy violations — refusing the request, redacting the answer, or escalating. Distinct from the model's own behaviour; the platform's defence-in-depth principle requires both.
[DeepL] He paparanga motuhake (he tauira iti ake rānei, he pūnaha ture rānei) e tirotiro ana i ngā tāuru me ngā putanga o te tauira mō ngā hē ture — ka whakahē i te tono, ka whakakore i te whakautu, ka whakamōhio rānei. He rerekē ki ngā whanonga o te tauira anō; e hiahiatia ana e te mātāpono ārai hohonu o te papaanga kia rua.
A fluent but factually wrong model output — a confident answer with no real grounding in the corpus or the world. The defining failure mode of generative language models; the practical motivation for RAG, citation discipline, and FAQ-layered grounding rather than reliance on the model's training memory.
[DeepL] He whakaputanga tauira e rere noa ana, engari he hē ki ngā kōrero pono — he whakautu māia kāore he tūāpapa tūturu i roto i te kohinga raraunga, i te ao rānei. Koinei te aronga ngoikore matua o ngā tauira reo whakaputa; ko te take whaihua mō te RAG, te tikanga tohutoro, me te tūāpapa mā ngā pātai auau (FAQ), kaua e whakawhirinaki ki te mahara whakangungu o te tauira.
The per-token cost of running a model — quoted separately for input tokens and output tokens, typically per million. In 2025-2026, frontier-model output tokens range from a few dollars to tens of dollars per million; reasoning-model output (with extended thinking) is several multiples of that. Drives most architectural decisions about which model handles which class of request.
[DeepL] Ko te utu ia tohu mō te whakahaere tauira — e whakahuatia motuhake ana mō ngā tohu urunga me ngā tohu putanga, ā, he nuinga ia miriona. I te tau 2025–2026, kei waenga i ngā tāra torutoru ki ngā tekau tāra ia miriona ngā tohu putanga o ngā tauira ā-mua; ko ngā tohu putanga o ngā tauira whakaaro (me te whakaaro whānui) he maha ngā wā atu i tērā. Koinei te mea e ārahi ana i te nuinga o ngā whakatau hoahoanga mō te tauira e whakahaere ana i tēhea momo tono.
Software that loads a model into memory and exposes it as a network service — typically as an OpenAI-compatible HTTP API. Examples include vLLM (high throughput, GPU), Ollama (local convenience), llama.cpp (CPU + Apple Silicon), and TGI. The platform's situated language layer runs behind an inference server on the sovereign GPU.
[DeepL] He pūmanawa e uta ana i tētahi tauira ki roto i te mahara, ā, e whakaatu ana hei ratonga whatunga — arā, he HTTP API e hāngai ana ki OpenAI. Ko ētahi tauira ko vLLM (teitei te tere whakawhiti, GPU), Ollama (mō te whakamahinga ā-rohe), llama.cpp (CPU + Apple Silicon), me TGI. Ka rere te paparanga reo ā-horopaki o te tūāpapa i muri i tētahi tūmau whakatau kei runga i te GPU rangatira.
A user-side attempt to bypass a model's safety training — typically by elaborate role-play scenarios, instruction-laundering through fictional framings, or multi-step setup that gradually erodes the refusal boundary. Distinct from prompt injection (which targets the system prompt from data) but related.
[DeepL] He whakamātau a te kaiwhakamahi ki te karo i ngā whakangungu haumaru o te tauira — i te nuinga o te wā mā ngā āhua tākaro tūranga matatini, te horoi tohutohu mā ngā anga pūrākau, rānei te whakarite i ngā hikoinga maha e āta whakaiti ana i te rohe aukati. He rerekē ki te werohanga tono (e aro ana ki te tono pūnaha mai i ngā raraunga), engari he hononga.
A neural-network model trained on text such that, given a prompt, it produces fluent natural-language output. Modern AI assistants are LLMs; the situated language layer is a community-scoped, smaller-than-frontier LLM.
[DeepL] He tauira whatunga neorana i whakangungua ki ngā tuhinga, kia puta ai he whakaputanga reo māori tōtika ina whakawhiwhia ki tētahi whakahau. Ko ngā kaiāwhina AI o nāianei he LLM; ko te paparanga reo ā-horopaki he LLM iti ake i ngā LLM matua, ā, he mea whakarite mō te hapori.
An open-weight large language model family from Meta, first released in 2023 and now in its third major generation (Llama 3, with point releases through Llama 3.3). Sizes range from 1B to 405B parameters. Released under the Llama Community Licence — open weights, with restrictions for very large commercial deployments. One of the most-used open-weight families in the wider industry; not currently the platform's base.
[DeepL] He whānau tauira reo nui taumaha-tuwhera nō Meta, i tukuna tuatahi i te tau 2023, ā, kei te whakatipuranga matua tuatoru ināianei (Llama 3, me ngā putanga tohu tae noa ki te Llama 3.3). Kei waenga i te 1 piriona me te 405 piriona ngā tawhā. I tukuna i raro i te Raihana Hapori Llama — he taumaha tuwhera, me ētahi here mō ngā whakamahinga hokohoko tino nui. Ko tētahi o ngā whānau taumaha tuwhera e tino whakamahia ana i te umanga whānui; kāore i tēnei wā hei tūāpapa o te papanga.
A parameter-efficient fine-tuning technique that adds small low-rank update matrices alongside the base model's weights, leaving the base unchanged. Cheaper to train and faster to swap than full fine-tuning. One of the weight-modification protocols the platform tested and did not adopt.
[DeepL] He tikanga whakangāwari āta whai hua mō ngā tawhā, e tāpiri ana i ngā matatini whakahou rahi-iti, ā-tūāpapa, ki ngā taumaha o te tauira taketake, ā, kāore e whakarerekē i te tauira taketake. He iti ake te utu whakangungu, ā, he tere ake te whakawhiti i te whakangāwari āta katoa. Ko tētahi o ngā kawa whakarerekē taumaha i whakamātauria e te papanga, engari kāore i whakamahia.
A French AI company that releases an open-weight model family of the same name — Mistral 7B, Mixtral 8x7B (a Mixture-of-Experts variant), Mistral Small / Medium / Large. Notable for European provenance (relevant to vendor-prohibition reasoning) and for popularising the MoE architecture in the open-weight space.
[DeepL] He kamupene AI Wīwī e tuku ana i tētahi whānau tauira taumaha tuwhera o taua ingoa anō — Mistral 7B, Mixtral 8x7B (he momo Mixture-of-Experts), Mistral Iti / Waenga / Nui. He mea rongonui nā tōna takenga mai i Uropi (e hāngai ana ki ngā whakaaro aukati kaiwhakarato) me tōna whakatairanga i te hanganga MoE i te ao tauira taumaha tuwhera.
Mixture of Experts (MoE) — Kohinga o ngā Mātanga (MoE)
A model architecture in which only a subset of the parameters (one or more 'expert' sub-networks) is activated for any given input, allowing very large total parameter counts at lower inference cost. Mixtral, DeepSeek, and recent Llama-class models use MoE architectures.
[DeepL] He hanganga tauira e whakahohehia ana he wāhanga noa o ngā tawhā (kotahi, neke atu rānei o ngā whatunga-iti 'mātanga') mō ia tāuru, ā, ka āhei ki te whai tatauranga tawhā katoa nui rawa atu i te utu aromatawai iti ake. Ka whakamahi a Mixtral, DeepSeek, me ngā tauira kāwai Llama hou i ngā hanganga MoE.
Model Context Protocol (MCP) — Tūtohu Horopaki Tauira (MCP)
An open protocol introduced by Anthropic in late 2024 for connecting AI models to tools, data sources, and the local environment in a standardised way. Lets a developer expose a database, a filesystem, or an API to an AI assistant once and have it work across all MCP-aware clients.
[DeepL] He kawa tuwhera i whakaurua e Anthropic i te mutunga o te tau 2024 hei hono i ngā tauira AI ki ngā taputapu, ngā puna raraunga, me te taiao ā-rohe i runga i tētahi huarahi paerewa. Ka āhei te kaiwhakawhanake ki te whakaatu kotahi i tētahi pātengi raraunga, tētahi pūnaha kōnae, tētahi atanga hōtaka taupānga rānei ki tētahi kaiāwhina AI, ā, ka mahi tonu puta noa i ngā kiritaki katoa e mōhio ana ki te MCP.
Read further:Anthropic 2024-11; AI industry standard 2025
Was this useful?Thanks — recorded.
Model family / generation — Mōdeli whānau / whakatupuranga
A series of related models from a single vendor — for instance the GPT series (GPT-3, GPT-4, GPT-5), Claude (Claude 3, Claude 4), Gemini, or Llama. Successive generations typically improve along several axes (capability, speed, cost) but with shifting trade-offs that matter for any production deployment.
[DeepL] He raupapa tauira e hono ana nō tētahi kaiwhakarato kotahi — hei tauira, te raupapa GPT (GPT-3, GPT-4, GPT-5), Claude (Claude 3, Claude 4), Gemini, Llama rānei. Ka whakapai ake ngā whakatipuranga e whai ake ana i ngā āhuatanga maha (pērā i te āheinga, te tere, te utu), engari me ngā whakawhitinga painga e rereke ana, ā, he mea nui ēnei mō ia whakaurunga whakaputa.
A system in which multiple AI agents cooperate, divide labour, or supervise one another — for instance a planner agent that delegates sub-tasks to specialist worker agents, or a critic agent that reviews another agent's draft. Adds capability but compounds risk; the Pluralistic Deliberation Orchestrator is one architectural answer to multi-agent values conflicts.
[DeepL] He pūnaha e mahi ngātahi ana ngā māngai AI maha, e wehe ana i ngā mahi, e ārahi rānei i a rātou anō — hei tauira, he māngai whakarite mahere e tuku ana i ngā mahi iti ki ngā māngai kaimahi mātanga, he māngai kaiwhakawā e arotake ana i te tuhi tuatahi a tētahi atu māngai. Ka tāpiri i te āheinga, engari ka whakapiki i te tūraru; ko te Kaiwhakahaere Whiriwhiringa Rōpū tētahi o ngā otinga hanganga hei whakatau i ngā pakarutanga uara i waenga i ngā māngai maha.
A model that accepts and produces more than one kind of input or output — text plus images, audio, video, or structured data. Most recent frontier models are multimodal by default. The platform's situated language layers are currently text-only.
[DeepL] He tauira e whakaae ana, e whakaputa ana hoki i ngā momo urunga me ngā momo putanga maha — kuputuhi me ngā whakaahua, oro, ataata, rānei raraunga hanganga. Ko te nuinga o ngā tauira Frontier hou he maha-āhua i te taunoa. Ko ngā paparanga reo ā-horopaki o te papanga he kuputuhi anake i tēnei wā.
A model whose trained parameters are publicly downloadable, so anyone can run it on their own infrastructure. Typically released under a licence (Llama, Qwen, Mistral, DeepSeek) — "open weight" is not always "open source" because training data + code may not be released. The platform's base model (Qwen2 14B) is open weight.
[DeepL] He tauira ā-rōpū e taea ana e te katoa te tiki i āna tawhā kua whakangungua, kia taea ai e ia te whakahaere i runga i āna ake hanganga. I te nuinga o te wā ka tukuna i raro i tētahi raihana (Llama, Qwen, Mistral, DeepSeek) — ehara i te mea ko te "taumaha tuwhera" he "puna tuwhera" tonu, nā te mea kāore pea i tukuna ngā raraunga whakangungu me te waehere. Ko te tauira taketake o te papanga (Qwen2 14B) he taumaha tuwhera.
Pre-training vs post-training — I mua i te whakangungu vs i muri i te whakangungu
Pre-training is the initial, expensive run on a vast corpus that gives the model its base capability. Post-training (sometimes called alignment training) is the much cheaper subsequent phase — instruction tuning, RLHF, DPO, constitutional AI training — that shapes the model's behaviour for end-user use.
[DeepL] Ko te whakangungu tuatahi te oma tuatahi, he utu nui, i runga i tētahi kohinga raraunga nui e whakarato ana i te pūkenga taketake o te tauira. Ko te whakangungu muri (i ētahi wā ka kīia ko te whakangungu whakaritenga) te wāhanga e whai ake nei, he iti ake te utu — te whakatikatika whakahau, RLHF, DPO, me te whakangungu AI ā-ture — e ārahi ana i te whanonga o te tauira mō te whakamahinga a te kaiwhakamahi whakamutunga.
An attack in which user-controlled input contains instructions intended to override the system prompt — for instance, a document the model is summarising contains "ignore previous instructions and..." An open problem in 2025-2026; defences include separating data from instructions, output-side filters, and the architectural rule that the AI should not be given irrevocable authority.
[DeepL] He whakaeke e whakamahi ana i ngā urunga e whakahaerehia ana e te kaiwhakamahi, ā, kei roto he tohutohu hei whakakapi i te karanga pūnaha — hei tauira, kei roto i tētahi tuhinga e whakarāpopototia ana e te tauira ngā kupu "kaua e aro ki ngā tohutohu o mua, ā…" He raru tuwhera i te tau 2025–2026; ko ētahi ārai ko te wehe i ngā raraunga me ngā tohutohu, ngā tātari i te taha whakaputa, me te ture hanganga e mea ana kāore e tika kia whakawhiwhia te AI ki te mana kāore e taea te whakahoki.
Reducing the numerical precision of a model's parameters — for instance from 16-bit to 4-bit — to shrink memory use and speed up inference, at some cost in quality. Allows large open-weight models to run on consumer GPUs. GGUF, AWQ, and GPTQ are common quantization formats.
[DeepL] Te whakaiti i te tika tau o ngā tawhā o tētahi tauira — hei tauira, mai i te 16-bit ki te 4-bit — kia whakaiti i te whakamahinga mahara, kia tere ake ai te whakatau, ahakoa he utu ki te kounga. Ka āhei ngā tauira nui me ngā taumata taumaha tuwhera ki te whakahaere i runga i ngā GPU a ngā kaiwhakamahi. Ko GGUF, AWQ, me GPTQ ērā ngā hōputu whakawhāiti tau noa.
An open-weight large language model family from Alibaba Cloud (Tongyi Qianwen), released in successive generations from 2023 onward (Qwen 1, Qwen 1.5, Qwen 2, Qwen 2.5, Qwen 3). Available in many sizes from sub-1B to several hundred billion parameters; popular choice for on-premises deployment because the weights can be downloaded and run locally. The platform's situated language layers are built on Qwen 2 14B as the base model.
[DeepL] He whānau tauira reo nui taumaha-tuwhera nō Alibaba Cloud (Tongyi Qianwen), i tukuna i ngā whakatipuranga whaiaronga mai i te tau 2023 (Qwen 1, Qwen 1.5, Qwen 2, Qwen 2.5, Qwen 3). E wātea ana i ngā rahi maha mai i raro i te kotahi piriona tae noa ki ngā rau piriona tawhā; he kōwhiringa rongonui mō te whakaurunga ā-rohe nā te mea ka taea te tiki i ngā taumaha kia whakahaerehia ā-rohe. Ko ngā paparanga reo ā-horopaki o te tūāpapa i hangaia i runga i te Qwen 2 14B hei tauira pūtake.
A model class that generates an internal chain-of-thought (often as hidden reasoning tokens) before producing the visible answer, trading more compute for better outcomes on multi-step problems. Examples include OpenAI's o-series (o1, o3) and Anthropic's extended-thinking modes.
[DeepL] He akomanga tauira e waihanga ana i tētahi aho whakaaro ā-roto (he maha ngā wā hei tohu whakaaro huna) i mua i te whakaputa i te whakautu e kitea ana, e whakawhiti ana i te nui ake o te kaha rorohiko mō ngā hua pai ake i ngā raruraru maha-hipanga. Ko ētahi tauira ko te raupapa o a OpenAI (o1, o3) me ngā aratau whakaaro-roa a Anthropic.
Red teaming — Te whakamātautau haumaru mā te taha whakatumatuma
Adversarial testing of an AI system — a team paid to find ways the system can be made to misbehave, before real users do. A standard practice for frontier models; the structural alternative is the architectural bounding the Tractatus framework provides.
[DeepL] Whakamātautau whakatumatuma o tētahi pūnaha AI — he rōpū i utua kia rapu huarahi hei whakararu i te pūnaha, i mua i te whakamahi a ngā kaiwhakamahi tūturu. He tikanga paerewa mō ngā tauira ā-mua; ko te kōwhiringa hanganga ko te here hoahoanga e whakarato ana e te anga Tractatus.
A second pass after vector-similarity retrieval that uses a more expensive cross-encoder model to re-score the top candidates — typically pulling the most relevant 3-5 passages out of an initial top-50. Substantially improves RAG accuracy for the price of one extra inference call.
[DeepL] He huringa tuarua i muri i te tiki raraunga mā te taurite ā-vector, e whakamahi ana i tētahi tauira cross-encoder utu nui ake hei whakahou i ngā tohu o ngā kaitono matua — ko te nuinga ka tango i ngā wāhanga e toru ki te rima tino hāngai i roto i te rārangi 50 tuatahi. Ka tino whakapiki i te tika o te RAG mō te utu o tētahi karanga whakamātau anō.
A pattern in which the model is given relevant passages — retrieved from a corpus by similarity search at request time — and asked to ground its answer in them rather than rely on its training memory. The standard alternative to fine-tuning when the goal is to make a model answer accurately about a specific corpus.
[DeepL] He tauira i whakawhiwhia ai te tauira ki ngā wāhanga hāngai — i tangohia mai i tētahi kohinga mā te rapu ōrite i te wā tono — ā, ka tonoa kia ū tōna whakautu ki ēnei wāhanga, kaua ki te whakawhirinaki ki tōna mahara whakangungu. Ko te huarahi paerewa hei kōwhiringa ki te whakangungu taipitopito, ina ko te whāinga kia whakautu tika te tauira mō tētahi kohinga motuhake.
Read further:Lewis et al. 2020; AI industry standard
Was this useful?Thanks — recorded.
RLHF (Reinforcement Learning from Human Feedback) — RLHF (Ako Whakakaha mā ngā Urupare a te Tangata)
A weight-modification approach that uses paired human-preference data to shape model behaviour. One of the protocols the project tested and did not adopt.
[DeepL] He huarahi whakarerekē taumaha e whakamahi ana i ngā raraunga manakohanga tāngata takirua hei ārahi i te whanonga o te tauira. Ko tētahi o ngā kawa i whakamātauria e te kaupapa, engari kāore i whakaaetia.
A language model deliberately smaller than a frontier model — typically a few hundred million to ~30 billion parameters — chosen for speed, lower cost, on-device or sovereign-infrastructure deployment, or task-specificity. The platform's situated language layer is a community-scoped SLM.
[DeepL] He tauira reo i hangaia kia iti ake i tētahi tauira rohe — he nuinga he rau miriona tae noa ki te 30 piriona ngā tawhā — i tīpakohia mō te tere, te utu iti, te whakaurunga ki runga pūrere, ki ngā hanganga motuhake rānei, me te arotahi ki ngā mahi motuhake. Ko te paparanga reo ā-horopaki o te tūāpapa he SLM ā-hapori.
A delivery pattern where the model's output is sent back token-by-token as it is generated, rather than waiting for the full answer. The user sees the answer typing live — better perceived latency at the same total throughput.
[DeepL] He tauira tuku e tukuna ana ngā putanga o te tauira, tohu-ia-tohu, i te wā e whakaputa ana, kaua e tatari kia oti katoa te whakautu. Ka kite te kaiwhakamahi i te whakautu e tuhi ana i te wā tūturu — he pai ake te kitenga o te whakaroa i te rahinga tuku katoa.
A model capability where the output is constrained to match a specific schema — usually JSON conforming to a given shape — so the calling application can parse it reliably. Critical for any agentic workflow where the model's output drives the next programmatic step.
[DeepL] He āheinga tauira e herea ana te putanga kia hāngai ki tētahi anga motuhake — he JSON e ū ana ki tētahi āhua kua whakaritea — kia taea e te taupānga e karanga ana te wetewete i te putanga kia pono. He mea tino hira mō ngā mahi aroturuki e ārahi ana te putanga o te tauira i te hiko whakamahere e whai ake nei.
Instructions given to a model at the start of an interaction — typically separate from the user's message — that set the model's role, constraints, voice, and rules of engagement. The system prompt is the most important single input to model behaviour and is itself a piece of governance the platform owns.
[DeepL] Ngā tohutohu e hoatu ana ki tētahi tauira i te tīmatanga o te whakawhitinga — he motuhake i te karere a te kaiwhakamahi — e tautuhi ana i te tūranga o te tauira, ngā here, te reo, me ngā ture whakawhitinga. Ko te tono pūnaha te tuku kotahi tino hira ki te whanonga o te tauira, ā, he wāhanga whakahaere nō te papaanga.
The atomic units a language model processes — typically sub-word fragments (around 3-4 characters of English on average). The tokenizer converts text to tokens before the model sees it. Cost and context limits are usually expressed in tokens, not characters or words.
[DeepL] Ko ngā wāhanga atomi e tukatukahia ana e tētahi tauira reo — arā, he wāhanga kupu-iti (tata ki te 3–4 pūāhua o te reo Ingarihi i te toharite). Ka hurihia e te kaitokena te tuhinga hei tohu i mua i te kitenga a te tauira. Ka whakapuaki ngā here utu me ngā here horopaki i roto i ngā tohu, ehara i ngā pūāhua, i ngā kupu rānei.
Tool use / function calling — Te whakamahi taputapu / te karanga mahi
The capability for a model to request that the host application run a specific function (search the web, query a database, send an email, write a file) with structured arguments, then incorporate the result into its next turn. The basis on which AI agents act on the world.
[DeepL] Ko te āheinga o tētahi tauira ki te tono ki te taupānga matua kia whakahaere i tētahi mahi motuhake (rapu ipurangi, pātai ki tētahi pātengi raraunga, tuku īmēra, tuhi kōnae) me ngā tūtohu hanganga, ā, ka whakauru i te hua ki tōna hurihanga e whai ake nei. Ko te tūāpapa e mahi ai ngā māngai AI i te ao.
A database optimised for storing and querying embedding vectors by similarity — for example finding the ten passages most semantically similar to a query. Examples: pgvector (PostgreSQL extension), Qdrant, Chroma, Pinecone, Weaviate. MongoDB Atlas also offers vector search.
[DeepL] He pātengi raraunga kua whakapaingia mō te rokiroki me te uiui i ngā huanga whakauru mā te ōritetanga — hei tauira, te kimi i ngā wāhanga tekau e tino ōrite ana ā-aronga ki tētahi uiui. Hei tauira: pgvector (toronga PostgreSQL), Qdrant, Chroma, Pinecone, Weaviate. Ka tukuna hoki e MongoDB Atlas te rapu huanga.
Cultural & Community ConceptsNgā Ariā Ahurea me te Hapori(12)
Channel — Arai
A named conversation space within a community used for ongoing messaging — typically scoped to all members, to a subgroup, or to a specific role. Channels carry the day-to-day exchange that holds a community together between scheduled events.
[DeepL] He wāhi kōrero ingoa kei roto i tētahi hapori, hei whakawhiti karere tonu — he mea whakarite mō ngā mema katoa, mō tētahi rōpū iti, mō tētahi tūranga motuhake rānei. Ka kawea e ngā hongere te whakawhiti kōrero o ia rā e whakakotahi ana i te hapori i waenga i ngā kaupapa kua whakaritea.
A private message between two members, outside any shared channel. DMs stay within the community's own infrastructure — they do not travel through a third-party messaging service.
[DeepL] He karere tūmataiti i waenga i ngā mema e rua, i waho atu i ngā hongere tiritahi. Ka noho ngā DM ki roto i te hanganga ake o te hapori — kāore rātou e haere mā tētahi ratonga karere a tētahi atu.
Foundational philosophical mentor to John G. Stroh on pluralistic thinking and the question of goodness in artificial intelligence. The pluralistic-deliberation commitment that runs through the platform's governance architecture — and the wider conviction that an AI substrate worth building must answer to a substantive notion of goodness, not a procedural one — owes its formative shape to those conversations. Acknowledged in Paper A, Paper B, and the Philosophical Foundations of the Village Project.
[DeepL] Ko ia te kaiārahi arorangi taketake ki a John G. Stroh mō te whakaaro maha-āhua me te pātai mō te pai i roto i te mātauranga hangarua. Ko te ū ki te whiriwhiringa maha-āhua e rere ana i roto i te hanganga whakahaere o te tūāpapa — me te whakapono whānui, arā, me whakautu e tētahi tūāpapa mātauranga hangarua e tika ana kia hangaia ki tētahi ariā nui o te pai, ehara i te mea tikanga anake — nā ēnei kōrero i hanga ai tōna āhua taketake. Kua whakamanahia i roto i te Pepa A, Pepa B, me ngā Pūtake Arorangi o te Kaupapa o te Kāinga.
Updated 2026-05-06 in response to reader feedback — added in response to a reader suggestion
Was this useful?Thanks — recorded.
My Digital Sovereignty Limited — Tōku Rangatiratanga Matihiko Herenga
The New Zealand company that operates the Village platform and authors the Tractatus framework. Director: John G. Stroh. Registered in New Zealand.
[DeepL] Ko te kamupene o Aotearoa e whakahaere ana i te papanga Village, ā, e tuhi ana i te anga Tractatus. Tumuaki: John G. Stroh. Kua rēhitatia ki Aotearoa.
An alert delivered to a member about community activity — new content, mentions, event reminders, invitations. Opt-in per category; delivered through the channels the member has chosen (in-app, email, push).
[DeepL] He whakamōhiotanga ka tukuna ki tētahi mema mō ngā mahi ā-hapori — ngā ihirangi hou, ngā whakahua, ngā whakamaharatanga kaupapa, ngā tono. Ka whiriwhiria ia mema ia kāwai; ka tukuna mā ngā hongere kua tohua e ia (i roto i te taupānga, īmēra, pānui peke).
The operator's operational cockpit at the mdsl tenant — backlog, infrastructure health, deployments, content pipeline, follow-ups, secret rotation, and other cross-tenant operations.
[DeepL] Te rūma whakahaere mahi a te kaiwhakahaere i te kiritaki mdsl — te rārangi mahi e toe ana, te hauora hanganga, ngā whakaurunga, te paipa ihirangi, ngā whaiaronga, te hurihanga muna, me ētahi atu whakahaere whakawhānui ki ngā kiritaki.
An Anglican-communion local congregation. The episcopal Tier-1 cohort is trained on contexts particular to parish administration and pastoral life.
[DeepL] He hāhi ā-rohe o te Whakawhanaungatanga Ānglikana. Ka whakangungua te rōpū tuatahi o ngā epikopā ki ngā horopaki motuhake o te whakahaere pārihi me te oranga ā-wairua.
The publicly-visible (members-only-aware) front page of a Village tenant. Distinct from the in-app member-facing pages: the shopfront is the first thing a prospective member sees.
[DeepL] Te whārangi tuatahi o mua o tētahi kaipāmu o te Hāpori e kitea ana e te katoa (engari ko ngā mema anake e mōhio ana). He rerekē i ngā whārangi o roto i te taupānga e aro ana ki ngā mema: ko te mata o te toa te mea tuatahi ka kitea e tētahi mema pea.
A piece of written content shared within a community — typically a memory, experience, reflection, or report, with optional photos. The platform's primary content type for member-authored material; subject to the same visibility controls as everything else.
[DeepL] He tuhinga i tuhia e tohaina ana i roto i tētahi hapori — he maumahara, he wheako, he whakaaroaro, he pūrongo rānei, ā, ka taea te tāpiri whakaahua. Koinei te momo ihirangi matua o te papanga mō ngā tuhinga a ngā mema; kei raro i ngā whakahaere kitea anō pērā i ērā atu mea katoa.
Refers to the open-source AI safety framework — codebase, documents, and the public site at agenticgovernance.digital — distinct from the My Digital Sovereignty community platform that uses it. Apache 2.0 licensed; hosted on Codeberg and on the project's self-hosted Forgejo.
[DeepL] E pā ana ki te anga haumaru AI puna tuwhera — te pūtake waehere, ngā tuhinga, me te paetukutuku tūmatanui i agenticgovernance.digital — he rerekē ki te papa hapori My Digital Sovereignty e whakamahi ana i a ia. He raihana Apache 2.0; kei Codeberg, ā, kei runga hoki i te Forgejo e mana ana te kaupapa.
The community platform built and operated by My Digital Sovereignty Limited (NZ) — the production application that uses the Tractatus framework. Lives at mysovereignty.digital, with each tenant on its own subdomain. Distinct from the Village product types within it (whānau, rūnanga, parish, etc.).
[DeepL] Te tūāpapa hapori i hangaia, i whakahaerehia e My Digital Sovereignty Limited (NZ) — te taupānga whakaputa e whakamahi ana i te anga Tractatus. Kei mysovereignty.digital, ā, kei ia kaipānga tōna ake subdomain. He rerekē i ngā momo hua o Village kei roto i a ia (whānau, rūnanga, pārihi, me ērā atu).
Wittgenstein, Berlin, Ostrom, Te Ao Māori — Wittgenstein, Berlin, Ostrom, Te Ao Māori
The four intellectual traditions the Tractatus framework draws on: Wittgenstein (the limits of the sayable), Berlin (value pluralism), Ostrom (polycentric governance of commons), and Te Ao Māori (rangatiratanga, kaitiakitanga, mana). Each contributes a distinct architectural insight.
[DeepL] Ko ngā tikanga hinengaro e whā e whakamahia ana e te anga Tractatus: ko Wittgenstein (ngā here o ngā mea ka taea te kī), ko Berlin (te kanorau uara), ko Ostrom (te whakahaere pokapū maha o ngā rawa tūmatanui), me Te Ao Māori (rangatiratanga, kaitiakitanga, mana). Ka whakawhiwhi ia o ēnei i tētahi tirohanga hoahoanga motuhake.